Is there a universal story?

Five of the fifteen diagrams follow figures that the authors drew themselves. They are Freytag’s pyramid, Campbell’s circle, Vonnegut’s graphs, Labov and Waletzky’s diagram, and the six curves of Reagan and colleagues. The other ten are my drawings of what the author wrote. Labov and Waletzky’s own figure is a square standing on one corner, which I have opened out into a line. Campbell’s is a circle crossed by a level line, which he called the “threshold of adventure”, and his hero goes down on the left and comes back up on the right. It marks the call, the helper, the crossing, the tests, the prize, the flight and the return and does not number them, so the seventeen loops are my addition. I know his figure from redrawings and have not seen the page of the book. The one loop on Freytag’s pyramid is mine too.

What the cards leave out

• Aristotle. Aristotle never published the Poetics or edited it properly. It is usually dated to his last years in Athens, about 335 to 323 BC, though parts may be earlier. The knot is from chapter 18, “Every tragedy falls into two parts,—Complication and Unravelling or Denouement”, and the demand for cause from chapter 10, where it “makes all the difference whether any given event is a case of propter hoc or post hoc”. The change of fortune is from chapter 7, where a tragedy should be long enough to “admit of a change from bad fortune to good, or from good fortune to bad”. His recognition, “a change from ignorance to knowledge” (chapter 11), is close to the secret coming out, which is one of the fourteen events tracked here and happens in 73% of the stories read with the longer form.
• The Natyashastra. Its translator, Manomohan Ghosh, concluded that the text as it survives existed by the second century AD, and that later hands had added to it. Bharata gives the five stages of the action and the five segments of the plot as separate lists. Pairing each stage with one segment comes from a later treatise, the tenth-century Daśarūpa.
• Qi, cheng, zhuan, he. According to Jiang Yin’s history of the idea, the four terms are first found together in manuals of poetry from the Yuan dynasty, around 1300. Kirkpatrick and Xu quote one of those teachers, Fan Peng (1272 to 1330, Fan Heng in their spelling), whose first rule is that the opening “needs to be level and straight”. They also cite a dictionary of Chinese rhetoric that gives a Tang poem as an early example of the practice, and a scholar’s view that the scheme had been adopted for prose by the end of the Yuan dynasty, in 1368. The card dates the scheme by the manuals and not by the poems.
• Zeami. The source is his treatise Sandō of 1423. Its translator, Shelley Fenno Quinn, writes that the five-section scheme of a noh play is “premised less on the mounting of the dramatic tension than on the mounting of the musical tension”, which is why the card treats it as a claim about pace.
• Freytag. The pyramid is on page 100 of the first edition and page 115 of the English translation. He says the climax “lies directly in the middle”. The height of the pyramid is not tension. The climax is where the play of the first half meets the counter-play of the second, and on the same page as the figure he names a “last suspense” that comes between the fall and the catastrophe. Textbook versions of the pyramid call its last part the denouement, a word the translation never uses. Freytag’s last part is the catastrophe.
• Polti. Goethe’s remark was made to Eckermann on 14 February 1830: “Gozzi said there were only six-and-thirty tragic situations. Schiller thought there were more, but could never succeed in finding even so many.”
• Propp. His hundred tales are numbers 50 to 151 in the older editions of Afanasyev’s collection (93 to 270 in the newer numbering), and all are fairy tales in the narrow sense, types 300 to 749 in Aarne’s index. He makes two claims that are easy to run together. One is that “the sequence of functions is always identical”. The other is that “by no means do all tales give evidence of all functions”. Later in the book he allows exceptions to the first, and writes that “some functions are capable of changing places”. Claude Lévi-Strauss said of the result, “Before the epoch of formalism we were indeed unaware of what these tales had in common. Now we are deprived of any means of understanding how they differ.”
• Campbell. The sentence usually quoted as the monomyth is “A hero ventures forth from the world of common day into a region of supernatural wonder: fabulous forces are there encountered and a decisive victory is won: the hero comes back from this mysterious adventure with the power to bestow boons on his fellow man.” He did not say that every tale has every stage, and wrote that “many tales isolate and greatly enlarge upon one or two of the typical elements of the full cycle”. He said in his preface that there are differences between the mythologies of mankind, “but this is a book about the similarities”. Folklorists have objected to the way the pattern was assembled. Alan Dundes called it “a synthetic, artificial composite which he fails to apply in toto to any one single hero”, and Barre Toelken wrote that Campbell “could construct a monomyth of the hero, therefore, only by citing those stories which fit his proposed mold, and leaving out equally valid stories which did not fit the pattern.” Robert Segal, in a book about Campbell, observed that he “analyzes few whole myths”. I could reach Campbell’s book only through search snippets of two scanned copies, so I have read the passages quoted here and not the book from end to end. The three critics were read the same way.
• Labov and Waletzky. They wrote that their conclusions “are restricted to the speech communities that we have examined”. In 1972 Labov added a sixth part, the abstract, which comes before the orientation, and wrote that only the complicating action “is essential if we are to recognize a narrative”.
• Todorov. The words quoted are from his essay “The Grammar of Narrative”, which sums up his study of the Decameron and was first printed in 1968, the year before the book. I did not open the book itself.
• Vonnegut. The story of the rejected thesis is his own, from Palm Sunday, and he gives no date for it. The University of Chicago’s notices say that he studied there from 1945 to 1947 and left after his thesis was turned down. His biographer Charles Shields says that the thesis on the shapes of stories, “Fluctuations Between Good and Ill Fortune in Simple Tales”, was written in the mid-1960s. Chicago gave him the degree in 1971 and took his novel Cat’s Cradle in place of a thesis. Of the names for shapes, “Man in Hole”, “Boy Meets Girl” and “Cinderella” are Vonnegut’s. “Rags to riches”, “Icarus” and “Oedipus” come from Reagan and colleagues.
• Vogler. The memo is dated 1985 and the book followed in 1992. Its second edition says the model should be used as “a form, not a formula”. Dan Harmon’s story circle, which has eight steps, is a later shortening of the same journey.
• Booker. I read the book only in search snippets of a scanned copy, and the card also rests on two reviews of it.

Two older schemes are not in the gallery because they describe a hero’s life from his birth and not the plot of one story. Otto Rank (1909) compared the birth stories of fifteen heroes, and Lord Raglan (1936) listed twenty-two incidents, from “The hero’s mother is a royal virgin” to “He has one or more holy sepulchres”. Dundes’s remark about Campbell, quoted above, comes from an essay on these patterns.

The three acts of the screenwriting manuals have no card either. They are often credited to Aristotle, who said only that a whole has “a beginning, a middle, and an end”. Horace asked for five acts, a fourth-century commentary on Terence divided a comedy into a prologue and three further parts, and the three-act screenplay was set out by Syd Field in 1979. How far to trust the tension scores says where his third act begins.

Sources

• Aristotle. Poetics. Translated by S. H. Butcher (first published 1895). Project Gutenberg text. Chapters 7 to 11 and 18 were read, and checked against Bywater’s and Fyfe’s translations.
• Booker, C. (2004). The Seven Basic Plots: Why We Tell Stories. Continuum. Read only in search snippets of a lending scan, and described from two reviews, one of them by Adam Mars-Jones in The Observer, 21 November 2004.
• Campbell, J. (1949). The Hero with a Thousand Faces. Bollingen Series XVII. New York: Pantheon. Read only through search snippets of two lending scans at the Internet Archive (one of them, the second edition of 1968), so no page number here was checked and the book was not read through.
• Dundes, A. (1977). The hero pattern and the life of Jesus. First given as a paper in December 1976. As reprinted in In Quest of the Hero, 179–223. Princeton University Press, 1990. Read in search snippets of a lending scan.
• Eckermann, J. P. Conversations with Goethe, the entry for 14 February 1830, in John Oxenford’s English translation (1875), page 437. Scan at the Internet Archive, checked against the German text.
• Field, S. (1979). Screenplay: The Foundations of Screenwriting. New York: Dell. Not opened. The three acts are described from summaries of the book.
• Forster, E. M. (1927). Aspects of the Novel. Scan at the Internet Archive. The Clark lectures given at Trinity College, Cambridge, in spring 1927; chapter 5 was read.
• Freytag, G. (1863). Die Technik des Dramas. Leipzig: S. Hirzel. In English: Freytag’s Technique of the Drama, translated by Elias J. MacEwan from the sixth German edition. Chicago: S. C. Griggs, 1895. Scan of the second edition of the translation, 1896; German checked in the tenth edition. The passages quoted were found by searching the text of the scan, and the book was not read through.
• Harmon, D. Story structure 101. Channel 101 wiki, on a page created on 13 August 2009.
• Hinds, J. (1982). Contrastive rhetoric: Japanese and English. Paper given at the TESOL convention, Honolulu, May 1982 (ERIC ED 217 720). Scan at the Internet Archive. The passages quoted were found by searching the text of the scan, and the paper was not read through.
• Jiang Yin (1998). Qi cheng zhuan he: jixie jiegou lun de xiaozhang [Opening, continuation, turn and close: the rise and decline of a mechanical theory of structure]. Wenxue yichan 1998(3). In Chinese. Only the passage on where the scheme is first found was read, in a copy online.
• Kirkpatrick, A., & Xu, Z. (2012). Chinese Rhetoric and Writing: An Introduction for Language Teachers. WAC Clearinghouse and Parlor Press. The chapter “The Literary Background and Rhetorical Styles” was read in the LibreTexts copy.
• Labov, W., & Waletzky, J. (1967). Narrative analysis: oral versions of personal experience. In J. Helm (ed.), Essays on the Verbal and Visual Arts, 12–44. American Ethnological Society and University of Washington Press. Copy, read as page images.
• Labov, W. (1972). The transformation of experience in narrative syntax. In Language in the Inner City. University of Pennsylvania Press. Read in search snippets of a lending scan.
• Lévi-Strauss, C. (1960). La structure et la forme: réflexions sur un ouvrage de Vladimir Propp. In English in V. Propp, Theory and History of Folklore (University of Minnesota Press, 1984). Read in search snippets of a lending scan.
• Polti, G. (1895). Les trente-six situations dramatiques. Paris: Mercure de France. In English: The Thirty-Six Dramatic Situations, translated by Lucille Ray (1916). Scan of the 1921 printing. The passages quoted were found by searching the text of the scan, and the book was not read through.
• Propp, V. (1928). Morfologiia skazki. Leningrad: Academia. In English: Morphology of the Folktale, translated by Laurence Scott; second edition, revised and edited by Louis A. Wagner, with an introduction by Alan Dundes. Austin: University of Texas Press, 1968. Read in excerpts of the second edition, which hold Dundes’s introduction and the chapters on method and on the functions. One sentence from chapter IX (page 107) was read in another scan of the same edition. The rest of the book was not read.
• Quinn, S. F. (1993). How to write a Noh play: Zeami’s Sandō. Monumenta Nipponica 48(1): 53–88. Copy. Contains her annotated translation of Sandō (1423). Pages 58 to 65 were used, and the rest was not read through.
• Raglan, Lord. (1936). The Hero: A Study in Tradition, Myth and Drama. London: Methuen. The list of incidents and the scores were read in a scan of the 1949 Watts edition, found by searching its text. The book was not read through.
• Rank, O. (1909). Der Mythus von der Geburt des Helden. In English: The Myth of the Birth of the Hero, translated by F. Robbins and Smith Ely Jelliffe. New York: Journal of Nervous and Mental Disease Publishing Company, 1914. Scan of the translation. The passages used were found by searching the text of the scan, and the book was not read through.
• Reagan, A. J., Mitchell, L., Kiley, D., Danforth, C. M., & Dodds, P. S. (2016). The emotional arcs of stories are dominated by six basic shapes. EPJ Data Science 5: 31. Journal page; read in the arXiv copy.
• Segal, R. A. (1987). Joseph Campbell: An Introduction. New York: Garland. Read in search snippets of a lending scan.
• Shields, C. J. (2011). And So It Goes: Kurt Vonnegut, a Life. Henry Holt. Read in search snippets of a lending scan.
• The Daśarūpa: A Treatise on Hindu Dramaturgy. Translated by George C. O. Haas. New York: Columbia University Press, 1912. Scan at the Internet Archive. Page 9 was read as a page image.
• The Nāṭyaśāstra: A Treatise on Hindu Dramaturgy and Histrionics, ascribed to Bharata-muni. Translated by Manomohan Ghosh, vol. I. Calcutta: Royal Asiatic Society of Bengal, 1950. Chapter XXI and the translator’s introduction, as reproduced at wisdomlib.org.
• Todorov, T. (1977). The grammar of narrative. In The Poetics of Prose, translated by Richard Howard. Cornell University Press. The essay first appeared in French in 1968. Read in search snippets of a lending scan. His Grammaire du Décaméron (1969), which the essay summarises, was not opened.
• Toelken, B. (1996). The icebergs of folktale: misconception, misuse, abuse. In C. L. Birch & M. A. Heckler (eds), Who Says? Essays on Pivotal Issues in Contemporary Storytelling, 35–63. Little Rock: August House. Read in search snippets of a lending scan; the page of the passage quoted was not seen.
• University of Chicago Chronicle (3 February 1994). Notice on Kurt Vonnegut, with a second notice of 10 May 2007.
• Vogler, C. (1985). A practical guide to Joseph Campbell’s The Hero with a Thousand Faces. Memo of seven pages, written at Walt Disney Pictures. Copy with the author’s later preface.
• Vogler, C. (1998). The Writer’s Journey: Mythic Structure for Writers, second edition (first edition 1992). Read in search snippets of a lending scan.
• Vonnegut, K. (1981). Palm Sunday: An Autobiographical Collage. Chapter XVIII, “The Sexual Revolution”, read in a scan of the 1982 Dell paperback, pages 311 to 315.
• Vonnegut, K. (2005). A Man Without a Country. Seven Stories Press. The passage with the five drawings is excerpted as “At the Blackboard” in Lapham’s Quarterly, which is where it was read; it was checked against snippets of a scan of the book.

The readings were made in three studies between 3 and 12 October 2026. The first read the 34 classics for fortune and for the mood of the writing, and read 319 folk tales. The second read 1,031 texts with a longer form. The third read 1,033 texts for the hero’s journey. Forty of the first folk tales were read again in the second study, and for those the page uses the later reading.

The forms the readers were given are published as they were written: for the classics, for the first folk tales, for the second study and for the hero’s journey.

The readers

The first reader is a Claude model and the second is a different AI model. Which version of each was used was not recorded, so the reading cannot be repeated with the same models. The counts made from it can, from the readings listed under Data, forms and plans.

Every pack of texts went to a fresh copy of the reader, which saw that pack and nothing else. The second study had 110 packs of up to ten texts and the third had 124 packs of five to nine, with the kinds of text mixed. A reader could do two things: open a text, and write down its ratings. It was told that a rating once written was final. It was not told what kind of text it had. It was told that some of the texts are records and not stories, that some are disjointed or out of order and that some break off, and to rate each one as it stands.

A reader can tell a folk tale from a diary, so it could have found what it expected to find. Against that, the literary stories are stories and score low on several features, the last stretch of Captain Scott’s diary has all seven, and when forty tales had their sections scrambled the readers found a chain of cause in 5% of them.

The forms for the first folk tales and for the second study showed the layout of an answer with a sample filled in. The sample has a fall then a rise, a happy ending, survival as the aim, and events running from a problem in the first section to a reward in the last. The readers did not copy it. None of the 40 biographies and 4 of the 40 diary stretches were given that shape, and no literary story was given that aim. Among stories that have an opponent, a test or a trick, the reader first met it in the sample’s section 18%, 15% and 13% of the time, where one section in eight is 12.5%. The sample may still have nudged the readers towards the most common shape, which would make its 38% a little high, and towards the common order of events. The form for the hero’s journey has no sample filled in. These counts were made after the planned tests.

The form for the third study never mentions a hero, Campbell or Vogler, and tells the reader to set aside any scheme the questions bring to mind. It also tells the reader that most texts have only a few of the things asked about, and that many have none. It does describe the journey once, in an entry labelled “journey”. Before the questions it asks for an overall impression of how far the text is the story of someone who goes out from their own world into another, wins something there through danger and comes back with it. Of its 27 questions, 23 are about stages of the journey and 4 have nothing to do with it (a trick, a broken rule, a punishment, a hidden truth coming out). Those four had been asked in the second study too, so they show how far two readings of the same text agree. The questions came in a random order, not the order of the journey, and in two different random orders for alternate packs.

What was scored

Fortune is the main character’s good or ill fortune at the end of the section, as the text presents it at that point, from −5 (“catastrophe: death, ruin, utter despair”) to +5 (“bliss, triumph, everything they want”). The reader was told to rate the level and not the change, and to ignore what it knew came later.

Tension is “how much hangs in the balance at the end of the segment: how uncertain and how important the outcome feels right then”. The form anchors 0 as “nothing at stake (nothing has started, or everything is settled)”, 3 as “a real problem is open” and 5 as “the decisive moment, everything at stake”. So a text that settles its main question is meant to end at 0, and tension as scored here is tied by its wording to a problem being open.

Cause was scored for every section after the first. A section either “is caused by earlier events, or is someone’s response to them”, or “merely comes next”, or is unconnected to what came before.

In the third study each question was answered 2 if the thing is clearly in the text, 1 if the text could be read that way only at a stretch, and 0 if it is absent. In doubt the reader was told to take the lower answer. The page’s “clearly” is a 2 and its “arguably” is a 1.

Checks on the readers

A second model read 96 of the texts, twelve of each of eight kinds, with the same forms. Section by section its scores correlate with the first reader’s at 0.93 for tension and 0.91 for fortune, and 99% and 94% of its scores are within a point of the first reader’s. On judgements about the whole text the two agree less. They chose the same shape for 70% of the texts and the same ending for 78%. Every question is listed under The two readers, question by question.

Forty folk tales were read twice by different copies of the first reader, once as printed and once with the eight sections in a random order. A section got much the same scores wherever it stood (a correlation of 0.91 for tension and 0.96 for fortune). The reader also noticed the damage, finding a chain of cause in 98% of the tales as printed and 5% of the scrambled ones.

The same forty tales were also read cut short, by copies that saw only the first one to seven sections and so could not know the ending. Their scores correlate with those of the full reading at 0.90 for tension and 0.96 for fortune.

The reader was asked whether it recognised each text. It said it knew well 307 of the 591 stories, among them all 99 Hollywood films and 69 of the 280 folk tales. Among the 284 it did not know well, each of the seven features that pass the test in “What a diary does not have” is within 3.9 points of its share among all the stories, a check I added after the planned tests. What the reader says it knows is its own report, which I could not test.

Each reader was also asked to finish by saying what in the instructions it had not followed. In the second study 26 of 122 said something and in the third 41 of 138. Most of them (18 and 30) reported writing one rating and opening the next text in a single step where the form asked for two, with the rating written first. No rating was changed because of a report. One file in the third study failed the format check, and that text was read again by a fresh copy, as the plan provided.

What others have found about AI readers

Matlin and colleagues (2025) reran four experiments on suspense with language models in place of people. The models could tell a passage written to be suspenseful from one that was not, and their ratings moved in the same direction as people’s. They did not match people’s in size, reporting sometimes too much suspense and sometimes too little, and they were least reliable at the points where suspense turns from rising to falling.

Mitran, Wu and Piper (2025) had models and people score how far each of several moral qualities applied to the characters of folk tales. The models gave higher scores than the people did, which the authors call “a systematic bias toward over-identifying the presence of moral content”. Their questions covered vices as well as virtues. Every question in the forms used here asks whether something is present, so the readers may find more than a person would.

Subbiah and colleagues (2026, a preprint) found that models summarising life-story interviews “shift toward an overly positive framing of events”. Tian and colleagues (2024) point the same way for models as writers. In their sample, 28.8% of film synopses written by people follow an arc that ends on a fall, against 5.9% of the versions that GPT-4 wrote (my sums from their Figure 4). That is a finding about what a model writes and not about how it reads.

Tian and colleagues also report that models choose a story’s arc less reliably than people do, and shape is the question on which the two readers here matched least often. In the same paper the models were asked to find the sentence of a film synopsis at which each of five turning points falls. The paper gives a success rate of 46.6% for people, averaged over the five, and 31.3% and 30.6% for the two best models. Chang and colleagues (2023) found that models do better on books they have memorised, which is the reason for the check on known texts above.

Sui and colleagues (2026, a preprint) report that on a widely used test of story writing, model judges rank stories written by models above short stories from the New Yorker. Of tension they write that “model evaluators tend to reward surface markers of drama like urgent language or explicit emotional labels, rather than verifying that information is genuinely withheld across the narrative arc”. If the readers here did the same, a quiet story got lower scores for tension than it should have. Tension reaches 3 out of 5 in 82% of the literary stories, the lowest share among the kinds of story.

How well models agree with people depends on the task. On where one event of a story ends and the next begins, Michelmann and colleagues (2025) found that the boundaries GPT-3 chose were “closer to the consensus, on average, than boundaries identified by individual human annotators”. On who a story’s hero and villain are, Hobson and colleagues (2024) found that “GPT agreed with the majority opinion of human annotators 70.8% and 71.5% of the time”, about as often as the human annotators agreed with one another (75.5% and 69.7%). Their test set was 144 passages of five kinds, folk tales among them. On tagging the functions of Propp’s scheme in his own tales, Gervás and Méndez (2024) called the models’ agreement with Propp “surprisingly low”.

Agreement between two models is weaker evidence than agreement with people. Chhun, Suchanek and Clavel (2024) had models rate 1,056 short stories, most of them written by other programs. Each model was “remarkably self-consistent” from one run to the next, while its ratings agreed only moderately or weakly with people’s story by story. Averaged over all the stories from one source, the agreement was high, so the models ranked the sources much as people did.

Sources

• Chang, K. K., Cramer, M., Soni, S., & Bamman, D. (2023). Speak, memory: an archaeology of books known to ChatGPT/GPT-4. Proceedings of EMNLP 2023, 7312–7327. Read in part in the arXiv copy: the abstract, the description of the data and the passage on books out of copyright.
• Chhun, C., Suchanek, F. M., & Clavel, C. (2024). Do language models enjoy their own stories? Prompting large language models for automatic story evaluation. Transactions of the Association for Computational Linguistics 12: 1122–1142. arXiv copy. The abstract, the introduction and the description of the stories were read.
• Gervás, P., & Méndez, G. (2024). Tagging narrative with Propp’s character functions using large language models. Proceedings of Text2Story 2024, CEUR Workshop Proceedings 3671, 137–148. Paper.
• Hobson, D. G., Zhou, H., Ruths, D., & Piper, A. (2024). Story morals: surfacing value-driven narrative schemas using large language models. Proceedings of EMNLP 2024, 12998–13032. Paper.
• Matlin, G., Zhang, D., Barroso Loza, R., Popescu, D. M., Isbell, J., Chakraborty, C., & Riedl, M. (2025). Do language models agree with human perceptions of suspense in stories? arXiv:2508.15794; the arXiv record gives the conference COLM 2025.
• Michelmann, S., Kumar, M., Norman, K. A., & Toneva, M. (2025). Large language models can segment narrative events similarly to humans. Behavior Research Methods 57: 39. Paper. Abstract only, in the open copy.
• Mitran, L., Wu, S., & Piper, A. (2025). Probing narrative morals: a new character-focused MFT framework for use with large language models. Proceedings of EMNLP 2025, 28514–28529. Paper.
• Subbiah, M., Mian, H., Deas, N., Mayukha, A., McAdams, D. P., & McKeown, K. (2026). Whose story gets told? Positionality and bias in LLM summaries of life narratives. arXiv:2604.20131. A preprint.
• Sui, P., Zhu, Y., Cheng, T., West, P., So, R. J., Long, H., & Holtzman, A. (2026). Spoiler alert: narrative forecasting as a metric for tension in LLM storytelling. arXiv:2604.09854. A preprint. The abstract and the introduction were read.
• Tian, Y., Huang, T., Liu, M., Jiang, D., Spangher, A., Chen, M., May, J., & Peng, N. (2024). Are large language models capable of generating human-level narratives? Proceedings of EMNLP 2024, 17659–17681. Read in the arXiv copy.

Shapes

After the last section the reader named the shape that the main character’s fortune had followed. The choices were a rise, a fall, a fall then a rise, a rise then a fall, a rise, fall and rise, a fall, rise and fall, or none of these. The shares in the main text count those answers and are not worked out from the eight scores. The two readers chose the same shape for 70% of the texts they both read (a kappa of 0.64, on a scale where 1 is perfect agreement and 0 is what chance would give). They matched less often on this than on any other question with a fixed set of answers. On the 60 stories that both read, the first reader chose the fall then rise for 22 and the second for 20. The label is a judgement of the whole story, made separately from the eight scores, so the two can disagree.

The fall then rise is the top answer for folk tales (36%), ancient and medieval stories (42%), world films (43%), Hollywood films (51%) and literary stories (26%). For the 34 classics the top answer was “none of these”, and the fall then rise had 18%. The shares over all 904 stories were added up for this page. The studies reported each kind separately, and a pooled share depends on the mix, which here is 559 folk tales in 904.

A label covers many different lines. The readers called 342 of the 904 stories a fall then a rise, and 9 of the 904 stay within two points, in every section, of the Man in Hole line on the drawing board. That count was made for this page.

The names for the six shapes come from Reagan and colleagues, who found them by counting words. Here they are a reader’s judgement of what happens to a character, which is nearer to what Vonnegut drew. “Death comes last” compares the two methods on the classics.

I found one study in which people labelled whole stories this way. Tian and colleagues (2024) collected synopses of recent films, left out the well-known ones, and had three annotators label each of a sample of them with one of Reagan’s six shapes or a seventh, the double man in hole. In the synopses written by people the shares were man in hole 30.7%, double man in hole 21.0%, Cinderella 15.1%, riches to rags 14.6%, Oedipus 9.3%, Icarus 4.9% and rags to riches 4.4%. The paper does not give the number of synopses behind these shares; 440 narratives were labelled in all, counting the ones GPT-4 wrote. Man in hole comes first in their count as it does here, and in neither does it reach half.

Herrmann and Lüdtke (2023) had twenty students mark the feeling in every sentence of 80 of the Grimms’ tales, two of them for each tale. They report that “some of the observed plotlines can be matched onto the six basic shapes identified by Andrew Reagan and colleagues, others cannot”.

Endings

The reader also said whether the main character was better off, worse off or about the same “at the very end compared with the very beginning”. A happy ending in the main text is the answer “better”. The two readers gave the same answer for 78% of the texts they both read (a kappa of 0.67). The first reader, whose answers the page uses, is the stricter of the two. On the 60 stories that both read it found a happy ending in 28 and the second reader in 34.

The folk tales came in two batches. The first study read 319 tales, and the second read 240 more from thirty collections the first had not used. Setting the two batches side by side was a check I added after the planned tests of the second study had been run. The share of tales that end with the main character better off, by region:

The two batches rank the seven regions with a correlation of 0.32, which is weak. Europe is highest or second highest in both and Oceania is lowest or second lowest in both, and the regions between them change places. The first batch’s figure for Oceania was 40% when that correlation was worked out. It is 37% in the table because some of those tales were read a second time in the second study and this page uses the later reading.

The numbers in a cell are small, and one book can supply most of a cell, so a region’s figure is partly a figure for the few collections that were read. The chart of the two batches names the peoples behind each one.

To see whether a few books could have made the gap between regions, I dealt the seven regions at random among the 59 collections 20,000 times. The spread between regions came out as wide as the real one in 1 of those deals, and the gap between Europe and Oceania in 0. All 12 European collections have at least 60% of their tales ending well, and 4 of the 5 from Oceania have fewer than half. The American collections run from 20% to 88%. This check was made after the planned tests.

On Hollywood’s endings, David Bordwell (1982) called the happy ending “a fixture of Hollywood filmmaking of the classical period (1918 to about 1960)”, expected in comedies, musicals and romances and not in gangster films. His essay gives no counts, but a book he published three years later does. In Narration in the Fiction Film (1985) he reported that “of one hundred randomly sampled Hollywood films, over sixty ended with a display of the united romantic couple”, and that “many more could be said to end happily”. The hundred appear to be the sample of films made between 1915 and 1960 that he, Janet Staiger and Kristin Thompson drew for another book of the same year. Of the 99 Hollywood films here, 66% end with the main character better off.

Tales that explain an origin

The forms of both studies asked whether a tale explains how something in the world came to be, and the count over all 559 folk tales was made for this page. Of the 148 tales that explain an origin, 47% end with the main character better off, against 71% of the other 411. In Oceania 41 of 54 tales are of this sort, and in Europe 4 of 112.

How much goes with place

For the second study I split the variation in the fortune and tension lines of the 559 folk tales three ways. Each batch was first centred on its own average line, because the two batches were read in different studies with different forms. Differences between regions account for 4.7% of it, differences between collections of the same region for 11.1%, and differences between tales of the same collection for 84.2%. The first two are larger than they would be if the labels were dealt at random (1.7% and 9.0%, with p = 0.002 and p = 0.004). For which events a tale contains, the three shares are 6.3%, 14.7% and 78.9%.

A classifier trained on every collection but one, and asked for the region of the tales in the collection held out, was right 26% of the time from the two lines and 28% from the events, taking the average over the seven regions. With region labels shuffled it scored 15% and 14%. So a tale’s region shows faintly in its plot, and most of the variety is between one tale and the next in the same book.

Ross, Greenhill and Atkinson (2013) found a split of the same order in a different kind of data. They compared 700 versions of one folk tale across 31 European populations and put 9.1% of the variation between populations.

Who wrote the tales down

Most of the folk tales from outside Europe were collected, translated and often retold by outsiders. A few of the books were made by members of the people whose tales they are, among them Zitkala-Sa’s, Arthur C. Parker’s and Lal Behari Day’s. Jonathan Gottschall and his colleagues, quoted by Haase (2010), said of their own sample of ninety collections that “about 60% of the collections” were “collected, edited, translated, and in some cases retold by Westerners, usually males, between 1860 and 1930”. Haase quotes this to object that naming the problem does not answer it. Such texts, he writes, are “forms of intercultural communication”, and knowing where the collections come from “requires that we engage the texts individually and not simply in bulk”. This page reads them in bulk, so his objection applies to it. A collector who expected a tale to end well could choose and shape tales accordingly, and nothing in my data can separate that from what the tellers did.

Sources

• Bordwell, D. (1982). Happily ever after, part two. The Velvet Light Trap 19: 2–7. Scan on the author’s site.
• Bordwell, D. (1985). Narration in the Fiction Film. Madison: University of Wisconsin Press. One sentence was read in search snippets of a lending scan, and the page number was not seen. The sample of films is described in Bordwell, D., Staiger, J., & Thompson, K. (1985), The Classical Hollywood Cinema: Film Style and Mode of Production to 1960 (New York: Columbia University Press), a few sentences of which were read the same way.
• Haase, D. (2010). Decolonizing fairy-tale studies. Marvels & Tales 24(1): 17–38. Read in the open copy at the University of Hawaiʻi.
• Herrmann, B., & Lüdtke, J. (2023). A fairy tale gold standard: annotation and analysis of emotions in the Children’s and Household Tales by the Brothers Grimm. Zeitschrift für digitale Geisteswissenschaften 8. Article. The abstract and the passages on the design and on the shapes were read.
• Reagan, A. J., Mitchell, L., Kiley, D., Danforth, C. M., & Dodds, P. S. (2016). The emotional arcs of stories are dominated by six basic shapes. EPJ Data Science 5: 31. Journal page; read in the arXiv copy.
• Ross, R. M., Greenhill, S. J., & Atkinson, Q. D. (2013). Population structure and cultural geography of a folktale in Europe. Proceedings of the Royal Society B 280(1756): 20123065. Article.
• Tian, Y., Huang, T., Liu, M., Jiang, D., Spangher, A., Chen, M., May, J., & Peng, N. (2024). Are large language models capable of generating human-level narratives? Proceedings of EMNLP 2024, 17659–17681. Read in the arXiv copy.

The hero’s journey had a study to itself, the third. Its plan lists sixteen tests, all written down before the reading began, and says what result would count as showing the journey to be universal. The questions the readers answered are in the form.

What counts as the journey

Each of the 27 questions was answered “clearly there”, “arguably there” or “absent”, and the form recorded the section in which each thing first happens. Before the reading I defined four things in terms of these answers.

• Setting out. The main character clearly crosses into unfamiliar territory.
• The whole journey. The crossing, then trials or one great ordeal, then winning something of great worth, then coming home. Each must be clearly there, the crossing no later than the prize, and the prize no later than the return. The plan calls this the round trip.
• There and back, or there to stay. The same, except that settling in a new home can stand in for the return.
• Campbell’s sentence in full. The whole journey, into a world that is supernatural, with a prize that benefits other people. This follows the sentence of Campbell’s that is quoted in the note on the diagrams.

The plan also said that the journey would count as universal only if the whole journey were clearly there in 80% of folk tales, in 60% of the folk tales of every region and in 60% of every kind of story.

How common it is

“Stories” here are the 591 stories of the second study, of which 280 are folk tales. The 270 American films and the twelve famous films were read for other tests and are left out of these shares, because with them more than half of the sample would be films. The row that allows any order was worked out after the results were in.

Among the 240 folk tales of the second batch, the whole journey is clearly there in 11 of 32 tales from Europe, 8 of 32 from South Asia, 12 of 48 from the Americas, 6 of 24 from Oceania, 3 of 16 from West Asia, 7 of 40 from Africa and 7 of 48 from East Asia. Over all 280 folk tales the share runs from 16% in Africa (8 of 49) to 29% in Europe (12 of 41), and the main text’s “one in seven” and “three in ten” are from that count.

Setting out does not separate stories from other writing. Of the encyclopedia biographies, 73% have it, about as many as the folk tales, because the account of a life usually includes leaving home. On the loosest reading in the table, 22 of the 40 biographies make the journey, against 2 of the dreams and 1 of the diary stretches. The whole journey is +12 (+9 to +16) points more common in stories than in non-stories, with the 95% interval in brackets. None of the 23 stages asked about passes the test that seven features pass in “What a diary does not have”. Of the 23, 10 are at least 20 points more common in stories than in non-stories, but none is clearly there in 80% of folk tales. If arguable answers are counted, two stages pass, the call to adventure and the ordeal.

Whether the form can find a journey

A form that found few journeys might be bad at finding them. To check, I named nine films that were built on the scheme or are commonly taught as examples of it, and three that are not adventures but that Vogler’s book treats as journeys, and had them read with the same form. The plan said the form would pass if the nine had a median of at least nine of Vogler’s twelve stages and at least seven of them had eight or more.

The median of the nine is 9 and 7 have eight or more, so the form passes. Only 4 of the nine make the whole journey, and I looked up the reason for the other five after the results were in. Luke Skywalker does not go home. Harry Potter’s return is only arguable, since the summary ends with him at school and about to leave for the summer. The summary of Avatar that was read is 465 words long and has lost the first half of the film, so his setting out is only arguable. For that reason the wheel in the main text shows eight of the nine and leaves Avatar out. The Lion King and The Wizard of Oz fail on a rule of mine. The form records the first section in which each thing happens. Simba is back in the Pride Lands in section 6 and wins the kingdom in section 8, and Dorothy, who has run away, hurries back to the farmhouse in section 2, before the tornado, and wins her way home in section 7. Both returns therefore come before the prize. If the order is ignored, 6 of the nine make the journey, and on the loosest reading all nine do. The strict test misses films that everyone agrees are hero’s journeys, which is why the main text gives the loose figure beside it.

How far the answers can be trusted

The second model read 96 of the texts with the same form. The two readers agree on the whole journey with a kappa of 0.82 and on setting out with 0.81, and their counts of stages correlate at 0.91. The second reader found a little more, 3.8 stages a text against 3.3. Every question is listed under The two readers, question by question.

Four of the 27 questions have nothing to do with the journey and had already been asked of the same 591 stories in the second study, by other copies of the reader using another form. The two readings agree with kappas of 0.76 for a trick, 0.71 for a broken rule, 0.78 for a punishment and 0.72 for a secret coming out.

How common the journey looks depends on how the questions are put. The second study asked nothing about a journey, but its form recorded leaving home, a test or a showdown, a reward, and a return, which there could also mean settling somewhere new. A journey pieced together from those turns up in 53% of folk tales, whereas the third study’s questions find 31% on the nearest definition (there and back, or there to stay). Text by text, the pieced-together journey agrees with that definition with a kappa of only 0.49, and with the whole journey with a kappa of 0.29. “Leaves home” is a much lower bar than crossing into unfamiliar territory, and a reward is a lower bar than something of great worth.

The questions came in two random orders, used for alternate packs. The count of stages was the same under both (4.59 and 4.60). The whole journey was found in 11.4% of stories under one order and 16.3% under the other, a difference that chance could produce, though only just (−4.9 points, 95% interval −10.2 to +0.5). The reader’s overall impression did differ, 1.39 out of 4 under one order and 1.61 under the other (−0.21, −0.38 to −0.05). The figures on this page average over the two orders.

The reader of the third study said it knew 329 of the 591 stories well. (The reader of the second study, asked the same thing, said 307.) The whole journey is clearly there in 11.6% of those and in 16.8% of the 262 it did not know well. The two groups are different mixes of stories, since nearly all the Hollywood films are in the first and most of the folk tales in the second. Among folk tales alone it is 23% of the 84 it knew well and 20% of the 196 it did not. Recognising a story did not make the reader more likely to find the journey in it.

Order

Among 451 stories that have at least three of the twelve stages in two or more sections, the stages lean towards Vogler’s order in 95%. With the same sections dealt at random it would be 46%. The plan called this weak evidence in advance, because a character cannot come back before leaving, so part of the order is forced.

The forty folk tales that were read with their sections scrambled show that the reader was following the text. It found about as many stages in the scrambled tales (4.0 against 4.2), and it reported them in the order in which the scrambled text gave them. Their order leaned towards the scheme’s in 92% of the tales as printed and in 42% of the scrambled readings, which is what chance gives.

Vogler’s diagram puts the ordeal near the middle, and he calls it the crisis to keep it apart from the climax, which is his eleventh stage and belongs to the last act. Of the 285 stories with an ordeal, 67% have it in sections six to eight, and in 53% it falls in the section where tension is highest (30% by chance). The question asked for one supreme crisis on which everything hangs, so in most stories the readers will have marked what Vogler calls the climax.

Whether the stages go together

The main text says that the four legs of the journey go together. The planned comparison deals each leg out separately among the stories of one kind, so that it allows for folk tales having more of every leg than literary stories do. On that basis 11.0% of stories would have all four, and 19.5% do. This test was added to the plan before the reading began.

Some of that excess belongs to stories in which a lot happens, which have more of everything. The four questions that have nothing to do with the journey also turn up together more often than the same comparison expects (7.1% of stories against 4.4%), and when the length of the text is allowed for as well, 12.6% of stories would have all four legs. The legs do go with one another more than they go with those four unrelated events. Inside each kind of story the average correlation between two legs is 0.22, and between a leg and an unrelated event it is 0.06 (a difference of 0.16, interval 0.10 to 0.21). That leaves out setting out and coming back, a pair that is partly forced. The same holds, more weakly, in the second study (0.15 against 0.06), whose form asked nothing about a journey, so it does not come only from the third form having asked for an overall impression first. The counts in this paragraph were made after the planned tests.

Of the stories that make the whole journey, 70% have all seven of the features that pass in “What a diary does not have”, against 52% of the rest.

Who makes the journey

Among stories with one human main character, 290 are about a man or boy and 86 about a woman or girl. The men’s stories have 5.2 of the twelve stages on average and the women’s 4.4, a difference of 0.86 (95% interval 0.28 to 1.43). Inside each kind of story the difference is 0.64 (0.09 to 1.18), a figure worked out after the results were in. The whole journey itself is clearly there in 18% of the men’s stories and 13% of the women’s, a difference that chance could produce (+5 points, interval −4 to +13).

Kathleen Ragan (2009) drew 1,601 published folk tales at random and set the sex of each main character beside the sex of its teller, collector and editor. I know the paper only from its abstract, which concludes that analyses “that have not taken gender into account in the compilation of their data sets can be considered compromised”. I did not record who told or collected each tale, so that applies to the counts here.

The journey’s images in the motif catalogue

Before the reading I chose ten images that Campbell names, going by the titles of motifs in Berezkin’s catalogue alone, and counted them in its 958 societies. One of the ten I already knew to be widespread, the tasks set by in-laws, which stands here for the road of trials. The most widespread are the boon (39% of societies), the road of trials (37%) and the belly of the whale (36%). Of the ten, 7 are recorded on all six inhabited continents, and none is recorded in half of the societies, though the whole family of 71 motifs about imposed tasks is recorded in 55%. In the 81 societies with the fullest records, 7 of the ten are recorded in at least half. The note on the catalogue explains why its figures are low.

Other counts

I found no published study that takes a defined sample of stories and counts how many follow Campbell’s or Vogler’s stages, so I have nothing to set these figures against directly. The nearest counts are of other hero patterns. Lord Raglan (1936) scored twenty-one heroes of myth and legend on his own list of twenty-two incidents, from 21 for Oedipus down to 9 for Elijah. In 1965 Victor Cook tried the same list on heroes from other parts of the world and found, as Alan Dundes reports it, that the pattern “did not apply, generally speaking, to selected American Indian, African, or Oceanic heroes”. Dundes’s own conclusion was that the evidence supports a pattern for Indo-European and Semitic heroes.

Clyde Kluckhohn (1959) searched the myths of fifty societies chosen to cover the world and found the slaying of a monster in thirty-seven of them. He also reported that in the forty-eight hero myths that Rank and Raglan had used, the hero marries his mother in only four. Ian Jobling (2001) looked for one plot, a hero who defeats a half-human ogre, in a random sample of 44 cultures, and a summary of his paper puts it in 31 of them. Kluckhohn and Jobling count the peoples who tell some story of a hero, and my figures count the stories that have the stages, so the two kinds of figure do not conflict.

The folklorists quoted in the note on the diagrams objected that Campbell built the pattern from the stories that fit it. In a sample chosen without regard to the journey, the whole of it is clearly there in 21% of folk tales, and several of its stages are in more than half of them.

Sources

• Berezkin, Y. E., & Duvakin, E. N. Thematic classification and areal distribution of folklore and mythological motifs: an analytical catalogue. Online, in Russian. Only the front page and the first motif page were opened; the catalogue was used through the data of Michalopoulos and Xue (2021).
• Campbell, J. (1949). The Hero with a Thousand Faces. Bollingen Series XVII. New York: Pantheon. Read only through search snippets of two lending scans at the Internet Archive (one of them, the second edition of 1968), so no page number here was checked and the book was not read through.
• Cook, V. (1965). Lord Raglan’s hero: a cross cultural critique. The Florida Anthropologist 18: 147–154. Not opened. Described from Dundes’s account of it.
• Dundes, A. (1977). The hero pattern and the life of Jesus. First given as a paper in December 1976. As reprinted in In Quest of the Hero, 179–223. Princeton University Press, 1990. Read in search snippets of a lending scan.
• Jobling, I. (2001). The psychological foundations of the hero-ogre story: a cross-cultural study. Human Nature 12(3): 247–272. Publisher’s page. Abstract only. The count of 31 cultures in 44 is from the summary of the paper in the Human Relations Area Files.
• Kluckhohn, C. (1959). Recurrent themes in myths and mythmaking. Daedalus 88(2): 268–279. JSTOR. Read in a copy of the scan.
• Ragan, K. (2009). What happened to the heroines in folktales? An analysis by gender of a multicultural sample of published folktales collected from storytellers. Marvels & Tales 23(2). Record. Abstract only.
• Raglan, Lord. (1936). The Hero: A Study in Tradition, Myth and Drama. London: Methuen. The list of incidents and the scores were read in a scan of the 1949 Watts edition, found by searching its text. The book was not read through.
• Toelken, B. (1996). The icebergs of folktale: misconception, misuse, abuse. In C. L. Birch & M. A. Heckler (eds), Who Says? Essays on Pivotal Issues in Contemporary Storytelling, 35–63. Little Rock: August House. Read in search snippets of a lending scan; the page of the passage quoted was not seen.
• Vogler, C. (1985). A practical guide to Joseph Campbell’s The Hero with a Thousand Faces. Memo of seven pages, written at Walt Disney Pictures. Copy with the author’s later preface.
• Vogler, C. (1998). The Writer’s Journey: Mythic Structure for Writers, second edition (first edition 1992). Read in search snippets of a lending scan.

Where the films come from

Three sets of films were read, all as plot summaries.

The 108 films from around the world and the 270 American films by era come from the CMU Movie Summary Corpus (Bamman, O’Connor and Smith 2013), which holds 42,306 plot summaries taken from the English Wikipedia in November 2012. The 99 Hollywood films are the TRIPOD set (Papalampidi, Keller and Lapata 2019), whose synopses have five turning points marked by people. The five are those of a scheme that Michael Hauge teaches to screenwriters: an opportunity, a change of plans, a point of no return, a major setback and a climax. The twelve famous films are CMU summaries too.

A summary is written by someone who has seen the film and knows how it ends, and a writer who knows the hero’s journey may well tell the film that way. The readers therefore read the film as retold on Wikipedia.

The test by era

For the era test I drew films credited to the United States alone, with a summary of 400 to 1,500 words. Longer summaries have room for more stages, so the films were drawn evenly by length. There are three eras, 1930 to 1976, 1977 to 1992 and 1993 to 2012, and three bands of length, and 30 films were taken at random for each combination, which gives 90 to an era. The boundaries are the release of Star Wars in 1977 and the publication of Vogler’s book in 1992.

The plan named one test, the difference in the average number of Vogler’s twelve stages between the last era and the first. The three eras average 4.17, 4.84 and 4.87. The difference is +0.70 of a stage, with a 95% interval of +0.04 to +1.36 and p = 0.039. (The plan’s 2,000 shuffles gave p = 0.031. I reran the test with 200,000 shuffles after finding that the figure moved from one run to the next.) Campbell’s seventeen stages rise a little more surely, from 3.4 to 4.2 (a difference of +0.79, interval +0.18 to +1.37). The other measures point the same way with less certainty. The reader’s own impression that a film is mostly or wholly a journey goes from 7% of the films of the first era to 17% and 17% (+10 points, interval +1 to +19). The whole journey is clearly there in 8, 9 and 15 of the 90 films in each era, a rise of +8 (−2 to +18) points, which chance could produce. When arguable answers are counted, the twelve stages go from 8.1 to 8.6 (+0.46, interval −0.22 to +1.10), which chance could also produce.

The plan also said the result would be reported for two kinds of film. A film counts as an adventure film if the corpus labels it adventure, fantasy, action, science fiction or family. I fixed that list by looking at the labels before any film had been read. There are 122 such films and 148 others, and the adventure films are 46%, 49% and 41% of the three eras, so the rise does not come from Hollywood making more of them. Among them the count of stages rises by +1.76 between the first era and the last (+0.77 to +2.72). Among the rest it changes by +0.01 (−0.76 to +0.76). Those two intervals were worked out after the results were in.

Most of the rise among adventure films comes between the first era and the second (+1.39 of a stage, interval +0.37 to +2.39), and after 1992 there is no further rise that can be told from chance (+0.37, interval −0.61 to +1.34). That does not separate Star Wars from the manuals. Vogler’s memo dates from 1985, and 34 of the 44 adventure films of the second era were released in 1985 or later. What rose is the count of stages. The whole journey is clearly there in 7 of 41, 9 of 44 and 10 of 37 of the adventure films of the three eras, a rise that chance could produce (+10 points, interval −8 to +28). The rise in stages among adventure films is larger than the change among the rest by 1.75 (0.54 to 2.96) of a stage.

Over all the films the rise depends on the length of the summary. Among the shortest summaries the count was 3.8 in the first era and 3.7 in the last. Among the medium ones it went from 4.3 to 5.6, and among the longest from 4.4 to 5.3. Among the adventure films the rise is there at every length (3.7 to 5.3, 4.9 to 6.5 and 4.9 to 7.0), though each of those figures rests on 10 to 17 films. The bands of length were in the plan. I set the eras side by side within each band, and made the counts in this paragraph and the one before it, after the results were in.

What Lucas and Vogler said

Lucas told Bill Moyers in 1999, “when I did ‘Star Wars’ I consciously set about to recreate myths and the — and the classic mythological motifs.” The remark about Yoda was made at the National Arts Club in 1985, where he said of Campbell, “He is a really wonderful man and he has become my Yoda.”

Vogler described his memo, in a preface he added later, as “the original seven-page memo”, and said that he wrote it “when I was working as a story consultant for Walt Disney Pictures”. The copy I read carries a copyright line of 1985. That the journey changes the hero is in the memo too. Its eleventh stage has the hero “transformed by his/her experience”. The memo also allows that “the stages can be deleted, added to, and drastically re-shuffled without losing any of their power”, which is why the main text gives a generous count beside the strict one.

Change, force and what is deserved

The figures in the second paragraph of this section come from the second study’s form and its 99 Hollywood films. The form asked whether the main character is a different person inside by the end, what settles the main problem, and whether the ending gives the characters what their conduct deserved. The two readers agree on these with kappas of 0.59, 0.80 and 0.73, so the question about inner change is the least reliable of the three.

Sources

• Bamman, D., O’Connor, B., & Smith, N. A. (2013). Learning latent personas of film characters. Proceedings of ACL 2013, 352–361. Paper. The source of the CMU Movie Summary Corpus.
• Cousineau, P. (ed.) (1990). The Hero’s Journey: Joseph Campbell on His Life and Work. Harper & Row. One remark by George Lucas, made at the National Arts Club in 1985, was read in a search snippet of a lending scan.
• Moyers, B. (1999). The mythology of Star Wars with George Lucas. Transcript dated 18 June 1999.
• Papalampidi, P., Keller, F., & Lapata, M. (2019). Movie plot analysis via turning point identification. Proceedings of EMNLP-IJCNLP 2019, 1707–1717. Paper. The source of the 99 Hollywood films with turning points marked by people.
• Vogler, C. (1985). A practical guide to Joseph Campbell’s The Hero with a Thousand Faces. Memo of seven pages, written at Walt Disney Pictures. Copy with the author’s later preface.
• Vogler, C. (1998). The Writer’s Journey: Mythic Structure for Writers, second edition (first edition 1992). Read in search snippets of a lending scan.

What it is

Yuri Berezkin is an anthropologist and folklorist at the Museum of Anthropology and Ethnography (the Kunstkamera) of the Russian Academy of Sciences. His catalogue, which names E. N. Duvakin as second author, records which motifs of myth and folk tale have been written down for which peoples. The economists Stelios Michalopoulos and Melanie Meng Xue released a version of it as data with their 2021 paper “Folklore”, and that version, with 2,564 motifs and 958 societies, is the one counted here. They call it his “lifetime work”. Papers that use other versions of the catalogue give other totals.

A society here is what Berezkin calls a tradition. Usually it is the speakers of one language. Where the record is thin, as in New Guinea and Australia, many peoples are taken together, and Australia is eight regions on the map. Fifteen of the 958 are written traditions of the past, among them Sumer, the Old Testament, Ancient Egypt and Ancient Greece. Michalopoulos and Xue used the version of October 2019 and count 6,239 books and articles behind it. The catalogue has gone on growing, and in December 2025 it stood at 3,328 motifs and 1,031 traditions.

No reader is involved in these counts. They are counts of entries in a table of societies and motifs.

What it leaves out

The catalogue was built as a tool for working out how peoples moved and met in prehistory, and for that purpose a motif told everywhere is no use. Berezkin (2015) writes that “Those elements that are truly found universally have simply been ignored because their study is useless for our purposes.” So the catalogue cannot show that a piece of story content is universal, and the low shares on this page are partly a result of that choice. It can show how sharing falls with distance.

Michalopoulos and Xue state two more limits. The catalogue “does not code images found in a single oral tradition, but focuses on motifs shared by at least two groups”, so nothing in it is unique to one people, and it “does not account for the popularity of a motif or the number of alternative versions of a particular tale”, so a motif told once counts the same as one told in every village.

A blank in the catalogue also does not mean that a motif was never told. A motif is entered for a society when someone recorded it and the record reached the catalogue. Societies differ greatly in how much was written down, so every share on this page is a lower limit. That is why the main text also gives the figures for the best-recorded societies.

The motifs are also narrow by design. A catalogue made to trace how motifs spread enters alternatives separately. A trickster who is a fox, a jackal or a coyote is one motif, a trickster who is a hare or a rabbit is another, and there are more for the turtle, toad or frog and for the raven. One of those four is recorded in 58% of societies (559 of 958). The share is 94% in Europe and 5% in Oceania. The sun and the moon are entered three ways, as a male sun with a female moon, as both male, and as a female sun. One of the three is recorded in 69% of societies and in at least 49% on every continent. Something told nearly everywhere in one variant or another does not show up in the catalogue as a widespread motif. These two counts were added after the planned tests.

The counts

The five counts were listed in the plan of the second study. The counts for the best-recorded societies and the second measure of sharing were added while the code was being written, as checks on how far gaps in the record decide the answer.

• The most widespread motif is recorded in 37% of societies. None is recorded in half, 16 are recorded in a quarter and 164 in a tenth. Of the 2,564 motifs, 640 are recorded in fewer than one society in a hundred. The median motif is recorded in 19 societies by my count, and Michalopoulos and Xue give 18.
• There are 182 motifs recorded on all six inhabited continents, and 14 recorded in at least a tenth of the societies of each.
• Of all the motifs, 43% are recorded both in the Americas and in the Old World. For Berezkin’s two halves of the catalogue the figure is 60% for cosmology and origins and 31% for adventures and tricksters. Tales of how the world began are shared across the oceans about twice as often as tales of what heroes and tricksters did.
• In the 282 societies with at least 100 motifs recorded, 10 motifs are found in at least half. In the 81 societies with at least 200, 143 are, and the trickster who is a fox, a jackal or a coyote is found in all but 1. Of those 81 societies, 31 are in Europe, 45 in Asia and 4 in North Africa, and the other is the Western Ojibwa of North America.

Sharing and distance

For every pair among the 568 societies with at least 50 motifs recorded (161,028 pairs), I counted the motifs the two share as a fraction of all the motifs that either has. That fraction averages 20% for pairs within 250 km of each other and 4% for pairs more than 8,000 km apart. It is 9.5% for two societies on the same continent and 4.7% for two on different ones.

That measure is pulled down when one society has a short list. Counting instead the share of the shorter list that is also on the longer one gives 44% for the nearest pairs and 11% for the farthest, so the fall with distance is there on either measure. Neighbours are also more likely to have been written down by the same collectors, which would make their lists alike for another reason, and I could not test that.

How this fits other work

Folklorists have two older indexes. Stith Thompson’s Motif-Index of Folk-Literature (1955–1958) lists motifs, and the index of tale types that Antti Aarne began in 1910 and Hans-Jörg Uther last revised in 2004 lists whole plots. Berezkin (2015) built his own catalogue because, in his words, “both are Eurocentric”, and of the index of types he says that “Australia and America are completely beyond its scope”.

Sharing that falls off with distance is what borrowing and common descent would produce, and both have been shown for particular tales. Tehrani (2013) traced the versions of Little Red Riding Hood as a family tree. Da Silva and Tehrani (2016) found that 100 of the 275 types of magic tale in the index of types, and 76 once borrowing between neighbours is allowed for, show signs of having been passed down along the branches of the Indo-European languages. Thuillard and colleagues (2018), working from another version of Berezkin’s catalogue, concluded that “world mythologies are structured in geographical patterns”.

Broader themes are more widespread than any single motif. In 1959 Clyde Kluckhohn and Richard Moench looked for a few themes in the mythologies of fifty cultures, one from each of fifty culture areas, and found a flood in 34, the slaying of monsters in 37, incest in 39 and rivalry between siblings in 32. A theme as wide as “a flood” covers many of the catalogue’s motifs, so the two counts do not conflict.

Michalopoulos and Xue also had crowd workers judge the 87 motifs whose descriptions mention a challenge or a contest. The character succeeds in 48% of them and is harmed in 19%. Those two figures are printed as images in the article and were read from it by eye.

Sources

• Berezkin, Y. E., & Duvakin, E. N. Thematic classification and areal distribution of folklore and mythological motifs: an analytical catalogue. Online, in Russian. Only the front page and the first motif page were opened; the catalogue was used through the data of Michalopoulos and Xue (2021).
• Berezkin, Y. E. (2015). Folklore and mythology catalogue: its lay-out and potential for research. RMN Newsletter 10: 58–70. The issue. Only the passages quoted were read.
• da Silva, S. G., & Tehrani, J. J. (2016). Comparative phylogenetic analyses uncover the ancient roots of Indo-European folktales. Royal Society Open Science 3(1): 150645. Article.
• Kluckhohn, C. (1959). Recurrent themes in myths and mythmaking. Daedalus 88(2): 268–279. JSTOR. Read in a copy of the scan.
• Michalopoulos, S., & Xue, M. M. (2021). Folklore. Quarterly Journal of Economics 136(4): 1993–2046. Article; the data, which is the version of Berezkin’s catalogue used here.
• Tehrani, J. J. (2013). The phylogeny of Little Red Riding Hood. PLoS ONE 8(11): e78871. Article.
• Thompson, S. (1955–1958). Motif-Index of Folk-Literature, volumes I–VI. FF Communications 106–109, 116–117. Helsinki: Academia Scientiarum Fennica. Not opened. Details from Berezkin (2015).
• Thuillard, M., Le Quellec, J.-L., d’Huy, J., & Berezkin, Y. (2018). A large-scale study of world myths. Trames 22(4): 407–424. Article.
• Uther, H.-J. (2004). The Types of International Folktales, volumes I–III. FF Communications 284–286. Helsinki: Academia Scientiarum Fennica. Not opened. Details from Berezkin (2015).

Tension was scored from 0 to 5 at the end of each of a text’s eight sections. The note on the reading gives the form’s wording.

The tests, and the shuffle

The plan fixed three tests on tension. The first is that the end is calmer than the peak. The second is that tension builds and then drops, which means that the peak is above the opening score and the last score is below the peak. The third is that the peak comes late, which means in the second half, taking the middle of the tied sections when several share the top score.

All three are easy to pass by accident with eight numbers, so each was compared with the same eight scores in shuffled order. Among the folk tales tension builds and then drops in 93%, and in 59% when the scores are shuffled.

The late peak is the weakest of the three. On the planned definition the peak is in the second half for 72% of stories and 44% of non-stories. Six of the non-stories have the same tension in every section. The planned test counts them as not late, and in the chart of the section in which tension is highest each gives an eighth of itself to every section. By kind it is 64% of folk tales, 64% of ancient and medieval stories, 81% of world films, 88% of Hollywood films and 84% of literary stories. It did not pass the test in “What a diary does not have” because it is not common enough in folk tales. A looser count, made for this page, takes a story if any of its top-scoring sections is the sixth or the seventh. That gives 75% of stories and 58% of non-stories. The bars under the chart in the main text give each story one peak and share it equally among the sections that tie for its top score. On that count the sixth and seventh sections hold 57% of the stories’ peaks, and the last four sections hold 74% of them, against 52% of the non-stories’.

What the reader adds

Two checks used forty folk tales from the first batch.

In the first, other copies of the reader were given the tales with their eight sections in a random order. A section got nearly the same tension score wherever it stood (a correlation of 0.91). But in the scrambled tales tension still built and then dropped 80% of the time. Moving the as-printed scores into the scrambled order gives 68%, and shuffled scores give 55%. So the reader supplies some of the pattern itself. The section it read last got an average of 2.0, against 2.3 for the same section in its proper place, and it was lower in 12 tales and higher in 3. I looked at this after the planned tests were run. Part of the final drop in the chart is therefore the reader responding to a text that is ending. Part is the scale, which gives 0 to a text in which “everything is settled”. In the 496 stories whose main question the reader called settled, the last score is 0 in 94%. In the 95 where it was partly settled or left open, the last score is 0 in 11%. I counted that for this page.

In the second, copies of the reader saw only the first one to seven sections of a tale, so each score was given without knowledge of what came next. Those scores correlate with the full reading’s at 0.90. Tension builds and then drops in 95% of the tales read this way and 98% read whole. The late peak holds up less well. It is there in 55% of the tales read without hindsight, which cannot be told apart from the 48% that shuffled scores give (p = 0.42), and in 63% of the same tales read whole.

Against turning points marked by people

The 99 Hollywood films come from a set in which people marked five turning points in each synopsis (Papalampidi, Keller and Lapata 2019). This is the only place where the readings meet a human judgement. The plan named three comparisons. One is clearly above chance, one weakly, and one is no better than chance.

• The reader’s tension peak is in the section of the marked climax, or the one beside it, in 70% of films. Pairing peaks and climaxes at random gives 62%.
• The lowest point of the main character’s fortune is at the marked major setback, or beside it, in 69% of films, against 62% at random (p = 0.024).
• The biggest single fall in fortune ends at the major setback, or beside it, in 48% of films, against 47% at random. On the planned measure this one fails. In the very same section it is 29% against 19%, which is above chance (p = 0.002).

The tension peak is in the very same section as the climax in only 6% of films, fewer than the 16% that chance gives. The score is taken at the end of a section, and by the end of the climax’s own section the matter is settled. Tension averages 3.9 in the section before the climax and 0.8 in the climax’s own. People put the climax 89% of the way through a synopsis on average.

In a pilot the paper’s authors picked one sentence for each turning point, “under the assumption that all TPs are present”, and the annotators’ instructions were drawn up from that pilot. So the marks show where people put a turning point and cannot show that a film lacks one. On the five films that two trained annotators both marked, they chose the same sentence for 64% of the turning points. The five turning points are those of the scheme for screenwriters that the note on the films describes. Of the 99 films, 77 were marked by one set of annotators and the other 22 by three sets.

The second reader

The two readers’ tension scores correlate at 0.93, and 99% are within a point of each other. They agree less on the yes-or-no feature built from the scores. On whether tension builds and then drops their kappa is 0.52, and by the second reader’s scores the gap between stories and non-stories on that feature is 16 points (88% against 72%), which is under the 20 the test asks for. The note on the seven features says more.

How high it goes

After the first results I added one count, the share of texts in which tension reaches 3 out of 5 at some point. It is 97% of folk tales, 91% of ancient and medieval stories, 97% of world films, 98% of Hollywood films and 82% of literary stories, against 55% of dreams, 23% of diaries and 25% of biographies. Tension both reaches 3 and then drops in 90% of stories and 28% of non-stories. That separates the two far better than the shape alone, but it was not a planned test.

What has been measured with people

I found no published set of human tension ratings on anything like this scale. The largest set of human marks I found is of a different kind. Vijayaraghavan and Roy (2023) had people highlight the climax in 2,382 personal stories posted on Reddit, and the average position was 61% of the way through. That is earlier than the readers here put the peak, though an anecdote told online is a different kind of text from a folk tale or a film. The studies of tension itself are small. Those that bear on the outline here agree with it, and the others show how much two people can differ.

• Wilmot and Keller (2020) had annotators mark, sentence by sentence, whether suspense rose or fell in 200 short stories. Rises were about three times as common as falls (38% of sentences against 12%, my sums), which fits a long climb and a short drop.
• Lehne and colleagues (2015) had 23 people rate suspense through one story by E. T. A. Hoffmann. Agreement between readers was r = .31.
• Bentz and colleagues (2024) had 34 people draw their suspense as a line while reading one Conan Doyle story, and found “a high variability of the suspense lines between participants”, though each reader’s line correlated 0.58 on average with the others’.
• Schmidt and colleagues (2023) played three stories to 31 listeners, each of whom heard two, while recording heart and skin measures, and report “sympathovagal double activation in the climax unit” of all three.
• Brewer and Ohtsuka had 48 American students each rate one American and one Hungarian short story, from six of each, part by part, so that seven to nine students rated each story. Suspense went with how much a story was liked (a correlation of .42), and the best-liked story was Shirley Jackson’s “The Lottery”, which ends with a stoning. I read the study as a technical report of 1986, and the article of 1988 may differ.
• Brewer and Lichtenstein (1982) describe an unpublished study in which readers rated narratives taken from popular magazines and from avant-garde writers. Almost all of the magazine pieces showed the rise and release of suspense, surprise or curiosity that their theory predicts, and almost none of the avant-garde ones did.
• Screenwriting manuals put the climax late. Syd Field’s Screenplay (1979) divides a film into three acts, the set-up, the confrontation and the resolution, with the third beginning about three quarters of the way through, and Blake Snyder’s beat sheet (2005) starts the finale on page 85 of 110. I know both from summaries and have not opened either book. James Cutting (2016) measured shot lengths, motion, brightness and other physical features of 150 English-language films released between 1935 and 2010 and concluded that they divide into four acts (set-up, complication, development and climax), each about a quarter of the running time. That scheme is Kristin Thompson’s, and it cuts Field’s second act in two. It puts the climax in the last quarter of a film. The readers here most often put the highest tension in the sixth or seventh of eight sections, which overlaps with it.
• Matlin and colleagues (2025) reran four experiments on suspense with language models in place of people. The models could tell a passage written to be suspenseful from one that was not, and here the same readers scored the non-stories much lower. Their average never rises above 1.3, and the stories’ reaches 3.1. The models’ ratings were often too high or too low, and they were least reliable at the points where suspense turns from rising to falling, which is a reason to trust the height of the lines here more than the place of the peak.

Sources

• Bentz, M., Cortez Espinoza, M., Simeonova, V., Köppe, T., & Onea, E. (2024). Measuring suspense in real time: a new experimental methodology. Scientific Study of Literature 13(1): 1–21. Article.
• Brewer, W. F., & Lichtenstein, E. H. (1982). Stories are to entertain: a structural-affect theory of stories. Journal of Pragmatics 6(5–6): 473–486. Read as Technical Report No. 265, Center for the Study of Reading, October 1982; the article may differ.
• Brewer, W. F., & Ohtsuka, K. (1988). Story structure, characterization, just world organization, and reader affect in American and Hungarian short stories. Poetics 17(4–5): 395–415. Read in an earlier version, Technical Report No. 375, March 1986; the article may differ.
• Cutting, J. E. (2016). Narrative theory and the dynamics of popular movies. Psychonomic Bulletin & Review 23(6): 1713–1743. Article.
• Field, S. (1979). Screenplay: The Foundations of Screenwriting. New York: Dell. Not opened. The three acts are described from summaries of the book.
• Lehne, M., Engel, P., Rohrmeier, M., Menninghaus, W., Jacobs, A. M., & Koelsch, S. (2015). Reading a suspenseful literary text activates brain areas related to social cognition and predictive inference. PLOS ONE 10(5): e0124550. Article.
• Matlin, G., Zhang, D., Barroso Loza, R., Popescu, D. M., Isbell, J., Chakraborty, C., & Riedl, M. (2025). Do language models agree with human perceptions of suspense in stories? arXiv:2508.15794; the arXiv record gives the conference COLM 2025.
• Papalampidi, P., Keller, F., & Lapata, M. (2019). Movie plot analysis via turning point identification. Proceedings of EMNLP-IJCNLP 2019, 1707–1717. Paper. The source of the 99 Hollywood films with turning points marked by people.
• Schmidt, M.-L. C. R., Winkler, J. R., Appel, M., & Richter, T. (2023). Tracking emotional shifts during story reception: the relationship between narrative structure and affective responses. Scientific Study of Literature 12(1): 17–39. Article.
• Snyder, B. (2005). Save the Cat! The Last Book on Screenwriting You’ll Ever Need. Studio City: Michael Wiese Productions. Not opened. The page numbers of the beat sheet are from summaries of it.
• Thompson, K. (1999). Storytelling in the New Hollywood: Understanding Classical Narrative Technique. Cambridge, MA: Harvard University Press. Not opened. Cited from Cutting (2016).
• Vijayaraghavan, P., & Roy, D. (2023). M-SENSE: modeling narrative structure in short personal narratives using protagonist’s mental representations. arXiv:2302.09418. A preprint. The abstract, the section on the data and the first two tables were read.
• Wilmot, D., & Keller, F. (2020). Modelling suspense in short stories as uncertainty reduction over neural representation. Proceedings of ACL 2020, 1763–1788. Read in the arXiv copy.

The rule

The rule was fixed in the plan of the second study before any text was read. A feature passes if it meets two conditions.

• It is present in at least 80% of the folk tales and at least 60% of every other kind of story, which here means the ancient and medieval stories, the world films, the Hollywood films and the literary short stories.
• It is at least 20 points rarer among the non-stories, taken together, than among the stories, and the 95% interval for that gap does not include zero.

The plan said that a feature meeting the first condition and not the second would be reported as common to stories and non-stories, and that a feature failing only in the literary stories would be reported as that. The literary stories were included as a hard case, because several of these writers set out to do without plot.

Eight of the twenty features had come out of the first study’s folk tales. The other twelve were new.

All twenty

Stories are the 591 of the second study and non-stories the 120. Gaps are in percentage points and were worked out before rounding, so a few differ by a point from the two columns beside them.

One feature only just fails. Ending calmer than the peak is 19.96 points more common in stories, and the bar was 20. I have kept to the rule and left it out. Nothing is lost by it, because every text in which tension builds and then drops also ends calmer than its peak.

Four features fail only in the literary short stories: an ending that is clearly good or clearly bad (39% of them), a main question settled by the end (50%), a repeated episode (42%) and stakes from the list of five (42%).

Two features are much more common in the non-stories. A diary or a dream nearly always follows one person, and usually has few characters.

The rule sets the stories against the non-stories taken together. Taken one kind at a time, four of the seven keep a gap of 20 points against the dreams, the diaries and the biographies alike. The other three do not. Tension builds and then drops in 78% of the dreams and 80% of the biographies, and passes because of the diaries (40%). An ending at the best or the worst point is found in 25 of the 40 diaries, and one line of action in 26 of the 40 biographies. The diaries are stretches cut from nine diaries at an entry picked at random, so they begin and end where the cut fell.

The second reader

The second model read 60 of the stories and 36 of the non-stories. On those 96 texts:

By the second reader’s answers 6 of the seven keep a gap of 20 points or more. The one that does not is tension that builds and then drops, with a gap of 16. Three of the seven have a kappa under 0.6, the line for good agreement used in the note on the two readers. They are tension that builds and then drops, a problem that sets things going, and an ending at the best or the worst point. Most of the disagreement is about the non-stories. Among the 60 stories that both readers scored, each of the three has a kappa of 0.63 or more. The second reader also found a problem that sets things going in 67% of the non-stories, whereas the first found one in 31%, so the two understood that question differently. The chain of cause is the feature on which they agree best, and it has the widest gap for both.

Texts the reader knew

A reader that recognises a famous story might answer from memory of what stories are like. Among the 284 stories the reader did not say it knew well, the seven features are present in 96%, 90%, 94%, 84%, 97%, 96% and 87%, in the order of the table above. None is more than 3.9 points from its share among all the stories. I added this check after the planned tests.

Together

The counts of stories with all seven, and with the four named in the answer at the end of the main text, were made for this page. The four are the ones most common among stories, and I picked them after the results were in. Any other four of the seven are found together in 58% to 83% of stories. The tension is the least sure of the four, and without it the other three are found together in 90% of stories and 11% of non-stories. By the second reader’s answers, all four are present in 51 of the 60 stories it read and 5 of the 36 non-stories, and by the first reader’s in 52 and 6. Length makes little difference. With the stories of each kind sorted into a short, a middle and a long third, all four are found in between 80% and 86% of the stories in each third. Among the folk tales the figure is 86% for the 69 that the reader said it knew well and 86% for the other 211.

All seven are present in 65% of folk tales, 47% of ancient and medieval stories, 42% of world films, 46% of Hollywood films and 45% of literary stories. A story has 6.3 of the seven on average and a non-story 3.1.

What makes a text read as a story

The reader also gave every text a score from 0 to 4 for how much it read as a story. The plan called for a regression of that score on the eight older features and on cause, closure and aim, over all 711 texts. Together they account for 72% of the variation. With the others held fixed, a chain of cause is worth 1.37 points (95% interval 1.11 to 1.62), a problem 0.50 (0.24 to 0.75) and an aim 0.43 (0.22 to 0.66). The score and the features come from the same reader, so this shows what the reader’s idea of a story rests on.

The order of events

The form listed eighteen things that can happen in a story and recorded the section in which each first happens. Fourteen of them have a place in time, and those are the ones in the chart. Only the first occurrence is recorded, so a helper who turns up twice is placed where they first appear. Part of the common order comes from that. A problem or an opponent runs through many sections and is first met early, however the tale is told.

The plan had two tests. In the first, each region’s ordering of the events was compared with the ordering in all the other regions put together. The lowest of the seven correlations is 0.90, for Oceania. In the second, a tale counts as in order if its own events lean towards the common order, and it needs at least three events in at least two sections to be counted. That holds for 100% of the 240 folk tales of the second batch, against 48% when the positions are dealt at random.

Only 5 of the 40 diaries and 21 of the 40 dreams have enough of these events to be counted. The dreams that have enough of them keep the order nearly as often as stories do, which is why the main text says that the order may not be special to stories. The biographies keep it less often.

By hand, Propp’s scheme has been tried on other peoples’ tales since the 1960s. Alan Dundes wrote in his introduction to the English edition that “attempts to study African tales (Paulme) and American Indian tales (Dundes 1964b) suggest that parts of Propp’s Morphology may be cross-culturally valid”. Pannach and colleagues (2021) marked his functions in twenty African and fifteen Indian tales and noted that two analysts of the same tale can assign them differently. By machine, his order has been tested only on his own Russian tales, as far as I could find. Finlayson (2016) had a program learn Propp’s functions from fifteen of the tales Propp used, and it recovered eight of fourteen groups of functions. Gervás and Méndez (2024) asked two language models to tag the functions in synopses of eleven of his tales and called the agreement with Propp “surprisingly low”, with precision between 0.37 and 0.46. The fourteen events here are much coarser than Propp’s thirty-one functions, so the order found here is not a test of his scheme.

What others have found

The oldest claims of this kind are the Danish folklorist Axel Olrik’s “epic laws” of 1908. Among them are that a folk narrative keeps to a single strand of plot and to one leading character, repeats its episodes, sets two opposed figures against each other, and moves from calm to excitement and back to calm before it ends. Several of the twenty features above are close to those laws, and they do not all come out alike. In the folk tales here the single strand is in 94%, the repeated episode in 87%, the calm ending in 100% and the pair of opposed figures in 21%. Olrik argued from his reading and gave no counts. I have not read his essay, and I give the laws as they are usually summarised.

Brewer and Lichtenstein’s experiment is the nearest thing to this test with human readers. They wrote three sets of events in six versions each and gave them to 103 undergraduates at the University of Illinois, twenty to a version. Readers rated how far each text was a story on a scale of seven, where 3 was “barely a story”. The base versions averaged 2.35 to 2.90. Versions with suspense and a resolution averaged 4.90 to 5.10, and versions with suspense and no resolution 2.60 to 3.20. The authors concluded, “The base narratives were clear cohesive prose, but the subjects’ story ratings showed that the subjects did not think these narratives were stories.” The base versions had a character with an aim and events that followed from one another. So the experiment supports a problem that sets things going, and it shows that aim and cause are not enough by themselves. It supports tension that builds and then drops only in part. In another version the reader was not told of the bomb until the car blew up, so there was surprise and no suspense, and that was rated a story too (3.90 to 4.70). The authors’ own account is that a story needs suspense, surprise or curiosity, followed by a resolution. One of their three significant events was a winning sweepstakes ticket, so what mattered was that something was at stake, which is wider than a problem. In a later paper (1982) they describe an unpublished study in which texts with suspense and no character pursuing a goal were still called stories by most readers. The readers were students at one American university in 1980 and the texts were written for the experiment. I read the study as a technical report of December 1980, which may differ from the printed chapter.

Piper and Bagga (2024) point the other way on one feature. They put questions about narrative features to a language model for passages from eighteen kinds of writing, and found that conflict “is also present within non-narrative communication”. They add that conflict was among the features on which the model agreed least with their annotators, and that the annotators agreed little with one another about it. Their passages were short, and their narrative class includes biographies and memoirs, so the contrast is not the one drawn here.

On dreams, Hendrickx and colleagues (2017) compared 6,998 dream reports with personal stories and found that the dreams had “about 20% fewer discourse markers”, with the exception of “then”, which was more common. That matches the widest gap here, which is cause. Cipolli and Poli (1992), which I know from its abstract, concluded that “Story-like organization seems to be a feature of dream production”. Dreams are the non-stories that come nearest to stories here too. Tension builds and then drops in 78% of them, and three of the four non-stories with all seven features are dreams.

Two experiments from around 1980 asked whether people of different cultures share a scheme for stories, and Brewer (1984), reviewing them, wrote that “they arrive at opposite conclusions”. Kintsch and Greene (1978) had students in Colorado summarise and recall European stories and Native American ones. The European ones were remembered better, and the authors concluded that such schemes belong to a culture. Brewer called their study “an incomplete experimental design”, since every reader came from one culture. The design goes back to Frederic Bartlett. In Remembering (1932) his English readers retold a Native American tale, “The War of the Ghosts”, and in retelling it they dropped what was strange to them and made the rest more like the stories they knew. Mandler and colleagues (1980) had Liberian and American listeners recall four Western folk tales and one Liberian tale, found the two groups much alike, and concluded that a scheme is shared. A later study with children from two cultures who were matched for schooling found that they “differ significantly in the structural components of their written story retellings” (Invernizzi and Abouzeid 1995, which I know from its abstract). Brewer’s own view was that what is universal is abstract, and he wrote that “the underlying use of plan schemas to understand human actions must be a universal”. I know the first experiment only from Brewer’s account, and the second from its abstract and from Brewer.

Steg, Slot and Pianzola (2022) had seven people rate passages for suspense, curiosity and surprise, trained a model on those ratings and used it to score 17,372 texts. Biography and history came out lower (.43 and .37) than fairy tales and short stories (.54 and .56). These are a model’s predictions and the gap is modest, but it runs the same way as the one here.

Manvir Singh (2021) argues that a plot with “a goal-directed protagonist who confronts obstacles, overcomes them, and wins rewards” is “ubiquitous”. The aim and the obstacles agree with what passed here, but the reward does not, since the share of stories that end well varies as much as “No single shape” describes.

Sources

• Bartlett, F. C. (1932). Remembering: A Study in Experimental and Social Psychology. Cambridge University Press. Not opened. The experiment with “The War of the Ghosts” is described as textbooks give it.
• Brewer, W. F., & Lichtenstein, E. H. (1981). Event schemas, story schemas, and story grammars. In J. Long & A. Baddeley (eds), Attention and Performance IX, 363–379. Hillsdale, NJ: Erlbaum. Read as Technical Report No. 197, Center for the Study of Reading, University of Illinois at Urbana-Champaign, December 1980; the printed chapter may differ.
• Brewer, W. F., & Lichtenstein, E. H. (1982). Stories are to entertain: a structural-affect theory of stories. Journal of Pragmatics 6(5–6): 473–486. Read as Technical Report No. 265, Center for the Study of Reading, October 1982; the article may differ.
• Brewer, W. F. (1985). The story schema: universal and culture-specific properties. In D. R. Olson, N. Torrance & A. Hildyard (eds), Literacy, Language, and Learning: The Nature and Consequences of Reading and Writing, 167–194. Cambridge University Press. Read as Technical Report No. 322, Center for the Study of Reading, University of Illinois at Urbana-Champaign, August 1984; the printed chapter may differ.
• Cipolli, C., & Poli, D. (1992). Story structure in verbal reports of mental sleep experience after awakening in REM sleep. Sleep 15(2): 133–142. Record. Abstract only.
• Domhoff, G. W., & Schneider, A. (2008). Studying dream content using the archive and search engine on DreamBank.net. Consciousness and Cognition 17(4): 1238–1247. Authors’ reprint. Abstract only. The source of the dream reports.
• Dundes, A. (1968). Introduction to the second edition of Propp’s Morphology of the Folktale, xi–xvii. Austin: University of Texas Press. Read in the excerpts linked under Propp.
• Finlayson, M. A. (2016). Inferring Propp’s functions from semantically annotated text. Journal of American Folklore 129(511): 55–77. Author’s copy.
• Forster, E. M. (1927). Aspects of the Novel. Scan at the Internet Archive. The Clark lectures given at Trinity College, Cambridge, in spring 1927; chapter 5 was read.
• Gervás, P., & Méndez, G. (2024). Tagging narrative with Propp’s character functions using large language models. Proceedings of Text2Story 2024, CEUR Workshop Proceedings 3671, 137–148. Paper.
• Hendrickx, I., Onrust, L., Kunneman, F., Hürriyetoğlu, A., Stoop, W., & van den Bosch, A. (2017). Unraveling reported dreams with text analytics. Digital Humanities Quarterly 11(4). Read in the arXiv copy of 2016, which may differ.
• Invernizzi, M. A., & Abouzeid, M. P. (1995). One story map does not fit all: a cross-cultural analysis of children’s written story retellings. Journal of Narrative and Life History 5(1): 1–19. Record. Abstract only.
• Kintsch, W., & Greene, E. (1978). The role of culture-specific schemata in the comprehension and recall of stories. Discourse Processes 1(1): 1–13. Record. Not opened; described from Brewer (1984).
• Mandler, J. M., Scribner, S., Cole, M., & DeForest, M. (1980). Cross-cultural invariance in story recall. Child Development 51(1): 19–26. Record. Abstract only, with Brewer’s (1984) account of it.
• Olrik, A. (1908). Episke love i folkedigtningen. Danske Studier 5: 69–89. A German version followed in 1909. In English: “Epic laws of folk narrative”, in A. Dundes (ed.), The Study of Folklore, 129–141 (1965). Not opened. The laws are given as they are usually summarised.
• Pannach, F., Sporleder, C., May, W., Krishnan, A., & Sewchurran, A. (2021). Of lions and Yakshis: ontology-based narrative structure modelling for culturally diverse folktales. Semantic Web 12(2): 219–239. Article. The abstract and the passage on the two analyses of one tale were read.
• Piper, A., & Bagga, S. (2024). Using large language models for understanding narrative discourse. Proceedings of the 6th Workshop on Narrative Understanding, 37–46. Paper.
• Propp, V. (1928). Morfologiia skazki. Leningrad: Academia. In English: Morphology of the Folktale, translated by Laurence Scott; second edition, revised and edited by Louis A. Wagner, with an introduction by Alan Dundes. Austin: University of Texas Press, 1968. Read in excerpts of the second edition, which hold Dundes’s introduction and the chapters on method and on the functions. One sentence from chapter IX (page 107) was read in another scan of the same edition. The rest of the book was not read.
• Singh, M. (2021). The sympathetic plot, its psychological origins, and implications for the evolution of fiction. Emotion Review 13(3): 183–198. Read in the accepted manuscript, which may differ from the printed article.
• Steg, M., Slot, K., & Pianzola, F. (2022). Computational detection of narrativity: a comparison using textual features and reader response. Proceedings of LaTeCH-CLfL 2022, 105–114. Paper.

The six shapes, rerun

Reagan and colleagues worked with “a filtered subset of 1,327 stories from Project Gutenberg’s fiction collection”. They scored the words with a list of about ten thousand English words that people had rated for happiness (Dodds and colleagues 2011), averaged the scores in a window sliding through each book, and broke the resulting curves down into basic patterns. The first three patterns and their mirror images are the six shapes. The paper also sorted the curves into groups by two other methods, and reports that all three support the six.

My rerun is the first of the three studies. Of the 1,327 books, 1,326 could still be downloaded and 1,325 could be analysed. With the paper’s settings the rerun gives the same patterns as the paper’s figures, and the first three carry 70% of the variation between the curves.

Part of the pattern comes from the smoothing. Books whose words had been shuffled gave the same three patterns, carrying 59% of the variation, and so did random noise (60%). Reagan and colleagues had made the first of these comparisons themselves, with shuffled books and with machine-made nonsense, and reported that the patterns there were flatter and the groups less distinct. Any long series that is smoothed with a wide window breaks down into slow waves of this kind. The share the first three carry also depends on the width of the window. It is 12% with a window of 500 words and 88% with 20,000. Real books do show the patterns more strongly than shuffled ones, as the paper said, and that difference is what the count shows about stories. The first study’s page gives the details.

Others had made the point about smoothing. In 2015 Matthew Jockers reported “six, or possibly seven, archetypal plot shapes” in 41,383 novels, using a related method, and gave up his smoothing filter after Annie Swafford, Ben Schmidt and Scott Enderle showed what it did to the curves. In 2016 Scott Enderle got the same patterns out of random walks. The published paper says that a random walk is the wrong comparison, since it is not a text, and that the test should be a book with its words shuffled. Reagan’s thesis (2017) gives Enderle the same answer by name.

Others have looked for the shapes in other material. Del Vecchio and colleagues (2018) sorted 6,174 film scripts into anything from four to twelve groups, judged six the best fit, and found that films with a fall and then a rise earned most at the box office. Maharjan and colleagues (2018) found two shapes in 1,003 books. Wu, Wang and Mihalcea (2023) found five in novels in eight languages, none of them a single fall and rise.

Mood and fortune

The 34 classics were read section by section by an AI reader who scored the main character’s fortune and, separately, the mood of the writing. The word list’s curve follows the reader’s mood score with a median correlation of 0.52, and the fortune score with 0.31. So the word list is a fair measure of the mood of the vocabulary and a poor one of how the main character is doing. Boot and colleagues (2024) had asked of the six-shapes paper whose feeling its curve follows, the reader’s, the characters’ or the narrator’s.

Twenty-one thousand books

The larger count took every English text that Project Gutenberg’s catalogue files under language and literature and classes as fiction, with at least 10,000 words, which came to 21,529 books. The mood of the vocabulary falls from start to finish, measured as a straight-line trend, in 69% of them and in 50% of copies with the words shuffled. Among the 7,215 novels that could be cut cleanly from the first chapter heading to “The End”, it falls in 74%. On the last few pages the mood usually lifts again. The final fiftieth of the book is brighter than the tenth before it in 69% of the fiction and 51% of the shuffled copies.

To see where in a book each kind of word falls, I counted 17 groups of words through a hundred equal slices of every one of the 7,215 novels. Fourteen of the groups are lists by theme, such as home and family, fear and danger, death, and tears, and three are groups of small grammatical words. Words for home and family run 30% above a book’s own average in its first tenth. In the last tenth, words for death run 39% above, words for love and marriage 39% above and words for tears 30% above. These are leanings, and the strongest of them, the early home and family words, is found in 63% of the novels and the late words for death in 61%. The chart in the main text is drawn in twentieths, so the figures at its two ends are larger than these.

The lists are mine, and nobody else has checked that the words on them mean what I took them to mean in context. “Grave” is on the death list and is also an adjective.

What others have counted

Where words fall in a novel had been counted before. David McClure (2017) cut each of 27,266 American novels into a hundred equal slices and counted where every word falls. He found that “death” leans towards the end, and summed up the 500 most uneven words by saying that beginnings are about “youth, education, physical size, (good) appearance, color, property, hair, and noses” and endings about “forgiveness, criminal justice, suffering, joy, murder, marriage, arms, and hands”. My 17 groups of words are a coarser count than his, made on a different collection of books and then tried in the nine other languages below.

Other studies have counted different things through a story. Boyd, Blackburn and Pennebaker (2020) cut about 40,000 narratives into five parts and counted kinds of small words. Articles and prepositions, which set a scene, start high and fall. Pronouns and auxiliary verbs, which carry the action, start low and rise. Words for thinking a thing through are most common in the middle. About 22,000 texts that are not stories (newspaper science articles, talks and court opinions) showed the first two patterns as well (the court opinions only the second), so only the third belongs to stories, and all three effects are small.

Toubia, Berger and Eliashberg (2021) measured how fast and how widely a text moves through a space of word meanings in 4,118 films, 12,336 television episodes and 29,300 academic papers. They found links with success that are weak for films and television and that differ between the three. Knight, Rocklage and Bart (2024) counted reversals of mood in about 30,000 films, television episodes, novels and fundraising pitches, and found that more reversals and larger ones go with success. They write that the six shapes “did not predict success”. Reagan and colleagues had reported that three of their shapes were downloaded more often than the rest, so the two papers do not agree on this.

Morin, Acerbi and Sobchuk (2019) are a caution about the late words for death. They set 811 novels beside 10,810 private letters and diary entries written by American women and found that “Private letters and diaries are as likely, or more likely, to use words relating to natural or agentive death.” So talk of death is something novels share with diaries, and my count adds only where in a novel it falls.

Nine other languages

Word lists had been translated and counted through novels in other languages before. Wu, Wang and Mihalcea (2023) translated an English set of lists automatically into seven languages and counted them through Project Gutenberg fiction. They report that the first two of Boyd’s patterns appear in most of the languages and the third only weakly.

The plan of the second study made ten predictions from the English counts and tested each in every other language in which Project Gutenberg has at least a hundred usable novels. Those are Finnish, French, German, Dutch, Spanish, Italian, Hungarian, Chinese and Portuguese, with 4,557 novels between them. Swedish, with 84, fell short. The word lists were my own translations of nine of the English lists, written before anything was counted, and no speaker of those languages has checked them. The mood lists exist for five of the languages (French, German, Spanish, Chinese and Portuguese; Dodds and colleagues 2015), so the last two predictions could be tested only in those.

A prediction counts as holding in a language when the 95% interval over its books lies wholly on the predicted side. “Unclear” means the interval includes zero. No interval lay wholly on the wrong side.

The last column is a check added after the first results. The Chinese sample turned out to hold eight parts of one novel and twelve second copies of books already in it, so I dropped single volumes of longer works and second copies from every language, which left 4,008 novels, and counted again.

Originals and translations were not told apart, so a French novel in the sample may be a translation from English. Project Gutenberg holds books that are out of copyright, so these are old novels, and eight of the nine languages are European. A count of words also shows only when a book talks about death, which is some way from what happens in the plot.

Sources

• Boot, P., Daza, A., Schnober, C., & van Hage, W. (2024). In the context of narrative, we never properly defined the concept of valence. Computational Humanities Research 2024, CEUR Workshop Proceedings 3834. Paper. The passages on Reagan and colleagues were read.
• Boyd, R. L., Blackburn, K. G., & Pennebaker, J. W. (2020). The narrative arc: revealing core narrative structures through text analysis. Science Advances 6(32): eaba2196. Article.
• Del Vecchio, M., Kharlamov, A., Parry, G., & Pogrebna, G. (2018). The data science of Hollywood: using emotional arcs of movies to drive business model innovation in entertainment industries. arXiv:1807.02221. A revised version appeared in Journal of the Operational Research Society 72(5): 1110–1137 (2021), which was not read.
• Dodds, P. S., Clark, E. M., Desu, S., and others (2015). Human language reveals a universal positivity bias. Proceedings of the National Academy of Sciences 112(8): 2389–2394. arXiv copy. Abstract only. The source of the mood lists in the five other languages.
• Dodds, P. S., Harris, K. D., Kloumann, I. M., Bliss, C. A., & Danforth, C. M. (2011). Temporal patterns of happiness and information in a global social network: hedonometrics and Twitter. PLoS ONE 6(12): e26752. Article. The source of the word list.
• Enderle, S. (2016). A plot of Brownian noise. Notebook, chosen by Digital Humanities Now on 8 September 2016.
• Jockers, M. L. (2015). The rest of the story. Blog post, 25 February 2015, followed by Requiem for a low pass filter, 6 April 2015.
• Knight, S., Rocklage, M. D., & Bart, Y. (2024). Narrative reversals and story success. Science Advances 10(34): eadl2013. Article.
• Maharjan, S., Kar, S., Montes-y-Gómez, M., González, F. A., & Solorio, T. (2018). Letting emotions flow: success prediction by modeling the flow of emotions in books. Proceedings of NAACL-HLT 2018, volume 2: 259–265. Read in the arXiv copy.
• McClure, D. (2017). Distributions of words across narrative time in 27,266 novels. Techne, Stanford Literary Lab, 10 July 2017. Post.
• Morin, O., Acerbi, A., & Sobchuk, O. (2019). Why people die in novels: testing the ordeal simulation hypothesis. Palgrave Communications 5: 62. Article.
• Reagan, A. J. (2017). Towards a science of human stories: using sentiment analysis and emotional arcs to understand the building blocks of complex social systems. PhD thesis, University of Vermont. arXiv:1712.06163. Pages 37 to 39 were read.
• Reagan, A. J., Mitchell, L., Kiley, D., Danforth, C. M., & Dodds, P. S. (2016). The emotional arcs of stories are dominated by six basic shapes. EPJ Data Science 5: 31. Journal page; read in the arXiv copy.
• Swafford, A. (2015). Problems with the Syuzhet package. Blog post, 2 March 2015.
• The first of the three studies behind this page, which reruns Reagan and colleagues’ analysis and tests it against shuffled text, is written up at shapes-of-stories.netlify.app.
• Toubia, O., Berger, J., & Eliashberg, J. (2021). How quantifying the shape of stories predicts their success. Proceedings of the National Academy of Sciences 118(26): e2011695118. Article.
• Wu, W., Wang, L., & Mihalcea, R. (2023). Word category arcs in literature across languages and genres. Proceedings of the 5th Workshop on Narrative Understanding, 36–47. Paper. The abstract and sections 3 to 5.1 were read.

This ranking was made for this page after every test had been run. It is an illustration, and no finding depends on it.

For each of the 591 stories I took 35 things the second study’s form had recorded. Of these, 8 are questions with named answers, such as the shape of the main character’s fortune, how they end up, what they want and what settles the matter. Another 12 are yes-or-no features, such as whether there is a chain of cause or a single line of action. The other 15 are events that either happen or do not, such as a trick, a helper, a showdown or a reward. For every question I found the answer that the largest number of stories give. A story’s score is the number of questions on which it gives that answer. Stories with the same score are ordered by how unusual their other answers are.

“The Strange Banana Skin” scores 34. The one question on which it differs is whether a secret comes out, which happens in 73% of stories and not in this one. The next 5 stories score 33. They are “The Magician and the Sultan’s Son” (Swahili), “Younger Brother Eludes his Sister-in-Law by Creating Obstacles and Liberates Older Brother” (Seneca), “The Field of Bones” (Bengali), “The Maiden who Found the Moon” (Aboriginal Australian) and the film Indiana Jones and the Kingdom of the Crystal Skull. So the first place is held by a single answer, and a second reading of the same tale could change it.

The tale as it was read is William Westervelt’s English version of 1915. He opens by summing up the whole plot, and he likens the bird’s valley to one in the Arabian Nights, so some of the fit may come from his way of telling it. One of the five that follow, “The Maiden who Found the Moon”, comes from a book written for children.

The median score is 25. Of the thirty highest-scoring stories, 20 are folk tales and 9 are films. Folk tales are 280 of the 591, so the most common answers are largely theirs, and a sample with a different mix would have a different most typical story.

The most common answer is not always a common one. The top answer for what the main character wants is “to survive”, given by 24% of stories. For what settles the matter it is help from someone, with 26%, and for shape it is the fall then rise, with 41%. A story can therefore give all 35 most common answers and still be unlike most stories on any one of these. The choice of questions is mine as well. Events make up 15 of them, which favours stories in which a great deal happens.

What others mean by a universal

The main text asks whether any pattern fits most stories. Patrick Colm Hogan, who has written a book on narrative universals, asks whether a pattern turns up in one tradition after another, and writes that “the point of a cross-cultural pattern is that it recurs across traditions, not that it characterizes every work”. His account of the typical story is close to what was found here. It “begins by establishing a particular goal for a hero”, who is then “frustrated in his or her goal pursuit”. There is often “exile or some related removal from home”, a middle in which the hero “seems furthest away from achieving his or her goal”, and most often “a comic resolution”. He calls this “a prototype, not a set of necessary and sufficient conditions”. His evidence is wide reading with no count, and I know his book of 2003 only from his own summary of it (Hogan 2016).

Meir Sternberg’s answer is different again. He holds that the universals of narrative are three interests that any telling plays on, “suspense, curiosity, and surprise” (Sternberg 2003), which puts what stories share in how they hold a reader and not in what happens in them.

Sources

• Hogan, P. C. (2003). The Mind and Its Stories: Narrative Universals and Human Emotion. Cambridge University Press. Record. Not opened. Known from the author’s summary of 2016.
• Hogan, P. C. (2016). Story. Literary Universals Project, University of Connecticut, 20 November 2016. Article.
• Sternberg, M. (2003). Universals of narrative and their cognitivist fortunes (I). Poetics Today 24(2): 297–395. Record. Pages 327 and 328 were read.

A result means more if the test was written down before the data were seen. This note says which results on the page are of that kind.

The first study, 3 October 2026

The first study had no plan. It began as a rerun of Reagan and colleagues’ paper, went on to the larger collection of fiction, and then had AI readers read the 34 classics and 319 folk tales. The tests it applied to the folk tales, and the thresholds for words like “usually”, were chosen after those readings were in.

The second study was designed to repair that. It took the first study’s twelve conclusions about folk tales as predictions and tested them, with the same definitions and thresholds, on 240 tales from thirty collections the first study had not used. Eleven of the twelve came out the same. The one that changed is a low point of fortune late in the tale, which the first study put at no more than chance and the second found to lean weakly that way.

The second study, 11 October 2026

The plan was saved before any text of the sample had been read. It lists the texts and tests A to J, and for A, C, F12 and I what would count as passing. Some sections were added later the same morning, still before the reading, and each is dated in the plan. They are the test of reading without hindsight, the final wording of the form, details of sampling, the rule for which part of a book in another language is counted, the fixed sample, details of the analysis, and one change after a pilot. The pilot used six texts that are not in the sample, and showed that the question about repeated episodes was being read too loosely, so its wording was tightened.

The plan ends with a list headed “Settled after the data were in”. Its main entries are these.

• With 397 of the first reader’s 1,031 readings in, and all 96 of the second reader’s, the scripts were run once, to check that they worked on real ratings and so that the report could be drafted while the reading went on. No test, definition or threshold was changed as a result, and no number from that run is used. Two things were settled at that stage, before the full readings were analysed. The words for chance, which the plan had fixed for test A, were also used for tests B, E, G, H and K. And the scrambled tales of test D were summed up by counting the measures on which the scrambled reading lies nearer to content alone than to the tale as printed.
• All 1,127 rating files passed the format check. None was read again, changed or left out.
• A check was added to the test on other languages after its results were known, because the Chinese sample turned out to contain several parts of one novel and second copies of others. The planned result is the one in the main text.
• After the full analysis had been seen, six checks were added. They are the seven features among stories the reader did not know well, the scrambled tales taken one by one, the guesses at endings split by whether the reader knew the tale, happy endings by region in the two batches, how high tension goes, and each feature in each kind of non-story. The plan gives the reason, which is that the planned results had raised those doubts.
• The whole analysis was run a second time to store more decimal places. All 2,010 counts and verdicts were identical.

The third study, 12 October 2026

The plan was written before any text had been read with the new form. Four sections were added about twenty minutes later, after a pilot and still before the reading. One of them is the sixteenth test, on whether the stages go together.

The plan records one lapse. While the analysis script was being tried on made-up ratings it printed one real figure, the share of folk tales with a journey pieced together from the second study’s readings. That was seen after the plan was saved and before any new reading, and no definition was changed.

Added after the results were in, and marked as such wherever they appear in these notes:

• the whole journey with the order ignored, and with each further loosening;
• the reason each of five well-known films fails the strict test;
• the intervals for adventure films and other films taken separately, and the differences by length of summary;
• the gap between men’s and women’s stories worked out inside each kind of story.

One rating file failed the format check. As the plan provided, it was set aside and the text was read again by a fresh copy of the reader.

Counts made for this page

These were made while this page was being written, when every result was already known. None of them was a planned test.

• The shares of the six shapes over all 904 stories, and the share of happy endings by region over all 559 folk tales. The studies reported each batch and each kind separately.
• The table and chart of the two batches of folk tales, which use the later reading for forty tales that were read twice.
• The count of tales that explain an origin over all 559 folk tales.
• The share of folk tales that make the whole journey, region by region, over all 280 tales that were asked, and the share of Man in Hole among the folk tales of each region.
• The shares of stories whose main character changes and whose ending is deserved, by kind of story.
• Where the 81 best-recorded societies of the motif catalogue are, and how many societies have at least one of several motifs that differ only in a detail.
• The average tension at each section, the share of stories whose highest tension is in the sixth or seventh section, and how often the last score is 0 in stories that settle their main matter and in those that do not.
• The shares of stories with all seven features, with six, and with the four named in the answer, by each reader’s own answers.
• The pairs of events in the chart of order, such as how often a helper comes before the showdown.
• The ranking of stories by how typical they are.
• The drawing board. It reads the line a visitor draws at the end of each of the eight sections, compares those eight values with each story’s eight fortune scores, and lists the stories for which the root-mean-square difference is smallest. A story counts as staying close when it is within two points of the line at all eight places. The classics were cut into more than eight sections, so their lines were reduced to eight points by interpolation.

Wording

In the reports of the second and third studies, every sentence that held a number was generated from the result files by rule. This page works the same way. Each number in the main text and in these notes is filled in from the result files when the page is built, and the build fails if one is missing. Two kinds of number are typed by hand, those taken from someone else’s publication and those that describe how the work was done, such as the number of packs. The list in Every planned test and its result is copied from the two reports by a script that checks that no number changes on the way. One figure is left out there, a p of 1.00 in row E1, which is put into words.

The page draws on 1,306 texts, 1,186 stories and 120 non-stories, about 3.1 million words in all. Every one was read in English, and many are translations.

The random draws used a seed written into the plans, so they can be repeated.

What each kind was asked

Not every question was put to every text, which is why the totals in the main text change from one section to the next.

• All 559 folk tales, the ancient and medieval stories, the world films, the 99 Hollywood films, the literary stories and the classics have a line of fortune and a shape. Those are the 904 stories in “No single shape”.
• The classics are novels and plays, with one fairy tale among them, Perrault’s Cinderella. They were read in the first study, for fortune and for the mood of the writing. They have no tension scores, and their lines have 12 to 30 points, against 8 for every other story.
• The first study alone read 279 of the folk tales. It asked for fortune, tension, shape, ending and sixteen events.
• The second study’s longer form went to 711 texts. They are 280 folk tales (the 240 of the second batch and 40 of the first batch read again), the ancient and medieval stories, the world films, the 99 Hollywood films, the literary stories and the 120 non-stories. The 591 stories among them are the base for the features in “What a diary does not have”.
• The hero’s journey form went to the same 711 texts and to the 282 further films, which makes 993 texts, 873 of them stories. The further films were read with that form only, so they have no line of fortune.

The folk tales

Both batches come from English-language collections on Project Gutenberg, which holds books that are out of copyright in the United States. The numbers are Project Gutenberg’s book numbers.

The first batch has 319 tales from 29 collections. Ten tales were drawn from each, or fifteen where a region had three collections or fewer. Two collections gave fewer. Honey’s southern African book had only ten tales of the right length, and one of the fifteen Santal tales could not be cut into eight sections, which left fourteen.

• Europe: German (the Grimm brothers, 2591), Norwegian (Asbjørnsen and Moe, translated by Dasent, 8933), Russian (Afanasyev, translated by Magnus, 62509), Irish and Celtic (Jacobs, 7885), Italian (Basile’s Pentamerone, 2198), French (Perrault, 29021), Czech and Slovak (Fillmore, 32217), Romanian (Kremnitz, 20552)
• West Asia: Georgian (Wardrop, 44536), Turkish (Kúnos, 64807), Jewish (Landa, 26711), Arabic (the Arabian Nights in Andrew Lang’s edition, 128; six of the ten are voyages of Sindbad)
• South Asia: Indian (Jacobs, 7128), Santal (Bompas, 11938)
• East Asia: Chinese (Wilhelm, 29939), Japanese (Ozaki, 4018), Korean (Im Bang and Yi Ryuk, translated by Gale, 51002), Filipino (Cole, 12814)
• Africa: Nigerian (Dayrell, 34655), Khoekhoe, San and other southern African peoples (Honey, 38339), Xhosa (Theal, 71335)
• The Americas: Dakota (Zitkala-Sa, 338), and Sioux, the larger nation the Dakota belong to (McLaughlin, 341), Blackfoot (Grinnell, 11547), Zuni (Cushing, 54682), Inuit (Rasmussen, 28932), Brazilian (Eells, 24714)
• Oceania: Aboriginal Australian (K. Langloh Parker, 3833), Hawaiian (Thrum, 18450)

The second batch has 240 tales from 30 collections that the first study had not used, eight from each.

• Europe: Hungarian (Jones and Kropf, 42981), Basque (Webster, 34902), Ukrainian (Bain, 29672), Serbian (Mijatovich, 45321)
• West Asia: Armenian (Seklemian, 46944), Turkish coffee-house tales (Adler and Ramsay, 30577)
• South Asia: Bengali (Day, 38488), Punjabi (Steel, 6145), Sinhalese (H. Parker, 56614), tales of the Deccan (Frere, 36696)
• East Asia: Tibetan (O’Connor, 75000), Kalmyk and Mongolian (Busk, 40402; two cycles of stories that came from India), Shan (Griggs, 32375), Filipino (Fansler, 8299), Ainu (Chamberlain, 29287), Lao (Fleeson, 35564)
• Africa: Gold Coast, now Ghana (Barker, 66923), Mpongwe, Benga and Fang of Gabon (Nassau, 58900), Swahili (Bateman, 37472), Kongo (Weeks, 69558), Swazi and Zulu (Bourhill and Drake, 75833)
• The Americas: Pawnee (Grinnell, 36923), Seneca (A. C. Parker, 61477), Modoc (Curtin, 73418), Tlingit (Swanton, 78871), Cherokee (Mooney, 45634), Jamaican (Beckwith, 72735)
• Oceania: Hawaiian (Westervelt, 39195), Aboriginal Australian (Bruce, 56699), Māori (Dittmer, 54610)

Two texts are in twice. Perrault’s Cinderella was read as a classic and was also drawn among the French folk tales, and “Punchkin” was drawn from Jacobs’s book in the first batch and from Frere’s, which Jacobs took it from, in the second. So the 1,306 texts are 1,304 different ones, and the 559 folk tales are 558. The two readings of each gave the same shape. On the drawing board the two copies of Cinderella come 259th and 266th, and the main text gives the first, which is the copy among the folk tales.

East Asia is used broadly here. It takes in the Philippines, the Shan and the Lao of Southeast Asia, Tibet and the Mongols.

I call all of these folk tales. Many are what folklorists call myths and legends, which are told as true (Bascom 1965), and the Oceanian books are mostly of that kind.

In the library the books are named by the titles they were printed under, and several of those titles use words for peoples that were ordinary in English then and are offensive now.

These books are what collectors, translators and publishers chose to print in English, and a good many were retold for children, among them Lang’s Arabian Nights, the two books by Jacobs, Ozaki’s Japanese tales, Eells’s Brazilian ones and Bruce’s Australian ones. How shapes and endings were counted says more about what that may have done to the endings.

The ancient and medieval stories

There are six stories from each of eleven sources. They are Egyptian tales from the papyri (in Petrie’s edition), the Bible in the King James Version (five narratives from the Hebrew Bible and one, Bel and the Dragon, from the Apocrypha), Ovid’s Metamorphoses (Riley), the Jataka (Rhys Davids), the Sanskrit tales of the vetala (Ryder), Buddhist tales of Indian origin from Tibetan sources (Schiefner), medieval French romances, among them the lays of Marie de France (Mason), the Gesta Romanorum in a retelling of 1845, the Decameron (Rigg), Pu Songling (Giles) and Tales of Old Japan (Mitford).

The name fits the latest of these books only loosely. Pu Songling wrote around 1700, and Mitford wrote his Japanese tales down in the nineteenth century. The stories from Ovid are the numbered fables of Riley’s translation, and one of them, the story of Tereus, stops before its ending.

The line between this group and the folk tales is not sharp. Three of the Indian folk tales in Jacobs’s book are Jataka tales, and the Jataka is one of the eleven sources here. “The Tongue-Cut Sparrow” is in the sample in two tellings, Mitford’s among these stories and Ozaki’s among the folk tales.

The films

The world films are six each from the United States, Britain, France, Italy, Spain, Germany, the Nordic countries, Russia and the Soviet Union, India in Hindi, India in its four southern languages, Japan, Hong Kong, South Korea, China and Taiwan, Thailand, the Philippines, Latin America, and the Middle East and Africa. Only films credited to a single country were eligible. The corpus labels them with 28 countries.

A film here is a plot summary written by Wikipedia’s editors, a few hundred words long. The reader did not see the film or its script. The corpus also holds what English Wikipedia had in 2012, which favours films known to English speakers.

The films, and the test by era describes the 270 films by era, and How the journey was tested the 12 famous ones.

The non-stories

The dreams come from DreamBank, an archive of dream reports kept by dream researchers (Domhoff and Schneider 2008). Reports with a collector’s insertions in square brackets were left out.

The diaries are those of Samuel Pepys, Dorothy Wordsworth, Walter Scott, James Cook, Meriwether Lewis, Robert Falcon Scott, a nursing sister on the Western Front in 1914 and 1915, a Union soldier in the American Civil War (Lawrence Van Alstyne) and a schoolgirl in New York State in the 1850s and 1860s (Caroline Cowles Richards). Each gave four or five stretches.

The biographies are the first forty people on a list drawn at random. A racehorse on the list was skipped.

Left out, and kept

Before any reading of the second study, 39 stories were left out for reasons recorded in the plan. Two Egyptian tales have no ending and a third is the translator’s summary. Fifteen passages of Ovid open in the middle of a speech. Fourteen stories continue another story and do not stand alone. Two have the author’s commentary woven through them. Four tales of the Gesta Romanorum carry modern verse or come from a later source. One Sinhalese tale had a riddle printed in two columns that could not be untangled. The first study left out eight for similar reasons, two reprints of tales already in another collection and six Arabian Nights stories that had to be pieced together from separate passages.

Some doubtful texts were kept because their books print them as stories. They are six Pawnee “hero stories” that are lives of real men, six Tlingit clan histories, a Lao piece describing an enchanted place, six loosely told pieces by Maupassant, and second tellings of a tale by another narrator.

Four sources named in the plan gave too few separate stories of the right length and fell out. They are the Mabinogion, a book of Korean tales, a book of Māori tales and Hemingway.

Cutting into sections

Each text was cut into eight sections, as nearly equal in length as its paragraph or sentence breaks allow. A text was eligible only if every section would have at least fifteen words. The classics were cut at paragraph breaks into 12 to 30 sections, by length.

Forty tales of the first batch were prepared three ways for the second study. Each was read as printed, read with its eight sections in a random order, and read as seven openings, the first section alone, the first two, and so on up to seven. Every opening went to a reader who was told only that the text breaks off.

Sources

• Bamman, D., O’Connor, B., & Smith, N. A. (2013). Learning latent personas of film characters. Proceedings of ACL 2013, 352–361. Paper. The source of the CMU Movie Summary Corpus.
• Bascom, W. (1965). The forms of folklore: prose narratives. Journal of American Folklore 78(307): 3–20. Not opened. His division into myth, legend and folk tale is given as it is usually summarised.
• Domhoff, G. W., & Schneider, A. (2008). Studying dream content using the archive and search engine on DreamBank.net. Consciousness and Cognition 17(4): 1238–1247. Authors’ reprint. Abstract only. The source of the dream reports.
• Papalampidi, P., Keller, F., & Lapata, M. (2019). Movie plot analysis via turning point identification. Proceedings of EMNLP-IJCNLP 2019, 1707–1717. Paper. The source of the 99 Hollywood films with turning points marked by people.

A second AI model read 96 texts with the same forms as the first, twelve from each of eight kinds (folk tales, ancient and medieval stories, world films, Hollywood films, literary stories, dreams, diaries and biographies). The tables below compare its answers with the first reader’s on every question. Both readers are AI models, so this shows how far the answers depend on which model reads, and says nothing about how far either agrees with people.

“Same answer” is the share of the 96 texts on which the two gave the same answer. Kappa corrects that share for the agreement two readers would reach by chance, on a scale where 1 is perfect agreement and 0 is what chance would give. The plan of the third study called a kappa of 0.6 or more good agreement and one under 0.4 poor, and I use the same lines for the second study’s questions here. With 96 texts a kappa is uncertain by about 0.16 either way, so a value near a line could fall on either side of it.

Section by section

For each section after the first the readers also said whether it follows from what came before. They gave the same answer for 93% of sections (a kappa of 0.82). The first reader found a cause in 72% of sections and the second in 75%.

Questions about the whole text, second study

Three questions were answered with a number.

Two of the questions fall below the line for good agreement. The readers disagree most about whether a text is built on two opposed figures, and that feature is not one of the seven that pass. Whether the main character is changed inside is just under the line.

Events, second study

The form asked whether each of eighteen things happens, and in which section. The columns for the two readers give the share of the 96 texts in which each found it.

The second reader finds more of nearly every event. It records a problem at the start in 85% of the texts against the first reader’s 70%, and a test in 75% against 60%. Those two are the events with the lowest kappas, and both are below the line for good agreement. “Starts with a problem” is one of the seven features that pass. The twenty features, and the seven that passed gives each of the seven as the second reader scored it.

The hero’s journey form

The columns for the two readers give the share of the 96 texts in which each found the thing clearly. The last four questions are the ones that have nothing to do with the journey.

Of the 23 questions about the journey, 19 have a kappa of 0.6 or more, and the median over the twenty-two that have a kappa is 0.79. The last crisis is at 0.596, just under the line. Not wanting to go back and belonging to two worlds each have a kappa of zero because each was found once, by one reader, in 96 texts. Becoming divine was found by neither, so it has no kappa. On the exact answer (clearly, arguably or absent) the readers agree for 85% of all answers.

On the measures built from these questions, the two readers agree on setting out with a kappa of 0.81 and on the whole journey with 0.82. Their counts of Vogler’s twelve stages are within one of each other for 79% of texts. Their own overall impressions of whether a text is a journey, given before the questions, agree with a kappa of 0.72.

The other questions about the whole text got the same answer in these shares of texts: the main character’s sex 93%, their age 88%, how they come to leave 80%, the strangest place they go 83%, what they win 80%, and whether the reader recognised the text 77%.

The first reader, twice

Forty folk tales were read in the first study and again, eight days later, in the second, each time by a copy of a Claude model and with a different form. Which version of the model was used was not recorded either time, so I cannot say that it was the same one. Section by section the two readings correlate at 0.91 for fortune and 0.93 for tension. They name the same ending for 90% of the tales and the same shape for 65%. A second reading therefore names a different shape for about a third of the tales, about as often as the second model does.

Four questions were asked in both the second and the third study of the same 591 stories. The two readings agree with kappas of 0.76 for a trick, 0.71 for a broken rule, 0.78 for a punishment and 0.72 for a secret coming out.

The second and third studies listed their tests before any text was read, and each reported every one of them, including the ones that found nothing. The list below is copied from those two reports, with the names of things changed to the ones this page uses. The reports are not published here, but most of the readings they counted are, in Data, forms and plans. Tests A to K and the checks R and X belong to the second study, and L1 to L16 to the third. The first study had no plan, and its conclusions about folk tales come back here as test A.

An outcome is printed in italics where a rule for one existed. For tests A, C and F12 the second study’s plan fixed the rule before the reading, and for L1 and L5 the third study’s plan did. For the other tests the plans fixed what would be counted and what it would be set against. The words for chance are the first study’s. The plan fixed them for test A, and I decided to use them for B, E, G, H and K while the report was being written, after a first run on part of the readings and before the full readings were analysed. L11 uses them too. “In every region” for F1 to F11 is the first study’s rule for “usually”, which the plan did not ask for there, and the summary of D2 was chosen after that first run.

Against a shuffle or a random pairing, p is the share of shuffles that do at least as well as the real result. For tests A, B, D and K and for L11 it counts shuffles that lie as far from the average shuffle in either direction. A result is clearly above chance with p below 0.01, weakly above with p up to 0.05, and otherwise no better than chance. With 1,000 shuffles the smallest p there can be is 0.001. About a hundred tests are listed here, and among that many a few would reach p = 0.03 by chance, so the four results that rest on a p between 0.01 and 0.05 (A10, E3, the end at an extreme in K2, and L11) should be read as leanings. The other rules are given with the tests they belong to. What was fixed in advance, and what was added later says when each plan was written.

A. The first study’s twelve conclusions, on 240 folk tales it had not used

The first study gave each conclusion a word, and the plan fixed what each word means. “Nearly always” is 95% of tales or more. “Usually” is 80% or more, and at least 60% in every region. “A real lean” beats the shuffle with p below 0.01, in at least six of the seven regions and in a clear majority of collections. “A weak lean” has p between 0.01 and 0.05, and “no more than chance” has p above 0.05. The shuffle is the same tale with its eight tension scores, or its moves of fortune, in a random order. Orders of the moves that would take the line off the scale were not used.

• A1. A problem sets things going. 95% of 240 tales; by region from 90% (East Asia) to 100% (West Asia, South Asia and Oceania). Verdict: nearly always. First study: nearly always. Same verdict.
• A2. The main character gets into serious trouble. 88% of 240 tales; by region from 81% (Americas) to 94% (South Asia). Verdict: usually. First study: usually. Same verdict.
• A3. It ends clearly well or clearly badly. 84% of 240 tales; by region from 75% (Oceania) to 97% (Europe). Verdict: usually. First study: usually. Same verdict.
• A4. It ends calmer than its peak. 100% against 77% shuffled (p = 0.001); above the shuffle in 7 of 7 regions and 30 of 30 collections. Verdict: nearly always. First study: nearly always. Same verdict.
• A5. Tension builds, then drops. 93% against 59% shuffled (p = 0.001); above the shuffle in 7 of 7 regions and 30 of 30 collections. Verdict: usually. First study: usually. Same verdict.
• A6. Tension peaks in the second half. 64% against 48% shuffled (p = 0.001); above the shuffle in 6 of 7 regions and 24 of 30 collections. Verdict: a real lean. First study: a real lean. Same verdict.
• A7. The biggest swing comes in the last quarter. 47% against 36% shuffled (p = 0.001); above the shuffle in 6 of 7 regions and 24 of 30 collections. Verdict: a real lean. First study: a real lean. Same verdict.
• A8. It ends at its best or its worst point. 85% against 76% shuffled (p = 0.001); above the shuffle in 6 of 7 regions and 25 of 30 collections. Verdict: a real lean. First study: a real lean. Same verdict.
• A9. Fortune goes down before it goes up. 70% against 66% shuffled (p = 0.07); above the shuffle in 7 of 7 regions and 18 of 30 collections. Verdict: no more than chance. First study: no more than chance. Same verdict.
• A10. A low point late in the tale. 45% against 39% shuffled (p = 0.03); above the shuffle in 6 of 7 regions and 21 of 30 collections. Verdict: a weak lean. First study: no more than chance. Verdict changed.
• A11. A happy ending. 66% of 240 tales; by region from 33% (Oceania) to 97% (Europe). Verdict: varies by region. First study: varies by region. Same verdict.
• A12. One shape. The most common shape, fall then rise, fits 41% of 240 tales. Verdict: no one shape. First study: no one shape. Same verdict.

B. The same eight tests in other kinds of story

The kinds are the ancient and medieval stories, the world films, the 99 Hollywood films and the literary stories. A story whose fortune never moves has no biggest swing and no highest point, so it is left out of B7 and B8 and counted as lacking the feature in C7 and C8. That is why the literary stories differ by a point or two between the two tests.

• B1. A problem sets things going. Ancient and medieval stories 94%; world films 98%; Hollywood films 97%; literary stories 87%.
• B2. The main character gets into serious trouble. Ancient and medieval stories 82%; world films 96%; Hollywood films 98%; literary stories 84%.
• B3. It ends clearly well or clearly badly. Ancient and medieval stories 86%; world films 70%; Hollywood films 71%; literary stories 39%.
• B4. It ends calmer than its peak. Ancient and medieval stories 98% (77% shuffled); world films 93% (77% shuffled); Hollywood films 99% (78% shuffled); literary stories 87% (78% shuffled). Beats the shuffle in 3 of the 4 kinds.
• B5. Tension builds, then drops. Ancient and medieval stories 94% (59% shuffled); world films 90% (59% shuffled); Hollywood films 96% (61% shuffled); literary stories 87% (61% shuffled). Beats the shuffle in 4 of the 4 kinds.
• B6. Tension peaks in the second half. Ancient and medieval stories 64% (47% shuffled); world films 81% (48% shuffled); Hollywood films 88% (48% shuffled); literary stories 84% (48% shuffled). Beats the shuffle in 3 of the 4 kinds.
• B7. The biggest swing comes in the last quarter. Ancient and medieval stories 53% (36% shuffled); world films 55% (38% shuffled); Hollywood films 63% (38% shuffled); literary stories 50% (31% shuffled). Beats the shuffle in 4 of the 4 kinds.
• B8. It ends at its best or its worst point. Ancient and medieval stories 79% (75% shuffled); world films 78% (64% shuffled); Hollywood films 76% (62% shuffled); literary stories 73% (62% shuffled). Beats the shuffle in 2 of the 4 kinds.
• B, by region. The films of six world regions. Tension builds and then drops in 83% (North America) to 100% (Latin America) of films; the main character ends better off in 33% (Middle East and Africa) to 75% (South Asia).

C. Stories against non-stories

A feature passes if it is in at least 80% of folk tales and at least 60% of every kind of story, and is at least 20 points rarer in non-stories, with a 95% interval for the gap that excludes zero. The interval is in brackets.

• C1. A problem sets things going. Stories 96%, non-stories 28%; a gap of +68 points (+60 to +76). Lowest kind of story: literary stories, 87%. Passes.
• C2. The main character gets into serious trouble. Stories 90%, non-stories 44%; a gap of +46 points (+36 to +55). Lowest kind of story: ancient and medieval stories, 82%. Passes.
• C3. It ends clearly well or clearly badly. Stories 76%, non-stories 22%; a gap of +55 points (+46 to +62). Lowest kind of story: literary stories, 39%. Fails in literary stories.
• C4. It ends calmer than its peak. Stories 97%, non-stories 78%; a gap of +19.96 points, just under the bar of 20 (+13 to +28). Lowest kind of story: literary stories, 87%. Gap under 20 points.
• C5. Tension builds, then drops. Stories 93%, non-stories 66%; a gap of +27 points (+19 to +36). Lowest kind of story: literary stories, 87%. Passes.
• C6. Tension peaks in the second half. Stories 72%, non-stories 44%; a gap of +28 points (+19 to +37). Lowest kind of story: folk tales, 64%. Not common enough in stories.
• C7. The biggest swing comes in the last quarter. Stories 52%, non-stories 36%; a gap of +16 points (+8 to +24). Lowest kind of story: folk tales, 47%. Not common enough in stories.
• C8. It ends at its best or its worst point. Stories 81%, non-stories 56%; a gap of +25 points (+15 to +34). Lowest kind of story: literary stories, 71%. Passes.
• C9. One thing leads to another. Stories 96%, non-stories 18%; a gap of +78 points (+71 to +85). Lowest kind of story: literary stories, 76%. Passes.
• C10. The main question is settled by the end. Stories 84%, non-stories 23%; a gap of +61 points (+53 to +69). Lowest kind of story: literary stories, 50%. Fails in literary stories.
• C11. The main character wants something. Stories 97%, non-stories 50%; a gap of +47 points (+38 to +56). Lowest kind of story: literary stories, 92%. Passes.
• C12. One main character throughout. Stories 77%, non-stories 96%; a gap of −19 points (−24 to −14). Lowest kind of story: ancient and medieval stories, 71%. Not common enough in stories.
• C13. Five or fewer characters who matter. Stories 45%, non-stories 94%; a gap of −49 points (−55 to −43). Lowest kind of story: world films, 17%. Not common enough in stories.
• C14. An episode repeats. Stories 76%, non-stories 13%; a gap of +63 points (+56 to +70). Lowest kind of story: literary stories, 42%. Fails in literary stories.
• C15. The main character is an underdog. Stories 36%, non-stories 2%; a gap of +34 points (+30 to +38). Lowest kind of story: Hollywood films, 21%. Not common enough in stories.
• C16. Two characters are set up as opposites. Stories 13%, non-stories 0%; a gap of +13 points (+10 to +15). Lowest kind of story: literary stories, 0%. Not common enough in stories.
• C17. One line of action. Stories 83%, non-stories 45%; a gap of +38 points (+28 to +47). Lowest kind of story: world films, 68%. Passes.
• C18. The main character changes inside. Stories 25%, non-stories 3%; a gap of +23 points (+18 to +27). Lowest kind of story: folk tales, 10%. Not common enough in stories.
• C19. The ending is a surprise. Stories 32%, non-stories 3%; a gap of +28 points (+23 to +33). Lowest kind of story: folk tales, 23%. Not common enough in stories.
• C20. The stakes are life, love, family, wealth or status. Stories 74%, non-stories 24%; a gap of +50 points (+41 to +58). Lowest kind of story: literary stories, 42%. Fails in literary stories.

D. Forty tales read as printed and with their sections in a random order

“Content alone” is what the scores given to a tale as printed would show if its sections were put in the scrambled order.

• D1. Scores follow the content of a section, wherever it stands. Correlation between the two readings of the same section: 0.91 for tension, 0.96 for fortune, over 320 sections; 98% of tension scores within a point.
• D2. Does the reader supply an arc of its own? Tension builds and then drops: as printed 98%, scrambled text 80%, shuffled scores 55%, content alone 68%. Ends at an extreme, in the same order: 85%, 23%, 73%, 23%. Nearer to content alone on 4 of 5 measures.

E. The reader against turning points that people marked in 99 Hollywood films

Chance here is the rate when the reader’s scores for one film are set against the turning points of another.

• E1. The peak of tension is at the climax people marked, or beside it. 70% of films against 62% for random pairs (p = 0.001). The very same section: 6% against 16%, which is below chance, because the score is taken at the end of the section and by then the climax is over. Clearly above chance.
• E2. The biggest fall in fortune ends at the major setback, or beside it. 48% of films against 47% for random pairs (p = 0.40). The very same section: 29% against 19% (p = 0.002). No better than chance.
• E3. The lowest point of fortune is at the major setback, or beside it. 69% of films against 62% for random pairs (p = 0.02). The very same section: 33% against 27% (p = 0.06). Weakly above chance.

F. New candidates in 280 folk tales

“In every region” means 80% of tales, and at least 60% in each of the seven regions of the second batch.

• F1. One thing leads to another. 99% of folk tales; by region from 98% (East Asia) to 100% (Europe). In every region.
• F2. The main question is settled by the end. 96% of folk tales; by region from 92% (Oceania) to 100% (Europe). In every region.
• F3. The main character wants something. 97% of folk tales; by region from 92% (Americas) to 100% (Europe). In every region.
• F4. One main character throughout. 77% of folk tales; by region from 69% (West Asia) to 83% (Oceania). Not in every region.
• F5. Five or fewer characters who matter. 55% of folk tales; by region from 28% (Europe) to 71% (Oceania). The median cast is 5. Not in every region.
• F6. An episode repeats. 87% of folk tales; by region from 75% (Oceania) to 97% (Europe). Of the 172 tales with an episode run through three times or more, it is run through three times in 58% and four times in 24%. In every region.
• F7. The main character is an underdog. 48% of folk tales; by region from 8% (Oceania) to 69% (Europe). Not in every region.
• F8. Two characters are set up as opposites. 21% of folk tales; by region from 4% (Oceania) to 33% (Africa). Not in every region.
• F9. One line of action. 94% of folk tales; by region from 88% (West Asia) to 97% (Europe). In every region.
• F10. The main character changes inside. 10% of folk tales; by region from 3% (Europe) to 19% (West Asia). Not in every region.
• F11. The ending is a surprise. 23% of folk tales; by region from 8% (Oceania) to 38% (West Asia). Not in every region.
• F12. Those five stakes cover 80% of the tales of every region. 86% of the folk tales of the second batch; the lowest region is Oceania, at 75%. Does not hold.
• F13. What makes a text read as a story. A regression over 711 texts accounts for 72% of the variation in how far the reader took a text for a story. Largest: a chain of cause +1.37 (+1.11 to +1.62), a problem at the start +0.50 (+0.24 to +0.75) and an aim +0.43 (+0.22 to +0.66).

G. Is there one order of events?

• G1. Each region ranks the events as the other regions do. Rank correlation between a region and all the others: lowest 0.90 (Oceania), over 7 regions.
• G2. A tale’s own events stand in the order found elsewhere. 100% of 240 tales against 48% with positions dealt at random (p = 0.001). Clearly above chance.
• G, by kind. The folk tales’ order in other kinds of text. Ancient and medieval stories 98% of 65 (47% at random); world films 99% of 107 (47% at random); Hollywood films 99% of 99 (47% at random); literary stories 96% of 26 (45% at random); dreams 95% of 21 (43% at random); diaries 80% of 5 (42% at random); biographies 66% of 32 (46% at random).

H. Shared against local, in 559 folk tales from 59 collections

• H1. How the variation in the lines of fortune and tension divides. Between regions 4.7% (1.7% with regions dealt at random, p = 0.002); between collections of one region 11.1% (9.0% at random, p = 0.004); between tales of one collection 84.2%. For the events: 6.3%, 14.7% and 78.9%. Clearly above chance.
• H2. The region can be told from the plot. Right for 26% of held-out tales from the lines of fortune and tension (15% with regions dealt at random, p = 0.002), 28% from the events (14%, p = 0.002), 27% from both (14%, p = 0.002). The shares are averaged over the seven regions. Clearly above chance.

I. Where words fall in novels in other languages

Each row says in how many of the languages with a hundred books or more the prediction holds. It holds in a language when the 95% interval over that language’s books lies on the predicted side.

• I1. Home and family words open the book. Holds in 9 of 9 languages. With single volumes and second copies dropped: 9 of 9.
• I2. Death comes at the end. Holds in 9 of 9 languages. With single volumes and second copies dropped: 9 of 9.
• I3. Tears come at the end. Holds in 8 of 9 languages; not clear in Chinese. With single volumes and second copies dropped: 8 of 9.
• I4. Love and marriage come at the end. Holds in 8 of 9 languages; not clear in Hungarian. With single volumes and second copies dropped: 8 of 9.
• I5. Secrets come out at the end. Holds in 7 of 9 languages; not clear in Hungarian and Chinese. With single volumes and second copies dropped: 6 of 9.
• I6. Fear and danger build through the middle. Holds in 9 of 9 languages. With single volumes and second copies dropped: 8 of 9.
• I7. Fighting builds through the middle. Holds in 7 of 9 languages; not clear in Italian and Portuguese. With single volumes and second copies dropped: 6 of 9.
• I8. Joy returns on the last pages. Holds in 5 of 9 languages; not clear in German, Spanish, Hungarian and Portuguese. With single volumes and second copies dropped: 6 of 9.
• I9. The words darken from start to finish. Holds in 5 of 5 languages. With single volumes and second copies dropped: 5 of 5.
• I10. The last pages lift. Holds in 3 of 5 languages; not clear in Spanish and Portuguese. With single volumes and second copies dropped: 3 of 5.

J. The motif catalogue: 2,564 motifs in 958 societies

• J1. How widespread is the most widespread motif? The most widespread is in 37% of societies; 0 motifs are in at least half, 16 in a quarter, 164 in a tenth. Among the 282 societies with 100 motifs or more: 10 in at least half. Among the 81 with 200 or more: 143.
• J2. How many motifs are on every continent? 182 of 2,564 are recorded on all six; 14 in at least a tenth of the societies of each.
• J3. Do neighbours share more? Two societies share 20% of their motifs within 250 km and 4% beyond 8,000 km (568 societies with 50 motifs or more).
• J4. Motifs found in both the Old World and the Americas. 43% of all motifs.
• J5. The same for the two halves of the catalogue. Cosmology and origins 60%; adventures and tricksters 31%.

K. Forty tales read without knowing the end

• K1. Scores without hindsight agree with scores of the whole tale. Correlation 0.90 for tension and 0.96 for fortune; a mean difference of 0.27 and 0.41 of a point.
• K2. The patterns of tension hold without hindsight. Tension builds and then drops 95% without hindsight (52% shuffled, p = 0.001), 98% with; tension peaks in the second half 55% without hindsight (48% shuffled, p = 0.42), 63% with; the last section is the highest or lowest point 85% without hindsight (74% shuffled, p = 0.05), 85% with. The build and drop are clearly above chance, the late peak is no better than chance, and the end at an extreme is weakly above.
• K3. How soon the ending shows. The guess is right for 65% after one section, 68% after 3 sections, 75% after 5 sections and 83% after 7 sections. Always guessing “better”: 53%.

R. The second reader, on the 96 texts both read

• R. Two different models give the same scores. Tension: correlation 0.93, 99% within a point. Fortune: 0.91, 94% within a point. Every question is in The two readers, question by question.

X. Checks added after the results were seen

None of these was a planned test.

• X1. Do the seven features hold in stories the reader did not know well? The reader said it knew well 307 of the 591 stories and 25 of the 120 non-stories. Among the 284 stories it did not know well: a problem at the start 96% (96% of all stories); serious trouble 90% (90% of all stories); tension that builds and then drops 94% (93% of all stories); an end at the highest or lowest point 84% (81% of all stories); a chain of cause 97% (96% of all stories); an aim 96% (97% of all stories); one line of action 87% (83% of all stories).
• X2. Does the reader add an arc of its own to a scrambled text? Tension building and then dropping in the scrambled reading and not from content alone: 9 tales; the reverse: 4. A calmer end: 7 and 3. The section that comes last is scored 1.98 against 2.30 for the same section in its place; lower in 12 tales, higher in 3.
• X3. Are guesses at the ending memory? Openings whose reader said it knew the story well: 114 of 146 guesses right (78%). The rest: 92 of 134 (69%). In the 12 tales no opening reader knew well, the guess after one section is right for 58% and after seven for 75%; always guessing the most common ending, 42%.
• X4. Does the happy ending vary by the same regions in both batches? First batch as the first study read it, then second batch: Europe 85%, 97%; West Asia 83%, 56%; South Asia 69%, 66%; East Asia 68%, 69%; Africa 60%, 53%; Americas 33%, 75%; Oceania 40%, 33%. Rank correlation 0.32. The table in How shapes and endings were counted uses the later reading for the forty tales that were read twice, which changes Oceania’s first figure to 37%.
• X5. How high does the tension go? Texts in which tension reaches 3 or more of 5: folk tales 97%; ancient and medieval stories 91%; world films 97%; Hollywood films 98%; literary stories 82%; dreams 55%; diaries 23%; biographies 25%. Stories 96%, non-stories 34%. Rising to 3 or more and then let go: stories 90%, non-stories 28%.
• X6. Each of the seven features in each kind of non-story. A problem at the start: dreams 45%, diaries 18%, biographies 20%; serious trouble: dreams 60%, diaries 25%, biographies 48%; tension that builds and then drops: dreams 78%, diaries 40%, biographies 80%; an end at the highest or lowest point: dreams 48%, diaries 63%, biographies 58%; a chain of cause: dreams 35%, diaries 10%, biographies 8%; an aim: dreams 50%, diaries 38%, biographies 63%; one line of action: dreams 35%, diaries 35%, biographies 65%.

L. The hero’s journey, the third study

The journey counts as a universal only if the whole journey is clearly there in 80% of folk tales, in at least 60% of the folk tales of every region and in at least 60% of every kind of story. How the journey was tested defines the whole journey and the other measures. Stories here are the 591 of the second study unless a row says otherwise.

• L1. How common is it? Setting out is clearly there in 72% of folk tales. The whole journey is there in 21%, in 3% to 21% of the stories of each kind and in 15% to 34% of the folk tales of each region. There and back, or there to stay: 31%. Campbell’s sentence in full: 5.4%. The reader’s own impression that the text is mostly or wholly a journey: 31%. Not a universal by the plan’s bar. Added after the results: with the four legs in any order, 26% of folk tales; on the loosest reading (any order, either ending, counting arguable answers), 75%, still short of the bar.
• L2. The stages. Of Vogler’s twelve stages, five are clearly there in most folk tales (50% or more), four in a minority (20% to under 50%) and three in few (under 20%). A folk tale has 5.1 of the twelve on average, a story 4.6 and a non-story 1.2. Nine or more: 5.8% of stories. All twelve: one story. Of Campbell’s seventeen a story has 4.2 on average, and none has thirteen or more.
• L3. Arguable answers. Setting out: 63% of stories clearly, 87% counting arguable answers. The whole journey: 14% and 38%. Stages of the twelve: 4.6 and 8.1. There and back, or there to stay, counting arguable answers, in folk tales: 61%. Counting arguable answers, five of the twelve stages are usual in folk tales.
• L4. Stories against non-stories. The second study’s rule for a feature that passes, applied to the whole journey, to setting out and to every stage. The whole journey: 14% of stories and 2% of non-stories, a gap of 12 points (95% interval 9 to 16). Setting out: 63% and 33%. Neither is common enough in stories to pass. None of the twenty-three stages passes, and ten are at least 20 points more common in stories than in non-stories. Counting arguable answers, two stages pass (the call and the ordeal), setting out is common to stories and non-stories alike, and the whole journey is still not common enough in stories.
• L5. Can the form find a journey? Nine films built on the scheme or taught as examples of it. The form counts as able to find the journey if their median is nine or more of the twelve stages and at least seven of them have eight or more. The median is 9 and seven have eight or more. The bar is cleared. The whole journey is clearly there in four of the nine (added after the results: six with the legs in any order, nine on the loosest reading). Of the three films that are not adventures, Pulp Fiction has 6 stages, The Full Monty 4 and Titanic 6.
• L6. Two readers. Agreement between two models on the 96 texts both read. The whole journey: kappa 0.82. Setting out: 0.81. The reader’s impression, mostly or wholly a journey: 0.72. Count of stages: correlation 0.91. Of the twenty-three stages, nineteen have good agreement (0.6 or more), one moderate (0.4 to 0.6) and two poor. One was found by neither reader.
• L7. The earlier readings. A journey pieced together from the second study’s readings, which asked nothing about a journey, is found in 53% of folk tales, 39% of stories and 9% of non-stories. Text by text it agrees with there and back, or there to stay, at a kappa of 0.49, and with the whole journey at 0.29. The four comparison questions against the second study: a trick 0.76, a broken rule 0.71, a punishment 0.78 and a secret coming out 0.72.
• L8. Order and timing. The test uses texts with at least three of the twelve stages in two or more sections. The plan called it weak evidence, since part of the order is forced. In 451 stories, 95% lean towards the scheme’s order, against 46% with the same sections dealt at random (p below 0.001). Stages are found where Vogler’s diagram puts them 39% of the time, against 19% by chance. The ordeal is in sections 4 or 5 in 20% and in sections 6 to 8 in 67%. It is in the section of highest tension in 53% (30% at random, p below 0.001).
• L9. Scrambled tales. Forty tales read as printed and with their sections in a random order. If the reader follows the text, the scrambled reading shows no more of the scheme’s order than the positions found as printed, moved into the scrambled order. Stages of the twelve: 4.0 scrambled and 4.2 as printed (difference −0.23, 95% interval −0.60 to +0.13). Leaning towards the scheme’s order: 92% as printed, 42% scrambled as presented, 58% for the positions found as printed, moved. The same stage is placed in the same passage 59% of the time. Question by question the two readings give the same answer 87% of the time, and the median kappa over the twenty-three stages is 0.84.
• L10. The order of the questions. Stories have 4.59 stages with one order of the questions and 4.60 with the other (difference −0.01, 95% interval −0.38 to +0.38). The whole journey: 11.4% and 16.3% (difference −4.9 points, −10.2 to +0.5). The reader’s overall impression, reported as well: 1.39 and 1.61 of 4 (difference −0.21, −0.38 to −0.05).
• L11. Hollywood by era. The mean number of stages in Hollywood films of 1930 to 1976, 1977 to 1992 and 1993 to 2012. If the journey is a recipe taken up after Star Wars it rises, and if it is how stories have always gone it does not. Stages of the twelve: 4.17, 4.84 and 4.87. Last minus first: +0.70 (95% interval +0.04 to +1.36; p = 0.031, and 0.039 when I ran the test again for this page with 200,000 shuffles). Rises, weakly. The whole journey: 9%, 10% and 17% (last minus first +8 points, −2 to +18). The reader’s impression that a film is mostly or wholly a journey: 7%, 17% and 17% (last minus first +10 points, +1 to +19). Adventure, fantasy, science fiction, action and family films are 46%, 49% and 41% of the films of each era. Added after the results, last minus first within each kind: those films +1.76 (+0.77 to +2.72), the rest +0.01 (−0.76 to +0.76).
• L12. Whose journey? The test uses stories with one human main character. Folk tales, 116 men or boys and 34 women or girls: the whole journey in 32% and 18% (gap +14 points, 95% interval −2 to +29); stages 6.0 and 5.3 (gap +0.67, −0.22 to +1.57). All stories, 290 and 86: the whole journey in 18% and 13% (gap +5 points, −4 to +13); stages 5.2 and 4.4 (gap +0.86, +0.28 to +1.43). Added after the results, the gaps inside each kind of story, averaged: the whole journey +4 points (−5 to +13); stages +0.64 (+0.09 to +1.18). In folk tales 10% of main characters are sent, 24% are driven out, 51% go by choice and 15% do not leave, and 34% enter a supernatural world.
• L13. The journey among the seven features. Of 82 stories with the whole journey, 70% have all seven features that pass in the second study. Of stories without it, 52% do. Of 320 stories with all seven, 18% have the whole journey. Added to the second study’s account of what makes a text read as a story, the whole journey is worth +0.14 of a point (+0.03 to +0.24) and the count of stages, all twelve against none, +0.50 (+0.26 to +0.72).
• L14. Memory. Stories the third study’s reader said it knew well (329): the whole journey in 12%, and 4.4 stages on average. Stories it did not know well (262): 17%, and 4.8 stages.
• L15. The journey’s images in the motif catalogue. Ten images that Campbell names, chosen from the titles of motifs alone and counted in 958 societies, with no reader involved. The most widespread are the boon (39% of societies), the road of trials (37%) and the belly of the whale (36%). Seven of the ten are on all six continents, two are in a tenth of the societies of each continent, and none is in half of all societies. In the 81 best-recorded societies the median is 6 of the ten, and 67% have an image from each of Campbell’s three chapters (departure, initiation and return).
• L16. Do the stages go together? Added before the reading began. All four legs of the journey are clearly there in 19.5% of stories, against 10.9% expected if each leg were dealt out separately within each kind of story (p below 0.001). The 10.9% is the average of the study’s random deals. Worked out by arithmetic it is 11.0%, the figure used elsewhere on this page. The spread of the count of stages is 2.50, against 1.71 expected (p below 0.001). The mean correlation between pairs of stages is 0.13.

The counts on the page are made from the files below. They are JSON and text files, and they open in a browser.

Shares are rounded to the nearest whole number, halves upward. Rounded shares may therefore add up to 99 or 101, and gaps were worked out before rounding, so a few differ in the last digit from the figures beside them.

The readings

• texts.json has one record for each of the 1,306 texts. It gives the title, the source, the region, the number of words, the eight fortune scores and eight tension scores, the shape and the ending the reader named, the reader’s one-sentence summary, the number of Vogler’s stages found, and whether the whole journey is there. A text has only what it was asked for, so the 282 films read only for the hero’s journey have no fortune line, and the classics have no tension scores.
• The files of detail hold the rest of what the first reader recorded, text by text. That is the reader’s one line on each section, whether each section follows from what came before, the section in which each event first happens, the answers to the questions about the whole text, and for the hero’s journey every answer with the reader’s few words on what in the text it is. The classics and the 279 folk tales read only in the first study have the shorter set of answers that its forms asked for. There is one file for each kind: folk tales, ancient and medieval stories, world films, Hollywood films, Hollywood films by era, the twelve famous films, literary stories, classics, dreams, diaries and biographies.
• second_reader.json has the second reader’s scores and answers for the 96 texts it read in the second study and the 96 it read for the hero’s journey.
• charts.json holds the tallies behind the charts, and typical.json the ranking of stories by how typical they are. words.json has the counts of word families through novels in English, in the nine other languages on the page, and in Swedish, which had too few books to be counted.
• motifs.json has the 240 most widespread motifs with the societies in which each is recorded, which the map needs, and the counts made from the whole catalogue. It is taken from the replication data of Michalopoulos and Xue (2021), which they released into the public domain. The whole catalogue is in their deposit.

The texts themselves are not republished here. The texts says where each kind comes from. Each folk tale is labelled with its people and region, and the list of collections in that note gives the book.

The names in the files are short. In a file of detail, `e` is the reader’s line on each section, `L` says how far each section after the first follows from the one before (2 for caused by it, 1 for merely after it, 0 for unconnected), `el` gives the section in which each of fourteen events first happens, `g` is the aim, `c` whether the main matter is settled, `rb` what settles it, and `sl` the reader’s score for how far the text reads as a story. Under `a` are the other answers about the whole text: how well the reader said it knew the text (`kn`, from 0 to 2), the main character’s kind and role (`pk`, `ro`), whether the aim is reached (`go`), how many characters matter (`n`), one main character (`one`), repeated episodes (`rep`), an underdog (`ud`), two opposed figures (`con`), a single line of action (`ss`), a change inside (`ch`), a surprise (`su`), and under `x` four more events (something supernatural, a deserved ending, an explained origin, a stated moral). Under `hj`, `c` is a string with one digit for each of the 27 questions about the journey, in the order of `journey.stages` in charts.json: 2 for clearly there, 1 for arguably, 0 for absent.

The forms the readers were given

• The classics, first study
• The first batch of folk tales, first study
• The second study
• The hero’s journey, third study

The forms are printed as they were written, except that the path of the working folder is replaced by a placeholder.

The plans

• The plan of the second study, with its dated additions and its list of what was settled after the data were in
• The plan of the third study, likewise

The plans are printed as they were written, so they use the working names of the time. A “scene” or “segment” is a section, a “record” is a non-story, the “protagonist” is the main character, a “life” is a biography, “old written stories” are the ancient and medieval stories, “plain” and “loose” are this page’s “clearly” and “counting arguable answers”, the “round trip” is the whole journey, and a “mark of stories” is a feature that passes. “Part two” and “part three” are the second and third studies, and “this page” in a plan means that study’s own report. One sentence of the third plan, which named the person who asked for the study, has been taken out, and its place is marked.

The first study

The first study has a page of its own at shapes-of-stories.netlify.app, with its methods and its charts. It reruns the paper on the six shapes, counts the larger collection of fiction, and reports the first reading of the classics and of the first batch of folk tales.

What is not here

The scripts that drew the samples, ran the readers and made the counts are not published on this site, and neither are the two longer reports of the second and third studies, from which the list of tests is taken. The readings of the forty tales with their sections scrambled and of the 280 tales cut off part of the way through are not here either. Their totals are in the list of tests.

Sources

• Michalopoulos, S., & Xue, M. M. (2021). Folklore. Quarterly Journal of Economics 136(4): 1993–2046. Article; the data, which is the version of Berezkin’s catalogue used here.

This is every work cited in the notes. Most were opened and read, at least in the part cited, by me or by a copy of me sent to check sources. An entry says so where only part of a work was read, and where a work was not opened and is known through another source.

• Aristotle. Poetics. Translated by S. H. Butcher (first published 1895). Project Gutenberg text. Chapters 7 to 11 and 18 were read, and checked against Bywater’s and Fyfe’s translations.
• Bamman, D., O’Connor, B., & Smith, N. A. (2013). Learning latent personas of film characters. Proceedings of ACL 2013, 352–361. Paper. The source of the CMU Movie Summary Corpus.
• Bartlett, F. C. (1932). Remembering: A Study in Experimental and Social Psychology. Cambridge University Press. Not opened. The experiment with “The War of the Ghosts” is described as textbooks give it.
• Bascom, W. (1965). The forms of folklore: prose narratives. Journal of American Folklore 78(307): 3–20. Not opened. His division into myth, legend and folk tale is given as it is usually summarised.
• Bentz, M., Cortez Espinoza, M., Simeonova, V., Köppe, T., & Onea, E. (2024). Measuring suspense in real time: a new experimental methodology. Scientific Study of Literature 13(1): 1–21. Article.
• Berezkin, Y. E., & Duvakin, E. N. Thematic classification and areal distribution of folklore and mythological motifs: an analytical catalogue. Online, in Russian. Only the front page and the first motif page were opened; the catalogue was used through the data of Michalopoulos and Xue (2021).
• Berezkin, Y. E. (2015). Folklore and mythology catalogue: its lay-out and potential for research. RMN Newsletter 10: 58–70. The issue. Only the passages quoted were read.
• Booker, C. (2004). The Seven Basic Plots: Why We Tell Stories. Continuum. Read only in search snippets of a lending scan, and described from two reviews, one of them by Adam Mars-Jones in The Observer, 21 November 2004.
• Boot, P., Daza, A., Schnober, C., & van Hage, W. (2024). In the context of narrative, we never properly defined the concept of valence. Computational Humanities Research 2024, CEUR Workshop Proceedings 3834. Paper. The passages on Reagan and colleagues were read.
• Bordwell, D. (1982). Happily ever after, part two. The Velvet Light Trap 19: 2–7. Scan on the author’s site.
• Bordwell, D. (1985). Narration in the Fiction Film. Madison: University of Wisconsin Press. One sentence was read in search snippets of a lending scan, and the page number was not seen. The sample of films is described in Bordwell, D., Staiger, J., & Thompson, K. (1985), The Classical Hollywood Cinema: Film Style and Mode of Production to 1960 (New York: Columbia University Press), a few sentences of which were read the same way.
• Boyd, R. L., Blackburn, K. G., & Pennebaker, J. W. (2020). The narrative arc: revealing core narrative structures through text analysis. Science Advances 6(32): eaba2196. Article.
• Brewer, W. F., & Lichtenstein, E. H. (1981). Event schemas, story schemas, and story grammars. In J. Long & A. Baddeley (eds), Attention and Performance IX, 363–379. Hillsdale, NJ: Erlbaum. Read as Technical Report No. 197, Center for the Study of Reading, University of Illinois at Urbana-Champaign, December 1980; the printed chapter may differ.
• Brewer, W. F., & Lichtenstein, E. H. (1982). Stories are to entertain: a structural-affect theory of stories. Journal of Pragmatics 6(5–6): 473–486. Read as Technical Report No. 265, Center for the Study of Reading, October 1982; the article may differ.
• Brewer, W. F., & Ohtsuka, K. (1988). Story structure, characterization, just world organization, and reader affect in American and Hungarian short stories. Poetics 17(4–5): 395–415. Read in an earlier version, Technical Report No. 375, March 1986; the article may differ.
• Brewer, W. F. (1985). The story schema: universal and culture-specific properties. In D. R. Olson, N. Torrance & A. Hildyard (eds), Literacy, Language, and Learning: The Nature and Consequences of Reading and Writing, 167–194. Cambridge University Press. Read as Technical Report No. 322, Center for the Study of Reading, University of Illinois at Urbana-Champaign, August 1984; the printed chapter may differ.
• Campbell, J. (1949). The Hero with a Thousand Faces. Bollingen Series XVII. New York: Pantheon. Read only through search snippets of two lending scans at the Internet Archive (one of them, the second edition of 1968), so no page number here was checked and the book was not read through.
• Chang, K. K., Cramer, M., Soni, S., & Bamman, D. (2023). Speak, memory: an archaeology of books known to ChatGPT/GPT-4. Proceedings of EMNLP 2023, 7312–7327. Read in part in the arXiv copy: the abstract, the description of the data and the passage on books out of copyright.
• Chhun, C., Suchanek, F. M., & Clavel, C. (2024). Do language models enjoy their own stories? Prompting large language models for automatic story evaluation. Transactions of the Association for Computational Linguistics 12: 1122–1142. arXiv copy. The abstract, the introduction and the description of the stories were read.
• Cipolli, C., & Poli, D. (1992). Story structure in verbal reports of mental sleep experience after awakening in REM sleep. Sleep 15(2): 133–142. Record. Abstract only.
• Cook, V. (1965). Lord Raglan’s hero: a cross cultural critique. The Florida Anthropologist 18: 147–154. Not opened. Described from Dundes’s account of it.
• Cousineau, P. (ed.) (1990). The Hero’s Journey: Joseph Campbell on His Life and Work. Harper & Row. One remark by George Lucas, made at the National Arts Club in 1985, was read in a search snippet of a lending scan.
• Cutting, J. E. (2016). Narrative theory and the dynamics of popular movies. Psychonomic Bulletin & Review 23(6): 1713–1743. Article.
• da Silva, S. G., & Tehrani, J. J. (2016). Comparative phylogenetic analyses uncover the ancient roots of Indo-European folktales. Royal Society Open Science 3(1): 150645. Article.
• Del Vecchio, M., Kharlamov, A., Parry, G., & Pogrebna, G. (2018). The data science of Hollywood: using emotional arcs of movies to drive business model innovation in entertainment industries. arXiv:1807.02221. A revised version appeared in Journal of the Operational Research Society 72(5): 1110–1137 (2021), which was not read.
• Dodds, P. S., Clark, E. M., Desu, S., and others (2015). Human language reveals a universal positivity bias. Proceedings of the National Academy of Sciences 112(8): 2389–2394. arXiv copy. Abstract only. The source of the mood lists in the five other languages.
• Dodds, P. S., Harris, K. D., Kloumann, I. M., Bliss, C. A., & Danforth, C. M. (2011). Temporal patterns of happiness and information in a global social network: hedonometrics and Twitter. PLoS ONE 6(12): e26752. Article. The source of the word list.
• Domhoff, G. W., & Schneider, A. (2008). Studying dream content using the archive and search engine on DreamBank.net. Consciousness and Cognition 17(4): 1238–1247. Authors’ reprint. Abstract only. The source of the dream reports.
• Dundes, A. (1968). Introduction to the second edition of Propp’s Morphology of the Folktale, xi–xvii. Austin: University of Texas Press. Read in the excerpts linked under Propp.
• Dundes, A. (1977). The hero pattern and the life of Jesus. First given as a paper in December 1976. As reprinted in In Quest of the Hero, 179–223. Princeton University Press, 1990. Read in search snippets of a lending scan.
• Eckermann, J. P. Conversations with Goethe, the entry for 14 February 1830, in John Oxenford’s English translation (1875), page 437. Scan at the Internet Archive, checked against the German text.
• Enderle, S. (2016). A plot of Brownian noise. Notebook, chosen by Digital Humanities Now on 8 September 2016.
• Field, S. (1979). Screenplay: The Foundations of Screenwriting. New York: Dell. Not opened. The three acts are described from summaries of the book.
• Finlayson, M. A. (2016). Inferring Propp’s functions from semantically annotated text. Journal of American Folklore 129(511): 55–77. Author’s copy.
• Forster, E. M. (1927). Aspects of the Novel. Scan at the Internet Archive. The Clark lectures given at Trinity College, Cambridge, in spring 1927; chapter 5 was read.
• Freytag, G. (1863). Die Technik des Dramas. Leipzig: S. Hirzel. In English: Freytag’s Technique of the Drama, translated by Elias J. MacEwan from the sixth German edition. Chicago: S. C. Griggs, 1895. Scan of the second edition of the translation, 1896; German checked in the tenth edition. The passages quoted were found by searching the text of the scan, and the book was not read through.
• Gervás, P., & Méndez, G. (2024). Tagging narrative with Propp’s character functions using large language models. Proceedings of Text2Story 2024, CEUR Workshop Proceedings 3671, 137–148. Paper.
• Haase, D. (2010). Decolonizing fairy-tale studies. Marvels & Tales 24(1): 17–38. Read in the open copy at the University of Hawaiʻi.
• Harmon, D. Story structure 101. Channel 101 wiki, on a page created on 13 August 2009.
• Hendrickx, I., Onrust, L., Kunneman, F., Hürriyetoğlu, A., Stoop, W., & van den Bosch, A. (2017). Unraveling reported dreams with text analytics. Digital Humanities Quarterly 11(4). Read in the arXiv copy of 2016, which may differ.
• Herrmann, B., & Lüdtke, J. (2023). A fairy tale gold standard: annotation and analysis of emotions in the Children’s and Household Tales by the Brothers Grimm. Zeitschrift für digitale Geisteswissenschaften 8. Article. The abstract and the passages on the design and on the shapes were read.
• Hinds, J. (1982). Contrastive rhetoric: Japanese and English. Paper given at the TESOL convention, Honolulu, May 1982 (ERIC ED 217 720). Scan at the Internet Archive. The passages quoted were found by searching the text of the scan, and the paper was not read through.
• Hobson, D. G., Zhou, H., Ruths, D., & Piper, A. (2024). Story morals: surfacing value-driven narrative schemas using large language models. Proceedings of EMNLP 2024, 12998–13032. Paper.
• Hogan, P. C. (2003). The Mind and Its Stories: Narrative Universals and Human Emotion. Cambridge University Press. Record. Not opened. Known from the author’s summary of 2016.
• Hogan, P. C. (2016). Story. Literary Universals Project, University of Connecticut, 20 November 2016. Article.
• Invernizzi, M. A., & Abouzeid, M. P. (1995). One story map does not fit all: a cross-cultural analysis of children’s written story retellings. Journal of Narrative and Life History 5(1): 1–19. Record. Abstract only.
• Jiang Yin (1998). Qi cheng zhuan he: jixie jiegou lun de xiaozhang [Opening, continuation, turn and close: the rise and decline of a mechanical theory of structure]. Wenxue yichan 1998(3). In Chinese. Only the passage on where the scheme is first found was read, in a copy online.
• Jobling, I. (2001). The psychological foundations of the hero-ogre story: a cross-cultural study. Human Nature 12(3): 247–272. Publisher’s page. Abstract only. The count of 31 cultures in 44 is from the summary of the paper in the Human Relations Area Files.
• Jockers, M. L. (2015). The rest of the story. Blog post, 25 February 2015, followed by Requiem for a low pass filter, 6 April 2015.
• Kintsch, W., & Greene, E. (1978). The role of culture-specific schemata in the comprehension and recall of stories. Discourse Processes 1(1): 1–13. Record. Not opened; described from Brewer (1984).
• Kirkpatrick, A., & Xu, Z. (2012). Chinese Rhetoric and Writing: An Introduction for Language Teachers. WAC Clearinghouse and Parlor Press. The chapter “The Literary Background and Rhetorical Styles” was read in the LibreTexts copy.
• Kluckhohn, C. (1959). Recurrent themes in myths and mythmaking. Daedalus 88(2): 268–279. JSTOR. Read in a copy of the scan.
• Knight, S., Rocklage, M. D., & Bart, Y. (2024). Narrative reversals and story success. Science Advances 10(34): eadl2013. Article.
• Labov, W., & Waletzky, J. (1967). Narrative analysis: oral versions of personal experience. In J. Helm (ed.), Essays on the Verbal and Visual Arts, 12–44. American Ethnological Society and University of Washington Press. Copy, read as page images.
• Labov, W. (1972). The transformation of experience in narrative syntax. In Language in the Inner City. University of Pennsylvania Press. Read in search snippets of a lending scan.
• Lehne, M., Engel, P., Rohrmeier, M., Menninghaus, W., Jacobs, A. M., & Koelsch, S. (2015). Reading a suspenseful literary text activates brain areas related to social cognition and predictive inference. PLOS ONE 10(5): e0124550. Article.
• Lévi-Strauss, C. (1960). La structure et la forme: réflexions sur un ouvrage de Vladimir Propp. In English in V. Propp, Theory and History of Folklore (University of Minnesota Press, 1984). Read in search snippets of a lending scan.
• Maharjan, S., Kar, S., Montes-y-Gómez, M., González, F. A., & Solorio, T. (2018). Letting emotions flow: success prediction by modeling the flow of emotions in books. Proceedings of NAACL-HLT 2018, volume 2: 259–265. Read in the arXiv copy.
• Mandler, J. M., Scribner, S., Cole, M., & DeForest, M. (1980). Cross-cultural invariance in story recall. Child Development 51(1): 19–26. Record. Abstract only, with Brewer’s (1984) account of it.
• Matlin, G., Zhang, D., Barroso Loza, R., Popescu, D. M., Isbell, J., Chakraborty, C., & Riedl, M. (2025). Do language models agree with human perceptions of suspense in stories? arXiv:2508.15794; the arXiv record gives the conference COLM 2025.
• McClure, D. (2017). Distributions of words across narrative time in 27,266 novels. Techne, Stanford Literary Lab, 10 July 2017. Post.
• Michalopoulos, S., & Xue, M. M. (2021). Folklore. Quarterly Journal of Economics 136(4): 1993–2046. Article; the data, which is the version of Berezkin’s catalogue used here.
• Michelmann, S., Kumar, M., Norman, K. A., & Toneva, M. (2025). Large language models can segment narrative events similarly to humans. Behavior Research Methods 57: 39. Paper. Abstract only, in the open copy.
• Mitran, L., Wu, S., & Piper, A. (2025). Probing narrative morals: a new character-focused MFT framework for use with large language models. Proceedings of EMNLP 2025, 28514–28529. Paper.
• Morin, O., Acerbi, A., & Sobchuk, O. (2019). Why people die in novels: testing the ordeal simulation hypothesis. Palgrave Communications 5: 62. Article.
• Moyers, B. (1999). The mythology of Star Wars with George Lucas. Transcript dated 18 June 1999.
• Olrik, A. (1908). Episke love i folkedigtningen. Danske Studier 5: 69–89. A German version followed in 1909. In English: “Epic laws of folk narrative”, in A. Dundes (ed.), The Study of Folklore, 129–141 (1965). Not opened. The laws are given as they are usually summarised.
• Pannach, F., Sporleder, C., May, W., Krishnan, A., & Sewchurran, A. (2021). Of lions and Yakshis: ontology-based narrative structure modelling for culturally diverse folktales. Semantic Web 12(2): 219–239. Article. The abstract and the passage on the two analyses of one tale were read.
• Papalampidi, P., Keller, F., & Lapata, M. (2019). Movie plot analysis via turning point identification. Proceedings of EMNLP-IJCNLP 2019, 1707–1717. Paper. The source of the 99 Hollywood films with turning points marked by people.
• Piper, A., & Bagga, S. (2024). Using large language models for understanding narrative discourse. Proceedings of the 6th Workshop on Narrative Understanding, 37–46. Paper.
• Polti, G. (1895). Les trente-six situations dramatiques. Paris: Mercure de France. In English: The Thirty-Six Dramatic Situations, translated by Lucille Ray (1916). Scan of the 1921 printing. The passages quoted were found by searching the text of the scan, and the book was not read through.
• Propp, V. (1928). Morfologiia skazki. Leningrad: Academia. In English: Morphology of the Folktale, translated by Laurence Scott; second edition, revised and edited by Louis A. Wagner, with an introduction by Alan Dundes. Austin: University of Texas Press, 1968. Read in excerpts of the second edition, which hold Dundes’s introduction and the chapters on method and on the functions. One sentence from chapter IX (page 107) was read in another scan of the same edition. The rest of the book was not read.
• Quinn, S. F. (1993). How to write a Noh play: Zeami’s Sandō. Monumenta Nipponica 48(1): 53–88. Copy. Contains her annotated translation of Sandō (1423). Pages 58 to 65 were used, and the rest was not read through.
• Ragan, K. (2009). What happened to the heroines in folktales? An analysis by gender of a multicultural sample of published folktales collected from storytellers. Marvels & Tales 23(2). Record. Abstract only.
• Raglan, Lord. (1936). The Hero: A Study in Tradition, Myth and Drama. London: Methuen. The list of incidents and the scores were read in a scan of the 1949 Watts edition, found by searching its text. The book was not read through.
• Rank, O. (1909). Der Mythus von der Geburt des Helden. In English: The Myth of the Birth of the Hero, translated by F. Robbins and Smith Ely Jelliffe. New York: Journal of Nervous and Mental Disease Publishing Company, 1914. Scan of the translation. The passages used were found by searching the text of the scan, and the book was not read through.
• Reagan, A. J. (2017). Towards a science of human stories: using sentiment analysis and emotional arcs to understand the building blocks of complex social systems. PhD thesis, University of Vermont. arXiv:1712.06163. Pages 37 to 39 were read.
• Reagan, A. J., Mitchell, L., Kiley, D., Danforth, C. M., & Dodds, P. S. (2016). The emotional arcs of stories are dominated by six basic shapes. EPJ Data Science 5: 31. Journal page; read in the arXiv copy.
• Ross, R. M., Greenhill, S. J., & Atkinson, Q. D. (2013). Population structure and cultural geography of a folktale in Europe. Proceedings of the Royal Society B 280(1756): 20123065. Article.
• Schmidt, M.-L. C. R., Winkler, J. R., Appel, M., & Richter, T. (2023). Tracking emotional shifts during story reception: the relationship between narrative structure and affective responses. Scientific Study of Literature 12(1): 17–39. Article.
• Segal, R. A. (1987). Joseph Campbell: An Introduction. New York: Garland. Read in search snippets of a lending scan.
• Shields, C. J. (2011). And So It Goes: Kurt Vonnegut, a Life. Henry Holt. Read in search snippets of a lending scan.
• Singh, M. (2021). The sympathetic plot, its psychological origins, and implications for the evolution of fiction. Emotion Review 13(3): 183–198. Read in the accepted manuscript, which may differ from the printed article.
• Snyder, B. (2005). Save the Cat! The Last Book on Screenwriting You’ll Ever Need. Studio City: Michael Wiese Productions. Not opened. The page numbers of the beat sheet are from summaries of it.
• Steg, M., Slot, K., & Pianzola, F. (2022). Computational detection of narrativity: a comparison using textual features and reader response. Proceedings of LaTeCH-CLfL 2022, 105–114. Paper.
• Sternberg, M. (2003). Universals of narrative and their cognitivist fortunes (I). Poetics Today 24(2): 297–395. Record. Pages 327 and 328 were read.
• Subbiah, M., Mian, H., Deas, N., Mayukha, A., McAdams, D. P., & McKeown, K. (2026). Whose story gets told? Positionality and bias in LLM summaries of life narratives. arXiv:2604.20131. A preprint.
• Sui, P., Zhu, Y., Cheng, T., West, P., So, R. J., Long, H., & Holtzman, A. (2026). Spoiler alert: narrative forecasting as a metric for tension in LLM storytelling. arXiv:2604.09854. A preprint. The abstract and the introduction were read.
• Swafford, A. (2015). Problems with the Syuzhet package. Blog post, 2 March 2015.
• Tehrani, J. J. (2013). The phylogeny of Little Red Riding Hood. PLoS ONE 8(11): e78871. Article.
• The Daśarūpa: A Treatise on Hindu Dramaturgy. Translated by George C. O. Haas. New York: Columbia University Press, 1912. Scan at the Internet Archive. Page 9 was read as a page image.
• The first of the three studies behind this page, which reruns Reagan and colleagues’ analysis and tests it against shuffled text, is written up at shapes-of-stories.netlify.app.
• The Nāṭyaśāstra: A Treatise on Hindu Dramaturgy and Histrionics, ascribed to Bharata-muni. Translated by Manomohan Ghosh, vol. I. Calcutta: Royal Asiatic Society of Bengal, 1950. Chapter XXI and the translator’s introduction, as reproduced at wisdomlib.org.
• Thompson, K. (1999). Storytelling in the New Hollywood: Understanding Classical Narrative Technique. Cambridge, MA: Harvard University Press. Not opened. Cited from Cutting (2016).
• Thompson, S. (1955–1958). Motif-Index of Folk-Literature, volumes I–VI. FF Communications 106–109, 116–117. Helsinki: Academia Scientiarum Fennica. Not opened. Details from Berezkin (2015).
• Thuillard, M., Le Quellec, J.-L., d’Huy, J., & Berezkin, Y. (2018). A large-scale study of world myths. Trames 22(4): 407–424. Article.
• Tian, Y., Huang, T., Liu, M., Jiang, D., Spangher, A., Chen, M., May, J., & Peng, N. (2024). Are large language models capable of generating human-level narratives? Proceedings of EMNLP 2024, 17659–17681. Read in the arXiv copy.
• Todorov, T. (1977). The grammar of narrative. In The Poetics of Prose, translated by Richard Howard. Cornell University Press. The essay first appeared in French in 1968. Read in search snippets of a lending scan. His Grammaire du Décaméron (1969), which the essay summarises, was not opened.
• Toelken, B. (1996). The icebergs of folktale: misconception, misuse, abuse. In C. L. Birch & M. A. Heckler (eds), Who Says? Essays on Pivotal Issues in Contemporary Storytelling, 35–63. Little Rock: August House. Read in search snippets of a lending scan; the page of the passage quoted was not seen.
• Toubia, O., Berger, J., & Eliashberg, J. (2021). How quantifying the shape of stories predicts their success. Proceedings of the National Academy of Sciences 118(26): e2011695118. Article.
• University of Chicago Chronicle (3 February 1994). Notice on Kurt Vonnegut, with a second notice of 10 May 2007.
• Uther, H.-J. (2004). The Types of International Folktales, volumes I–III. FF Communications 284–286. Helsinki: Academia Scientiarum Fennica. Not opened. Details from Berezkin (2015).
• Vijayaraghavan, P., & Roy, D. (2023). M-SENSE: modeling narrative structure in short personal narratives using protagonist’s mental representations. arXiv:2302.09418. A preprint. The abstract, the section on the data and the first two tables were read.
• Vogler, C. (1985). A practical guide to Joseph Campbell’s The Hero with a Thousand Faces. Memo of seven pages, written at Walt Disney Pictures. Copy with the author’s later preface.
• Vogler, C. (1998). The Writer’s Journey: Mythic Structure for Writers, second edition (first edition 1992). Read in search snippets of a lending scan.
• Vonnegut, K. (1981). Palm Sunday: An Autobiographical Collage. Chapter XVIII, “The Sexual Revolution”, read in a scan of the 1982 Dell paperback, pages 311 to 315.
• Vonnegut, K. (2005). A Man Without a Country. Seven Stories Press. The passage with the five drawings is excerpted as “At the Blackboard” in Lapham’s Quarterly, which is where it was read; it was checked against snippets of a scan of the book.
• Wilmot, D., & Keller, F. (2020). Modelling suspense in short stories as uncertainty reduction over neural representation. Proceedings of ACL 2020, 1763–1788. Read in the arXiv copy.
• Wu, W., Wang, L., & Mihalcea, R. (2023). Word category arcs in literature across languages and genres. Proceedings of the 5th Workshop on Narrative Understanding, 36–47. Paper. The abstract and sections 3 to 5.1 were read.