Shapes of Stories

Each line is one complete reading. The story was cut into 12 to 30 scenes and an AI reader went through it in order, marking where the protagonist's fortune stood at the end of every scene, from −5 (catastrophe) to +5 (bliss), without looking ahead. Pick one to put it on the board.

The paper's arithmetic holds up. Rerunning it on the same books (1,325 of its 1,327 can still be downloaded as text) gives the same patterns in the same order: the first three carry 70% of all the variation between books, and each of them, right side up or upside down, is one of the six named arcs.

The trouble is what else produces them. Shuffle every book's words into random order, so that no plot survives, and run the same analysis: the same six shapes come out, carrying 59%. Replace the books with random numbers: the same six, 60%. The shapes are what a 10,000-word moving average followed by a principal-components analysis draws from any sequence at all. They are the first three waves of a cosine series, and statisticians have known since the 1920s that moving averages conjure waves out of randomness.

None of this is new. Scott Enderle showed the same thing with random walks within weeks of the paper appearing in 2016, and a 2023 paper in PNAS showed why smooth data does this in general, calling the result "phantom oscillations". The only published reply is one paragraph in the lead author's doctoral thesis. The paper has been cited several hundred times since, mostly without the caveat.

The arc is the one the method draws: 10,000-word window, 200 points. The paper's label comes from its supplementary table; ours is from this rerun on the 2026 text. The two agree for 56% of books, which says how firmly a label attaches to a book.

Knocking down the six shapes does not leave nothing. At the paper's window, 87% of the variation in real arcs is beyond what shuffled words produce: books have long, slow emotional structure. It just does not come in six kinds.

One feature does stand out, and it is the lopsidedness you can see in the panels above, where "Tragedy" collects three times as many real books as "Rags to riches" while noise splits them evenly. Fit a straight line through each book and 70% slope downward. The average book opens bright, slides for most of its length, bottoms out about nine-tenths of the way through and lifts on the last page.

On a bigger shelf. The paper's books are a small, hand-filtered sample, so both tests were rerun on every English-language work of fiction that could be fetched from Project Gutenberg: 21,529 books. The answers hold. 69% slope downward, against 50% of their shuffled copies, and among the 7,215 novels that can be cut cleanly from the first chapter to "The End" it is 74%. Non-fiction does not do this: 54% of 2,163 histories, biographies and travel books fall, and 51% of 2,631 books of science, philosophy and instruction. The darkening belongs to fiction, not to long books in general. The six-shapes arithmetic also comes out the same way at this scale: in the 19,513 of these books long enough for the paper's window, the first three patterns carry 63% of the variation, and in their shuffled copies 51%.

The paper could only count words. A language model can read. So 34 famous stories were read from start to finish by an AI reader that rated, scene by scene, both the tone of the writing and the fortune of the protagonist, which is the thing Vonnegut's axis is about. A second reader, a different model, independently re-read eight of them.

Three things came out. The readers are consistent: the two agree on fortune at r = 0.94. The word list does hear the mood of a scene: it tracks the reader's sense of tone at a median r = 0.52. But mood is not fortune: against the protagonist's fortune the word list manages only r = 0.31. A scene can be bleak while things are going well, or sparkle while they fall apart, and comedy defeats the method entirely.

And the six labels do not survive contact with a reader. Label each story by the paper's rule and the result matches the shape the reader named in 8 of 33 stories. One reason is on the board above: a 10,000-word window is a third of A Christmas Carol, so the graveyard and Christmas morning are averaged into one point.

There is also a twist on the falling-arcs result. In these 34 stories the word list slopes downward in 71%, the same lean as the big corpus. But the reader has the protagonist ending better off than they began in 56%. Stories get darker in vocabulary as the stakes rise, whether or not they end well. The downward drift is a fact about how stories are told, and it is a mistake to read it as "most stories are tragedies".

Six shapes was the wrong answer, but the question under it is a good one: is there anything that stories everywhere do? Two tests follow, and neither counts happy and sad words. The first sends an AI reader through 319 folk tales from seven regions of the world and compares every tale with itself, its own events put in shuffled order. The second counts where ordinary words fall in 7,215 novels.

The short answer is yes, and what is shared is trouble and timing, not shape. Nearly every tale starts with something wrong (97%), has its protagonist in real trouble at some point (86%), raises the tension and then lets it go (94%), and ends somewhere decisive (85%). Each of the four is true of most tales in every region sampled; the lowest figure for any region is 73%. Two things often assumed to be universal are not. The happy ending runs from 85% of the European tales to 33% of those from the Americas. And no shape comes close to dominating: the commonest fits 32% of tales.

Folk tales are the nearest thing there is to a world sample of plots: short, complete, and shaped by retelling long before anyone wrote a manual on structure. Project Gutenberg holds English translations of collections from most of the world. From 29 of them 939 single tales were cut out, and 319 were drawn at random: 80 from Europe, 60 from the Americas, 40 each from West Asia, East Asia and Africa, 30 from Oceania and 29 from South Asia.

Each tale was cut into eight scenes of equal length. An AI reader marked, for every scene, where the protagonist's fortune stood at its end (−5 to +5) and how much hung in the balance (tension, 0 to 5), then recorded how the tale ends and which of sixteen plot elements it contains. A second reader, a different model working alone, rated 48 of the same tales. The two agree closely on fortune (r = 0.96, with 94% of scenes within one point) and on how a tale ends (92%), and less well on its overall shape (73%).

A claim about the order of events counts here only if it beats a shuffle. Each tale is compared with itself: the same rises and falls put in random order, so that the tale keeps its start, its end and the size of every move, and loses only its sequence. Whatever real tales do more often than their shuffled selves is the work of order, which is to say of plot.

Every tale the reader rated is here. Pick one to put it on the board at the top of the page, with its eight scenes, its tension and the word-list line. The word list does no better on folk tales than on the classics: it follows the reader's sense of tone at a median r = 0.41 and the protagonist's fortune at 0.26, and it has 63% of tales sloping downward although 64% end better.

The second way needs no reader and no feelings, only a count of where ordinary words fall. In 7,215 novels cut cleanly from chapter one to "The End", each family of words below was counted in a hundred slices of the book and compared with the book's own average rate. Shuffle a book's words and every row goes flat. Three of the rows (scene-setting grammar, action grammar and thinking words) repeat a 2020 study of 40,000 stories by Boyd, Blackburn and Pennebaker; the other fourteen families were chosen for this page.

What holds everywhere is close to what makes something a story at all: a problem, real trouble, tension that mounts and is settled, an ending that decides the matter. That is thinner than six shapes or seven basic plots, and firmer. It fits the best-supported published work. Boyd and colleagues found that stories set the scene first and then move into action, and Wu and colleagues found the same in five of six other languages. Boyd also found both movements in science journalism and in talks, and what set stories apart there was a peak of thinking words in the middle. In the count above, book-length non-fiction shares the opening scene-setting and shows no build-up of action or of thinking words. Tehrani, and da Silva and Tehrani, showed that particular tales can be traced across cultures like family lines.

That last finding is also a warning. Tales that resemble each other in Europe and West Asia may share ancestors, so agreement between neighbouring regions is weaker evidence than it looks; the agreement that counts most here is between regions far apart. The other warnings are about the sample. These are English translations chosen by collectors about a century ago, several of them retold for children. A tale no collector thought worth printing is not in it. And the readers are AI models, checked against each other and not against people.

Red flags. None of the misconduct kind. Three of the method kind: a result that a null model reproduces; a free parameter that sets the headline number; and a critique from 2016 that was never answered in print. No correction or formal replication was found.

In its favor. It asked a real question in a testable way, shared everything, and its instrument is not empty: the word list does register the mood of a scene, and the long downward lean of books is a genuine regularity that later work has also found. Other teams using other corpora and methods have reported one, two, five, or "five to seven" shapes, which is what you would expect if the number comes from the method.

This is a first-pass assessment built on one rerun and a reading of the public critiques. Citation counts and "no reply found" are as of early October 2026 and could be incomplete.