Start with an AI story generator. Type: A woman walks into a room. She does not know why she is there. Seconds later, the machine delivers a sequence. The woman enters. She notices some object—a photograph, a letter, a stain on the carpet. That object triggers a memory. The memory supplies a motive, and from the motive a quest unspools. By the end of the third paragraph, the generator has installed a goal, an obstacle, and a quiet tick of urgency. It has turned a situation into a story without being asked, and it has done so by relying on a very specific, very old, and very ideological definition of what a story is: beats, arcs, resolution.

Treating this as a neutral technical convenience misses the real operation. The AI story generator is not a blank page. It’s a filter. It takes the vast combinatorial space of possible narrative forms and shrinks it to the subset the model was trained to recognize—training scraped from screenwriting manuals, genre templates, film synopses, and the accumulated sediment of already-told stories. What the generator refuses to write is not a glitch. It’s a structural exclusion, and that exclusion carries a politics.

I want to read the AI story generator as a formal technology of constraint, one that belongs in the same critical lineage as censorship boards, production codes, and state script committees that have historically decided which stories reach the screen. Not because the algorithm has intentions, but because it has defaults. At scale, defaults function as law.

The Beat Sheet as Ideology

In 1979, Syd Field published Screenplay: The Foundations of Screenwriting, codifying the three-act structure and a paradigm of plot points that hardened into industry standard. His model was originally descriptive—he analyzed films that already existed—but it became prescriptive almost immediately. By the 1990s, the Save the Cat beat sheet had metastasized into a fifteen-point structural template. Script readers used it as a checklist. A screenplay that missed the beats registered as a failure. This is the tradition that StudioBinder’s screenplay guide reproduces when it explains that “one page of script format equals roughly one minute of screen time” and that “proper screenplay structure is crucial.” The guide is useful, clear, and entirely upfront about its assumptions. But those assumptions deserve a pause. The page-to-screen-time equation naturalizes a particular narrative rhythm, a tempo that feels “right” because it has been repeated thousands of times. It isn’t the only rhythm. It’s simply the one that got institutionalized.

When an AI story generator produces output, it channels this institutionalized rhythm. Ask it for a story, and you’ll get a protagonist with a want and a need. A midpoint reversal. A climax where the protagonist confronts the antagonist and either succeeds or fails, but always changes. The arc is mandatory. The arc is moral. It assumes that human experience is structured like a problem to be solved, and that narrative exists to deliver the solution.

But experience doesn’t always work that way. A life, or a moment, sometimes simply refuses to resolve. The films of Apichatpong Weerasethakul offer a counter-model. In Uncle Boonmee Who Can Recall His Past Lives (2010), narrative unfolds not through cause-and-effect chains but through drift, association, and the logic of dreams. A ghost sits down at a dinner table. A princess seduces a catfish. The film doesn’t explain these events; it arranges them in a pattern that rewards perceptual attention rather than plot-tracking. Weerasethakul has described his scripting process as a form of recollection—writing down dreams, memories, and images, then assembling them without forcing a dramatic arc. The resulting film is legible as cinema but illegible to a beat sheet. Feed the prompt A ghost appears at a dinner table into an AI story generator, and the output will likely be a story about a haunting that demands resolution: a mystery, a buried trauma, a confrontation. What it won’t produce is the ghost simply sitting down, eating, and leaving. That doesn’t register as a “story” within the model’s parameters.

This isn’t a technology failure. It’s an ideological success. The generator has been trained to recognize story as a particular shape, and it polices that shape with every output. It functions as a censor—not by prohibition, but by probability. It cannot write what it cannot see.

The Oulipo Constraint and the Algorithmic Cage

Constraint has a long, productive history in art. The Oulipo group—founded in 1960 by Raymond Queneau and François Le Lionnais—used mathematical constraints as generative devices. Georges Perec wrote an entire novel without the letter e. Italo Calvino built a narrative machine from tarot cards. In these cases, constraint wasn’t a limitation; it was a liberation from habitual patterns. The crucial difference: Oulipo constraints were chosen. They were visible, arbitrary, subject to the writer’s control. The constraints of an AI story generator are invisible, naturalized, and not subject to the writer’s control—unless the writer knows enough about the system to hack it, and most don’t.

Chilean filmmaker Raúl Ruiz, in his Poetics of Cinema, proposed a different kind of constraint: a “shamanic” method of constructing films through random encounters, aleatory combinations, and a refusal of central conflict. His 1978 film The Hypothesis of the Stolen Painting is structured as a series of tableaux vivants that a narrator interprets, misinterprets, and reinterprets. The film doesn’t resolve. It accumulates. It is a story generator of a different order—one that produces not closure but proliferation. Ruiz wrote against what he called the “central conflict theory” of American screenwriting, which he diagnosed as a “police theory” of narrative because it arrests the drift of images and forces them into a single, authoritative line.

The AI story generator is the apotheosis of central conflict theory. It cannot drift. It cannot proliferate. It cannot hold open the space of ambiguity that Ruiz and Weerasethakul cultivate. And because the generator is increasingly positioned as a tool for “idea generation” and “story development”—a pre-visualization device for the cinematic imagination—its constraints risk becoming the industry’s constraints. The more writers rely on generators for first drafts, the more those first drafts will converge on the same structural defaults. Narrative diversity contracts, not because anyone banned it, but because the tool that promises efficiency encodes one model of story as the only model.

The Material Infrastructure of Exclusion

To understand what the AI story generator refuses to write, look at its material infrastructure. The training data is not a neutral archive of human storytelling. It’s a dataset scraped from the internet, weighted toward English-language texts, genre fiction, screenwriting blogs, and film databases. The labor of prompt engineers—the humans who refine the model’s outputs—is itself a form of editorial gatekeeping. A prompt engineer decides which outputs are “good” and which are “hallucinations” or “nonsense.” But the criteria for “good” are inherited from the same manuals and templates that trained the model. A hallucination might be a genuinely novel narrative form, but the system is designed to suppress it in favor of coherence, where coherence means conformity to the expected pattern.

The Authors Guild’s AI best practices document addresses this from the angle of creative control and copyright, noting that “the labor of prompt engineers and the infrastructure of AI are ideological and economic factors in storytelling.” The document stresses the systemic impact of generative AI on cultural production—who gets paid, who gets credited, whose work is used for training without consent. But the question of form is equally urgent. The economic incentives of the AI industry push toward scale and standardization. A generator that produces reliably coherent outputs has more commercial value than one that produces strange, ambiguous, or unresolved texts. The market reinforces the ideological defaults. The algorithm is not just a censor; it’s a censor with a business model.

What the Generator Refuses: A Close Reading

Consider a specific example. I prompted a widely available AI story generator with the following: A man stands at a window. He does not move. Nothing happens. The generator produced a story about a man grieving his wife, remembering their life together, and finally walking away from the window to begin healing. Stillness was converted into backstory, backstory into an arc, arc into resolution. The prompt had explicitly asked for nothing to happen, but the generator couldn’t comply. It had to make something happen, because its model of story rests on the premise that narrative is change over time, and change is always directional.

This isn’t a trivial limitation. Entire cinematic traditions are built on the refusal of directional change. The films of Chantal Akerman, particularly Jeanne Dielman, 23 quai du Commerce, 1080 Bruxelles (1975), accumulate duration and domestic repetition until the breakdown, when it arrives, registers not as resolution but as rupture. The film’s politics reside precisely in its refusal to conform to the beat sheet. Akerman’s long takes and static frames are not a style; they are an argument about women’s time, domestic labor, and the violence of narrative expectation. An AI story generator could not produce Jeanne Dielman. It could not even recognize it as a story. The generator’s model of story is, in this sense, a model of exclusion that maps onto gender, culture, and form. What it refuses to write is not random. It refuses the feminine, the non-Western, the non-goal-oriented, the unresolved—not because it has been programmed to reject these things, but because the training data has already marginalized them.

Comparison with Historical Censorship Structures

The comparison with censorship boards is not metaphorical. The Hays Code, enforced in Hollywood from 1934 to 1968, was a set of rules that determined what could and could not be shown on screen. It prohibited “lustful kissing,” required that crime be punished, and mandated that the institution of marriage be upheld. These rules were explicit, but their deeper function was to enforce a particular moral and narrative order. The Production Code Administration was a script committee that read screenplays and demanded revisions. The AI story generator performs a similar function, but preemptively. It doesn’t need to read a finished script and demand changes; it generates the script already in compliance with the unstated rules of narrative coherence. The censorship is built into the output from the start.

State script committees in the Soviet Union and contemporary China operate on a similar principle. They evaluate scripts for ideological conformity—not just political content, but formal conformity to the socialist realist model of narrative. Socialist realism demanded positive heroes, progressive development, and a clear resolution that affirmed the state’s narrative of history. The AI story generator, trained on the global corpus of commercial storytelling, demands its own version of positivity: a protagonist who wants something, a conflict that escalates, a resolution that delivers emotional payoff. The form is different, but the function is the same. Both systems naturalize a particular model of story and render other models illegible.

The Interface as Ideological Surface

The interface design of AI story generators is itself a site of ideological work. Most generators present a blank text box and a button labeled “Generate” or “Write.” The blankness suggests openness, but the underlying architecture is closed. The user is invited to type anything, but the system will only return what it can recognize. This asymmetry is hidden behind the interface’s apparent neutrality. The user doesn’t see the training data, the model weights, or the fine-tuning parameters. The user sees a magic box that produces stories. And because the stories are coherent, they feel like the story, not a story. The interface naturalizes the output.

Contrast this with the working methods of a filmmaker like Apichatpong Weerasethakul, who begins not with a text box but with a notebook, a dream diary, a collection of found objects. His method is open in a way the generator is not, because it doesn’t pre-filter experience through a story template. The generator, by contrast, is always already filtering. The question it answers is not “What can be told?” but “What can be told as a story?” That “as a story” is the ideological operation. It is the point at which the algorithm becomes a censor.

Toward a Critical Practice of Algorithmic Refusal

So what is to be done? One response is simple refusal—to write against the tool, cultivating the narrative forms it cannot recognize. This remains the path of Ruiz and Weerasethakul, and it remains viable. But refusal alone doesn’t address the systemic drift toward algorithmic defaults in an industry that increasingly adopts these generators as pre-visualization devices. A more engaged critical practice would treat AI outputs not as drafts to be refined but as texts to be read against the grain. What did the generator exclude? What did it smooth over? Where did it force resolution? Reading the output diagnostically—as a document of the anxieties that produced it—opens a space for critique. The generator’s refusal becomes a site of inquiry. Why can’t it write stillness? Why does every character need a want? Why does every story need an ending?

These questions are not merely technical. They are political. They go to the heart of what cinema has been and what it might become. The moving image has always been a technology of memory, a way of constructing pasts and futures that exceed the documentary record. The AI story generator threatens to close that openness by training the imagination on a single, optimized model of story. The resistance is to insist on the forms that fall outside the model—the drift, the fragment, the unresolved image, the story that refuses to become a story. That resistance is not nostalgia for a pre-digital past. It is a demand that the future of narrative remain contested, multiple, and strange.