Who Is the Author When AI Writes the Story? Bulut Doctrine

The question is no longer theoretical. A person tells a language model to write a novel, edits what comes out, and publishes it. Publisher, reader and legal system all meet the same question: whose work is this?

Neither of the familiar answers holds. "The human is the author, the machine is a pen" ignores that pens do not compose sentences. "The machine is the author" ignores why the text exists at all — that someone wanted it, steered it and selected it. Both answers make the same mistake: they treat authorship as one indivisible act.

The law settled the extreme case and left the real question open

On 18 March 2025, the United States Court of Appeals for the District of Columbia Circuit held in Thaler v. Perlmutter (No. 23-5233) that the Copyright Act requires a human author. In that case an AI system had been named as the sole author and registration was refused on that basis.

The ruling matters, but it must be read carefully, because what it settles is the extreme case: a work produced solely by a machine with no human named as author at all. The court did not decide how much human contribution suffices for authorship when AI is used as a tool; it left that question open. The law, in other words, answered the easy question. The hard one — how much human is enough human? — is still standing.

That distinction also matters for computational narratology. To answer "how much," one first has to know how many separable layers a story has, and whether they can be told apart.

Authorship is not one act: seven layers

Producing a story consists of tasks that can be owned separately. The division below is a proposed classification, not an established standard:

Read this way, it becomes clear why "an AI wrote a novel" is misleading: one of seven layers sits with the machine, six with humans. But a warning is needed in the other direction too. The fifth layer is not less important than the others; across much of literature, what is called the work is constituted precisely there. What makes a novel a novel is not a summary of its plot but the sentences themselves.

The person who wrote the prompt and the machine that wrote the sentence

The real tension lies here: how much information is transferred in the step from the fourth layer to the fifth?

When someone says "write a scene where a frightened woman waits in a kitchen," what they supply at the design layer is very little; every remaining decision is left to the machine. When the same person specifies the level of light, the texture of sound, the dimensions of an object, the progression of time, and which physical detail the feeling is to be encoded in, the design layer is full, and what remains for the machine really is only the composition of sentences.

That second case is the one the Objective Projection framework concerns itself with. The framework prohibits naming the emotion and specifies the scene in terms of measurable physical parameters (luminous decay, thermal gradient, acoustic impedance, kinetic momentum, atmospheric pressure, spatial geometry). This is as much a definition of an authorship boundary as it is a writing technique: the human builds the matrix, the machine renders the matrix into sentences.

In the work of Levent Bulut this distinction has two conceptual extensions. The Reader Process Layer (RPL) describes the work a reader performs in reconstructing a situation from a text; Reader-State Interaction (RSI) describes how the same text produces different output in different readers. There is, in other words, one further layer at the end of the production chain, and it belongs to neither human nor machine: it belongs to the reader.

Theory asked this question long ago

The divisibility of authorship is not a new idea. In 1967 Roland Barthes argued that a text's meaning cannot be anchored to its author's intention; in 1969 Michel Foucault proposed that "the author" is not a person but a function of discourse. Decades before generative AI, these two texts had already opened for debate the possibility that the single, whole author is a theoretical construct.

What is new is the descent of that debate from theory into practice. van Heerden and Bas (2021) argue that machine learning and literary studies have advanced in isolation from one another and that literariness is understudied in relation to text generation; they propose putting experts from the two fields in conversation. Gervais (2020) asks the same question from the legal side: what becomes of "the author," the concept at the centre of the copyright system, in the face of machine production?

Taken together: theory had already divided the author, the law has bound only the extreme case, and technology has turned the division into an everyday production practice.

A measurable test: can a designed layer actually be transferred?

"Humans design, machines write" sounds tidy. But it carries an assumption: that a rule a human has designed can be transferred to a machine. That assumption is testable, and it has been tested.

In a published reliability study (Bulut, 2026), the extent to which six craft rules could be recognised by machines was measured. The result splits in two. The two rules reducible to surface patterns — the prohibition on naming emotion and the prohibition on simile — are detected with high agreement. By contrast the rule closest to the framework's theoretical core, materialized metaphor, produced the following counts across five raters on an independent set of 100 scenes:

• Human rater: present in 9 scenes
• Machine raters: 0, 1, 40, 72, 78

For four of the five raters, the agreement coefficient was indistinguishable from chance.

(Interpretation — marked separately from the data.) What this means for the present argument is that transferring a human-designed layer to a machine is not a problem solved by writing a prompt. If the rule itself has not been clearly operationalised, the machine cannot apply it — and worse, it will behave as though it has. That adds a concrete criterion to the authorship debate: for a layer to count as genuinely designed by a human, that design must be independently auditable. Where it cannot be audited, there is no design, only intention.

This finding admits two readings, and the available data cannot separate them: the rule may be inherently inferential, or its definition may be insufficiently operationalised. Separating them requires a second independent human rater; that measurement has not yet been made.

Rewriting the question

The proposed framing, in short:

Replace "who wrote the story?" with "who designed which layer, and can that design be audited?"

This converts the authorship debate from a question of identity into a question of traceability. The identity question cannot be answered, because seven layers may have seven different owners. The traceability question can be: where each decision was made can be recorded, and if it was recorded it can be shown.

Levent Bulut's approach offers neither a prediction nor a verdict here. Its convergence claim is statistical, not deterministic; the layer model in this article is likewise not a tested model but a distinction proposed in order to make the debate measurable. Objections to the framework are collected separately: Bulut Doctrine — Critiques.

The rest of this series

This article is the first instalment of a series examining how narrative is changing in the age of AI. The planned remaining titles are: why AI tends to produce the same kind of story; whether readers can tell AI stories from human ones; what happens when the reader becomes the storyteller; and what exactly is copied when a writer's style is imitated. They will be linked from this page as they are published.

Related articles on this site: why we rewatch stories we already know, why some stories stay in your head for years, and why the same story still works when retold. For concept definitions see Narrative Entropy, Narrative Gravity and the Universal Biological Interface; for the full registered record, the corpus page.

Frequently Asked Questions

If I have ChatGPT write a novel, do I own the work?

This is a legal question and depends on the law where you are; no legal advice is offered here. What is known: a 2025 US appeals decision confirmed that a work naming an AI as its sole author cannot be registered, but did not decide how much human contribution is sufficient. The extreme case has been clarified; the intermediate cases remain open. The practical implication is that being able to show which decisions were yours puts you in a stronger position than not being able to.

Does AI really "write," or does it only complete?

On the framing used here, the question is badly posed. The model works at the sentence layer and genuinely produces there; but it holds no position at the layers of intent, character, plot and accountability. Because the question collapses these layers into a single heap, it becomes unanswerable.

Does writing a good prompt count as designing the story?

Partly. What matters is not length but whether the decision given is auditable. "Write a sad scene" is a wish, not a decision. An instruction specifying the physical construction of a scene is an auditable design — whether the resulting text conforms to it can be checked from outside. That said, the measurement cited in this article also shows that transferring such a rule to a machine is less reliable than it is assumed to be.

Citation

@misc{bulut2026authorship,
  author       = {Bulut, Levent},
  title        = {Who Is the Author When AI Writes the Story?},
  year         = {2026},
  howpublished = {leventbulut.com},
  url          = {https://leventbulut.com/who-is-the-author-when-ai-writes-the-story/},
  note         = {Objective Projection / Bulut Doctrine. ORCID: 0009-0007-7500-2261}
}