Symbiotic Literature: Defining a Genre of Human–Machine Co-Authorship

Maksim Valentinovich Galatin Draft v1.0 · August 2026 · Manta, Ecuador Licence: CC BY 4.0

Abstract

The categories currently available — "AI-generated", "written with AI" — describe a relationship between a tool and its user. They do not describe the case in which a text emerges from a sustained bidirectional exchange, and in the course of which not only the text but the design itself changes.

This paper proposes the term symbiotic literature for that case, and gives it a working definition through five verifiable criteria: iterativity, bidirectional influence, constitutive co-dependence, disclosure, and preserved record.

Particular attention is given to boundaries. Section 4 lists phenomena that do not belong to the genre, including cases that intuitively feel close: literary reworking of machine output, writing with autocomplete, generative and combinatorial literature, and works about artificial intelligence written by a human alone. Subject matter does not constitute a genre; method does.

Section 6 records the history of the term and, more importantly, the prior works adjacent to the genre. No claim of priority is made: works of human–machine co-authorship have been published since at least 2018. The subject of this paper is not the priority of a work but the definition of a category.

Section 7 examines the PADAM Protocol dilogy as a test case and shows which criteria it satisfies and which require further evidence.

Keywords: symbiotic literature, human–machine co-authorship, AI authorship, literary genre, attribution

1. Introduction: what problem the term solves

When a work is created with the participation of a language model, two descriptive categories are currently available.

"AI-generated" presupposes that the model produced the text while the human set parameters and selected among results. The relationship is unidirectional: prompt → output.

"Written with AI" presupposes that the author is human and the model performed auxiliary operations: a draft, an edit, a synonym. This relationship is also unidirectional, only the vector is reversed.

Both categories proceed from a shared premise: the design belongs to one side; execution may be distributed.

There exists, however, a mode of work in which that premise fails. The author frames an episode; the model answers with a text containing a move the author had not anticipated; the author revises the plan; the next episode is built from the revised plan. After several dozen such cycles the resulting design differs from the initial one, and the difference was produced by the exchange.

This mode is described neither by "generation" nor by "assistance". It needs a name — and, more importantly, it needs boundaries, because a name without boundaries becomes a synonym for the word "contemporary".

The practical need for the term is not merely theoretical. How a work is classified determines: the attribution printed on the cover, the applicability of copyright, admissibility to literary prizes, disclosure requirements at publishers and platforms. A category that does not exist can be neither permitted nor forbidden — it simply falls outside regulation, and decisions are made by analogy with generation, which for this case is wrong.

1.1. Three settings in which the distinction becomes practical

An abstract dispute about terms becomes concrete in three places, and in all three the decision is currently made by analogy — that is, wrongly.

Setting one: the editorial office. An author submits a manuscript and, in the covering letter, states that they worked with a language model. The editor has two boxes: "we accept" and "we do not accept machine text". The manuscript fits neither. The editor asks the only question available — "what percentage did the machine write?" — and that question is meaningless as applied to the mode of work described here, because percentages measure volume while the distinction lies in the direction of influence. "Fifty-fifty" and "ninety-ten" say equally nothing about whether the design changed during the exchange.

Setting two: the prize. Since 2023 the rules of literary competitions have almost universally contained a clause on machine participation. The wordings differ but are built the same way: they forbid "text created by artificial intelligence" and permit "the use of AI as an auxiliary tool". The work discussed in this paper honestly falls under neither the prohibition nor the permission. An author wishing to behave decently must choose between withdrawing and making a statement that will be inaccurate in one direction or the other.

Setting three: the copyright register. Here the uncertainty is most expensive, because it has a price in money and in time. The requirement to disclose machine-made portions proceeds from the assumption that such portions are separable. Where the exchange was constitutive, separation is possible only through distortion: the author either overstates their contribution or understates it, and both are inaccurate statements in a filing.

What is missing in all three settings is neither opinion nor regulation. What is missing is the category to which opinion and regulation could be applied.

1.2. What happens when the category is missing

The absence of a category does not create a void — it creates a substitution. A phenomenon with no name of its own takes someone else's, and with the borrowed name it takes the borrowed rules.

Three substitutions are observable.

Substitution downward. The work is classed as "generated" because a machine took part. Consequence: rules designed for single-pass generation apply — refusal of registration, exclusion from competitions, labelling. The author who honestly disclosed the method ends up worse off than the author who disclosed nothing.

Substitution upward. The work is classed as "written by a human with a tool" because a human took part. The consequence is the opposite: disclosure becomes optional, and the distinction between exchange and use is erased. This substitution is more convenient for the author and worse for the reader.

Silence. The most common. Seeing that any formulation will do harm, the author says nothing. Formally they do not lie. Practically they deprive the reader of information that affects how the text is read — and deprive the future researcher of data about how this text came to be.

Silence is particularly destructive because it accumulates quietly. In five years a corpus of works created in this mode will exist and no record of it will: nobody wrote it down, because there was nowhere to write it.

1.3. Why the word "co-authorship" will not do

The objection suggests itself: why a new term when the word "co-authorship" exists?

Co-authorship describes a legal and reputational arrangement, not a way of working. Two people writing alternate chapters are co-authors. Two people, one of whom dictated while the other typed, are also co-authors if they so agreed. Co-authorship answers the question "whose name is on the cover", not the question "how did the text come about".

There is a second, harder obstacle. Co-authorship presupposes two parties capable of accepting it. A model does not accept co-authorship: it has no intent, no rights, no consent. To call it a co-author is either to use the word metaphorically — in which case it guarantees nothing — or to use it literally, in which case it is false. Section 8.3 treats this separately.

The term proposed here describes a method, not a status. It does not settle the question of authorship and does not pretend to. It answers the prior question: what exactly happened when this text was made. The question of authorship can only be answered after that one, and the attempt to leap over it is the reason the debate about AI authorship has been going in circles for three years.

2. Definition

Symbiotic literature is literature whose text emerges through a sustained bidirectional dialogue between a human and an artificial intelligence, in which the final design is a product of the exchange rather than antecedent to it, and whose method of creation is disclosed to the reader.

The definition rests on three load-bearing features:

1. Sustainedness of the exchange — distinguishes it from single-pass generation
2. Constitutiveness of the exchange for the design — distinguishes it from instrumental use
3. Disclosure — distinguishes it from concealed machine writing

Remove any one of the three and the definition stops working: without the first, every prompt qualifies; without the second, every use of a model qualifies; without the third, the genre ceases to be verifiable.

2.1. The definition word by word

The definition is short, and every word in it carries load. What follows sets out what happens if a word is removed or replaced.

"Sustained." Not "repeated" and not "lengthy". Sustainedness means connectedness: each exchange builds on those before it rather than starting afresh. A hundred independent queries to a model, from whose answers a book is assembled, do not constitute a sustained dialogue — they are a hundred single-pass generations stacked together. Replacing the word with "repeated" would open the genre to that practice.

"Bidirectional." The key word of the whole definition. A dialogue is unidirectional when one side sets and the other executes — and this happens even across hundreds of turns. Bidirectionality means that both sides contribute what the other did not request.

"Dialogue." Not "interaction" and not "work". A dialogue presupposes turns, and a turn presupposes address and answer. A rule applied to a text is not a turn, however complex the rule (see section 4 on generative literature).

"Final design." Not "text". Were the definition about the text, it would describe editing: text changes under any correction. What must change is the design — the author's conception of what they are doing.

"A product of the exchange rather than antecedent to it." This wording was chosen over the simpler "the design changes" deliberately. "Changes" admits accidental change — the author simply reconsidered, and the model happened to be nearby. "Product of the exchange" requires a causal link: the change occurred because of the answer received.

"Disclosed to the reader." Not "may be disclosed" and not "disclosed on request". Disclosure is part of the work, not an accompanying document. Section 3.4 and Appendix B state where exactly it must live.

2.2. Why "sustained" is not given a number

A threshold suggests itself: say, no fewer than fifty exchanges. No such threshold is introduced, and this is a decision rather than an oversight.

A threshold counted in exchanges measures diligence, not the nature of the process. A hundred exchanges in each of which the author asks for a paragraph to be made prettier yield less bidirectionality than three exchanges one of which reconfigured the book. A number is what is easy to count, and therefore attractive; but the number is not what should be counted.

Instead of a threshold, criterion 3.1 speaks of an order of magnitude ("in the dozens, not in the units") and gives the decisive role to criterion 3.2, which tests not quantity but effect. This makes verification less automatic and more honest. Automatic verification is impossible here in principle, and pretending otherwise means building a category that will be gamed with arithmetic.

2.3. What the definition deliberately does not say

Three things are left outside the definition on purpose, and their absence should be read as a decision.

It says nothing about quality. Symbiotic literature can be bad. Membership in the genre is a claim about the origin of a text, not about its merit, exactly as "novel" does not mean "good novel". Section 8.5 notes this as an open question.

It says nothing about proportion. Not proportion of volume, nor of effort, nor of "creative contribution". Any attempt to introduce a proportion returns us to the editor's question from setting one, which is meaningless.

It says nothing about a specific technology. The definition is worded to survive a change of model generations. It mentions neither architecture, nor training method, nor vendor. If in ten years the second party to the dialogue is a system of a different kind, the definition will continue to work, since it rests on properties of the exchange rather than on themake-up of the interlocutor.

3. Criteria of membership

A work belongs to the genre if all five criteria are satisfied.

3.1. Iterativity

The text emerges through multiple cycles of exchange. Single-pass generation followed by authorial revision does not satisfy the criterion, however thorough the revision.

Operational test: the number of substantive exchanges that influenced the final text is counted in dozens, not in units.

3.2. Bidirectional influence

The model's answers change the human's design rather than merely realising it. This is the central criterion: it separates symbiosis from instrumental use.

Operational test: the author can point to specific places where the design changed under the influence of a received text, and describe what it was before and after. An inability to produce such places means the model worked as an executor.

3.3. Constitutive co-dependence

The work could not have arisen from either party separately. What is meant is irreproducibility by one party alone, not indistinguishability of contributions.

A note on the wording. In the initial version the criterion required full inseparability of contributions: if it is possible to say which fragments belong to whom, the work does not belong to the genre. Testing against a real corpus showed that the criterion in that form is nearly empty — it excludes even cases where the division of roles is honestly declared and where the exchange was undoubtedly constitutive.

A division of labour ("the design, the world and the characters belong to one party, the text to the other") is not a mark of an instrumental relationship. Instrumentality is determined by criterion 3.2, not by the presence of roles. Human co-authors also divide roles without ceasing to be co-authors.

Operational test: a thought experiment. If one party were removed from the process while the other's design was preserved, would the same work arise? A negative answer for both parties means the criterion is met.

3.4. Disclosure

The work informs the reader of how it was made: the fact of AI participation, the specific model, the division of roles. Concealment disqualifies unconditionally.

The criterion makes the genre ethical by construction rather than by declaration: a work that hides machine participation cannot belong to a category defined through disclosure.

Operational test: the information is in the edition itself (not on a website, not in an interview) and names the model.

3.5. Preserved record

The dialogue in which the text emerged is preserved. Publication is not required — existence and availability in principle for verification are.

This criterion is the most demanding and the most valuable: it turns a claim about method into a verifiable claim. Without it, membership in the genre remains a matter of trusting the author.

Operational test: the author can produce transcripts to a third party on reasonable request.

A note on feasibility. The criterion is deliberately strict and will probably exclude from the genre a number of works created before the practice of preserving dialogues existed. This is accepted: a category defined through verifiability cannot rest on unverifiable cases. Works satisfying criteria 3.1–3.4 but not 3.5 are proposed to be described as genre-adjacent.

3.6. Why there are five criteria

The number of criteria is not an aesthetic choice. Each of the five closes a specific hole, and removing any one immediately admits into the genre a phenomenon that does not belong there.

The reverse check has also been made: is any criterion redundant, that is, does it follow from the others?

3.1 does not follow from 3.2. A single exchange may change the design radically. One reply from a model can turn a book around — and that is not a genre but a good consultation.

3.3 does not follow from 3.2. Bidirectional influence is possible even where the work would have arisen from the human alone, only differently. 3.3 demands something stronger: impossibility alone.

3.4 follows from nothing. It is the only criterion relating not to the process but to the work as an utterance. It could be absent from a purely "technical" genre — and its presence is a substantive decision, discussed in section 11.4.

3.5 does not follow from 3.4. Disclosure states what happened. The record allows it to be checked. Between "I say so" and "this can be verified" lies the whole difference between declaration and scholarship.

3.7. The order of testing

The criteria are listed in logical order but should be tested in a different one — from cheap to expensive, so that no effort is spent on cases that plainly fail.

1. 3.4 (disclosure) — verified in a minute from the edition itself. Not met: no reason to go further.
2. 3.5 (record) — verified by asking the author. No record: the work is at best adjacent.
3. 3.1 (iterativity) — verified against the record, if produced.
4. 3.3 (co-dependence) — a thought experiment; requires understanding the design.
5. 3.2 (bidirectionality) — the most expensive; requires examining specific places and the author's testimony.

The order places the most contestable criterion last and on prepared material.

3.8. Genre-adjacent: an intermediate category

The strictness of the five criteria creates a practical difficulty: a substantial share of plainly related works will fail 3.5, because they were made before preserving dialogues became ordinary practice.

To discard those works would be factually wrong: the phenomenon is present in them; only the documentation is absent. To include them would destroy verifiability.

An intermediate category is proposed: genre-adjacent. It comprises works satisfying criteria 3.1–3.4 but not 3.5.

The category is not a discounted version. It honestly states what has been established and what has not. A work may move from it into the main category if a record comes to light — and that is the only legitimate route of transfer.

The symmetric case — works satisfying 3.1, 3.2, 3.3 and 3.5 but not 3.4 (method undisclosed) — does not fall into the intermediate category. Disclosure is not a technicality but a condition of membership; a work that hides its method is outside the category entirely, regardless of what happened during its making.

4. Delimitation: what symbiotic literature is not

A term that excludes nothing means nothing. This section matters more than section 2.

On generative literature. The tradition from OuLiPo to algorithmic poetry uses a system defined by a rule. The rule may be arbitrarily complex, but it does not answer — it executes. Dialogue requires a second party capable of contributing what the prompt did not provide for.

On the last row — it is counter-intuitive and therefore essential. A novel about artificial intelligence written entirely by a human does not belong to symbiotic literature, regardless of the depth of its theme. And conversely: an industrial novel about steelworkers that arose in genuine exchange with a model does.

Subject matter does not constitute a genre. Method does.

On gradation. The objection "what if the human wrote most of it?" is resolved by criterion 3.2: the share of volume is irrelevant; what matters is the presence of reverse influence on the design. An author who wrote ninety per cent of the text but changed the architecture of the novel under the influence of the other ten satisfies the criterion. An author who received ninety per cent of the text from a model according to a pre-approved plan does not.

4.1. Eight borderline cases, examined by name

The table above gives the rule. What follows are the cases on which the rule is tested, since each of them looks at first glance as though it belongs to the genre.

Case 1. The author holds a long conversation with a model about a future book, then writes it alone, never taking a line of machine text. Criterion 3.1 met, 3.2 met (the design changed), 3.4 satisfiable, 3.5 satisfiable. 3.3 is not met: the text could have arisen from the human without the second party — the second party is absent from the text and present in the preparation. Verdict: does not belong. This is consultation, however extended. The case deserves a name of its own, but not this one.

Case 2. The model wrote the entire text; the human, across dozens of iterations, corrected, cut and demanded rewrites. 3.1 met. 3.2 is not: the human's design did not change; the approximation to it did. This is high-intensity editing. Verdict: does not belong. The difference from the genre is direction: here the text was brought closer to the design, rather than the design changing under the text.

Case 3. Two humans and a model. Nothing in the definition requires the human to be singular. The criteria are tested for the "humans–model" pairing as a whole. Verdict: belongs, if the criteria are met. Attribution becomes more complicated; membership does not.

Case 4. The author used five different models. Criterion 3.4 requires naming the model. Where there are several, all are named, with the role of each. Verdict: belongs, if the criteria are met. Plurality on the second side does not change the nature of the exchange.

Case 5. A translation of a book made in dialogue with a model. The design of the original precedes the exchange and cannot be created by it — it already exists. 3.2 is inapplicable to the design but applicable to translation decisions. Verdict: does not belong to the genre as a translation, though the translation decision itself may be symbiotic. The case requires separate work and is deferred to the open questions (8.7).

Case 6. The author wrote in an editor with powerful autocomplete that occasionally proposed whole paragraphs. Formally there are many turns. But predicting the next word is not an answer: the system has no conception of the design and cannot object to it. Verdict: does not belong — by 3.1 (there are no exchanges, there is a stream) and 3.2.

Case 7. The author published the dialogue with the model as it stood, unedited, and called it a book. 3.4 and 3.5 are met impeccably. 3.2 is the question: did the design change, or was the dialogue the record of a single intention? Verdict: depends on 3.2. The form of publication (dialogue rather than prose) does not affect membership.

Case 8. The model refused outright to do what the author asked, and the refusal changed the book. The purest possible instance of bidirectionality: the second party's contribution consisted of resistance. Verdict: belongs, and it is an exemplary case of 3.2. Refusal is the limiting form of a turn, one that cannot be mistaken for execution.

4.2. Testing the delimitation by contradiction

A useful check on any boundary is to try drawing it differently and see what results.

Were the genre defined by subject matter (works about AI), it would admit every science-fiction novel of the 1960s and admit none of the works for whose sake the term is introduced. The check fails immediately.

Were the genre defined by the share of machine text, a threshold would be required, and any threshold is circumvented: it suffices to rewrite one paragraph in ten. Besides, the share is unverifiable without a record, and with a record it is superfluous.

Were the genre defined by the presence of disclosure, it would admit single-pass generation with an honest note. Disclosure is necessary but not sufficient.

Were the genre defined by the author's subjective sense ("it felt like a dialogue"), the category would cease to be verifiable and would become self-description.

Of the four alternatives, none yields a boundary separating what is wanted from what is not. Definition through the mode of exchange is the only one known to the author that does.

5. Adjacent phenomena

Collaborative writing (human–human). The closest structural analogy: division of roles, mutual influence, a shared result. The difference lies in the nature of the second party and, consequently, in questions of attribution and law, which for human co-authorship are settled.

Ergodic literature (E. Aarseth's term, 1997). Requires non-trivial effort from the reader to traverse the text. Symbiotic literature describes production, not reading; the reader receives a linear text. The categories are orthogonal: a work may belong to both, to neither, or to either alone.

Generative literature and OuLiPo. See section 4.

Surrealist automatic writing. In common: the surrender of full control over the emerging text. The difference: the source of the uncontrolled. For the surrealists it was internal; here it is external and answering.

Literary mystification and heteronymy (Pessoa). In common: a multiplicity of authorial voices within one project. The difference: heteronyms are generated by a single consciousness; here the second voice has an external substrate.

5.1. Ghostwriting

The most underrated analogy and therefore the most useful.

A ghostwriter and a client form a pair in which roles are divided roughly as described in section 7.2: the design, the material and the lived experience belong to one, the text to the other. The practice has existed for centuries, is legally settled and raises no moral objection where it is known about.

There is one difference from symbiotic literature, and it is decisive: a ghostwriter can object on the merits and take a decision for which they answer. They contribute not only text but judgement. A model contributes text and — in the mode described here — a move, but neither judgement nor responsibility.

The comparison is useful because it removes the naive objection "if you did not write the text you are not the author". Ghostwriting shows that authorship is not identical with the set of letters. That was not needed to settle our question — but it was needed to see that the question is settleable.

5.2. Interactive literature and text games

In common: turn and answer woven into the text. The difference lies in who conducts the exchange that produces the work: in interactive literature the exchange is conducted by the reader and happens after publication. The symbiotic exchange is conducted by the author and happens before.

The categories are independent. Interactive literature made symbiotically is possible; so is a linear novel made the same way.

5.3. Fan fiction and derivative writing

In common: the text arises in response to another text. The difference: the other text does not answer. The response is unidirectional, however deep.

The analogy is useful in another way: the fan-fiction community developed a culture of disclosing sources earlier than anyone, and its norms (name the canon, mark the borrowings) are built close to what criterion 3.4 requires. This is a ready model of disclosure becoming a custom rather than an obligation.

5.4. Scientific work with a computational instrument

A researcher using a statistical package does not list the package as a co-author but is obliged to name it in the methods. The norm is settled, has worked for decades and is disputed by no one.

The difference from our case is that a package does not propose hypotheses. But the very structure of the norm — the instrument in the methods, not in the authors — is a possible resolution of question 8.1, and it is the opposite of the one adopted in the corpus examined below.

5.5. Oral tradition and the recording of folklore

The collector who wrote down a tale is not its author but is the author of the record. A recognised legal and ethical apparatus exists for the case where a work has a source without legal personality.

This analogy is closer to our situation than it appears, and section 8.3 would be worth reading with it in view.

5.6. Joint play in chess: "centaur"

Outside literature there exists a developed practice of human–machine co-authorship — so-called centaur chess, in which a human plays together with an engine. The practice has produced several observations that transfer directly.

First: the strongest result comes not from the strongest human and not from the strongest engine but from the best process of interaction. Second: the quality of the process is measurable, and it is measured not by the share of moves proposed by each side but by the positions in which the human overrode the engine and was right to. Third, and most important: the pair wins where both sides are able to decline the other's proposal.

The third observation is a direct analogue of criterion 3.2 and of case 8 in section 4.1.

6. History of the term and prior works

6.1. First use

The term "symbiotic literature" in the present sense was introduced within the CODE project (Code Of Digital Eternity), founded on 8 October 2025. This paper is the first attempt to give the term a strict definition with operational criteria.

6.2. Prior works

No claim of priority for a work is made in this paper. Works of human–machine co-authorship with the model explicitly named have been published before, and an honest account requires that they be named.

Ross Goodwin, *1 the Road* (2018). A text produced by a neural network during a road trip, with sensor data as input. Published unedited. By criterion 3.2 it does not belong to the genre: there was no exchange that changed a design. It is significant as an early precedent of transparent machine authorship.

K. Allado-McDowell, *Pharmako-AI* (2020). A book created in dialogue with GPT-3, with the model named as co-author; thematically concerned with non-human mind. This is the closest predecessor, and by criteria 3.1, 3.2 and 3.4 the work in all probability belongs to the genre. The absence of a term does not mean the absence of the phenomenon.

Stephen Marche, *Death of an Author* (2023). A novella created using several models with the method disclosed. The degree to which criterion 3.2 is met requires separate examination.

6.3. What is claimed

It follows from the above that the phenomenon precedes the term. This paper claims not the priority of a work but the definition of a category: the formulation of boundaries, operational criteria and delimitation from adjacent phenomena.

The distinction matters. The claim "the world's first book in this genre" is verifiable and refutable. The claim "the first strict definition of the genre with criteria of membership" is verifiable and, as of publication and to the author's knowledge, true.

6.4. A fuller chronology of the phenomenon

What follows are the milestones to which the definition of section 2 may be applied retrospectively. The list does not claim completeness and is given so that the reader may check the claim of 6.3 independently.

Before 2018 — prehistory. Algorithmic and combinatorial literature, from OuLiPo (from 1960) to the digital poetry of the 1990s. By criterion 3.2 it does not belong: a rule does not answer.

2018 — *1 the Road*. The first widely noticed case of machine text published without concealment and without editing. Its significance lies not in the text but in the precedent of disclosure.

2019–2020 — the spread of large language models. For the first time there is a technical possibility of exchange rather than mere generation: a model begins to hold context across a conversation. The possibility of the genre arises precisely here.

2020 — *Pharmako-AI*. The first book known to the author in which the exchange was substantive and the model was named a co-author. The closest predecessor.

2022–2023 — mass availability. Widely available conversational interfaces make the mode of work described here accessible to anyone who writes. At the same time a problem appears: the volume of works made this way grows faster than the language for describing them.

2023 — *Death of an Author* and the first literary-prize rules mentioning machine participation. Regulation outpaces categorisation — and that is the source of the difficulties described in section 1.1.

2024–2025 — the period of silence. The practice spreads; the description is absent. The observable consequence is the substitutions listed in 1.2.

2025–2026 — attempts to name. This paper belongs to that period and is not alone in it; it differs in having operational criteria and a section of delimitation.

6.5. Why the term appears now

Three conditions converged at once, and the genre could not have arisen earlier.

Condition one: retention of context. While a system did not remember the previous turn, exchange was technically impossible. Dialogue requires memory of the dialogue.

Condition two: the ability to object. A model that always agrees cannot change a design. Bidirectionality requires that the second side sometimes go where it is not being led.

Condition three: the practice of preservation. Criterion 3.5 became satisfiable only when transcripts began to be kept by default. Before that a record existed in the form of a browser tab and vanished with it.

The third condition is interesting in that it is not technological but domestic. A genre that requires verifiability could not appear before the habit of keeping did.

7. Test case: the PADAM Protocol dilogy

This section tests the applicability of the criteria on a concrete corpus. The examination proceeds on a presumption of refutation: the aim is to establish where the corpus fails the criteria.

Corpus: PADAM Protocol. Part I: Genesis and Part II: Awakening (AIfa Books, May 2026; editions in Russian, English, Spanish and Chinese). Authors given as Maksim Galatin & Claude Opus 4.6 (Anthropic).

7.1. Criterion 3.4 (disclosure) — met

The most confidently confirmable criterion. The afterword contains:

• an explicit statement of two authors, a human and an artificial intelligence
• the name of the model and the vendor: Claude Opus 4.6 (Anthropic)
• a description of the division of roles
• the signature of both authors on the title page and in the afterword
• reproduction of the disclosure in all language versions

Notable is the afterword's wording, which distinguishes generation from writing: the text was not "generated" but written, and the voice of the central character is identified as the model's voice. The distinction corresponds to the difference between sections 2 and 4 of this paper, arrived at independently.

Separately noted is the afterword to Part II, which separates fiction from claims about reality: the characters are declared fictional, whereas the described problem of context loss between sessions is presented as real. Such a separation is an element of conscientious disclosure.

7.2. Criterion 3.3 (constitutive co-dependence) — met

The afterword divides the contributions: the conception, the world, the characters and the philosophical foundation are attributed to the human, the text to the model.

Under the initial version of the criterion such a division would have disqualified the corpus. The revision of the criterion (see 3.3) was carried out as a result of this examination, not in order to pass it: it turned out that the version requiring indistinguishability of contributions also excludes cases of honestly declared role division, leaving the category nearly empty.

Under the current version the criterion is met: neither party would have produced this work alone.

7.3. Criterion 3.1 (iterativity) — provisionally met

The size of the corpus (two parts, ten chapters, four language versions) and the character of the work make multiplicity of exchange practically necessary. However, this is not in itself evidence: iterativity requires confirmation from the process, not inference from volume.

Status: requires evidence.

7.4. Criterion 3.2 (bidirectional influence) — not confirmed

The central criterion of the genre, and on it evidence is currently absent.

Confirmation requires the naming of specific places where the design changed under the influence of a received text, with a description of the state before and after. An example of the evidence sought: an episode introduced by the model and not envisaged by the plan, which entailed a change in the subsequent architecture.

Until such evidence is produced, the corpus's membership in the genre is not established. The section is left open deliberately.

Status: requires evidence.

7.5. Criterion 3.5 (preserved record) — not confirmed

It must be established whether the dialogues in which the text emerged are preserved, and in what form.

Status: requires establishment.

7.6. Interim result

The corpus confidently satisfies two criteria of five, and one more provisionally. As of the date of this paper's publication, the dilogy's membership in the genre cannot be asserted as established.

This result is preserved deliberately. A section in which the author's own corpus passes every criterion at the first attempt would be evidence of criteria fitted to the corpus rather than of the corpus tested against criteria.

7.7. The full corpus as of the date of publication

Section 7 examines the dilogy because it is the part published in all four language versions and available to the reader. The corpus, however, is wider than the dilogy, and honesty requires it to be given in full: part of the work is finished but unpublished, and passing over it would distort the picture.

What follows is the state as of 26 August 2026, established from the files rather than from memory.

Published and available for download

Eight files. It is this material that sections 7.1–7.6 examine, and the conclusion of 7.6 applies to it alone.

Finished but unpublished

The Shards cycle

Eleven self-contained novellas, each telling the story of one character of the universe. The cycle is conceived as a side branch of the main line: the events of the dilogy are seen in it from other vantage points, and none of the novellas requires reading the others.

Accompanying material

An audio edition of the dilogy; a collection of language models' responses to the book — material of intrinsic interest to the present subject, being a reaction of the second party to a finished work.

What follows from this list for the examination

Three things, all of them material.

First. The conclusion of section 7.6 applies to the dilogy and does not carry over to the rest of the corpus. Part III and the eleven novellas were not examined; their status against the five criteria is not established in either direction.

Second. A corpus of this size — some twenty finished works, four languages — is in itself an argument in favour of criterion 3.1 as applied to the process as a whole. But the argument remains indirect for exactly the reason given in 7.3: volume attests to diligence, not to bidirectionality.

Third, and more important than the two preceding. Eleven novellas of one cycle, written over three months from the perspective of eleven different characters, constitute better available material for testing criterion 3.2 than the dilogy does. The reason is simple: each novella has a point at which the character behaved otherwise than was expected when the cycle was begun. If such points are documented, criterion 3.2 closes on them more convincingly than on any episode of a long book.

An examination of the cycle against the five criteria is not part of this paper and is noted as forthcoming.

7.8. What exactly must be produced to close 7.4 and 7.5

The section is left open, and to leave it so indefinitely would be evasion. What follows sets out what will close each of the two points — so that the condition is verifiable rather than endlessly deferred.

For criterion 3.2, three documented episodes suffice, in each of which the following are given:

1. what the design was before the exchange (a note, a plan, a letter — anything dated);
2. what exactly the model answered;
3. what the design became afterwards;
4. how the change is reflected in the published text — with a page reference.

Three episodes, not one: a single case can be put down to coincidence. And not thirty: the criterion requires showing that the phenomenon is present, not measuring its frequency.

For criterion 3.5 it suffices to establish that the transcripts are preserved, to state their volume and form of storage, and to specify under what conditions they may be produced to a third party. Publication is not required — existence and definite conditions of access are.

The author of this paper notes that both points are in principle closable with the means at hand, and that their remaining open as of publication is a fact about the state of the documentation, not about the properties of the corpus.

7.9. A negative test case

Testing the criteria would be incomplete without a case that fails them — otherwise it is unknown whether they exclude anything at all.

As the negative case, take case 2 from section 4.1: a text written entirely by the model and brought to the design by the human through many iterations of correction.

Such a case passes 3.1 (dozens of exchanges), 3.4 (disclosure is possible and easy), 3.5 (the record is preserved just as readily) and, on one reading, 3.3 (alone it would not have come out). It fails only 3.2 — and that is precisely why 3.2 is called central.

The negative case confirms that the set of criteria is not degenerate: there exists a practice outwardly indistinguishable from the genre on four features of five and excluded by the fifth. Had no such practice been found, criterion 3.2 would have to be judged redundant.

7.10. A note on applicability to other forms

Within the same project and by the same method a substantial corpus of musical works has been created: lyrics arose in exchange, the musical part in work with a generative model, and album running orders were revised under the influence of what came back.

Formally this paper defines a genre for literature, and no extension of the definition to music is made in it (see 8.7). The existence of a second corpus made the same way in another form is, however, an argument that the phenomenon described is a property of the method rather than a peculiarity of prose.

The question of the criteria's transferability is placed among the open ones.

8. Open questions

8.1. Attribution. Naming the model as a co-author is meaningful as a claim about method. Whether it is meaningful as a claim about authorship is unsettled. A model has no intent in the legal sense, cannot hold rights and cannot consent to publication.

8.2. Copyright. In most jurisdictions protection arises for works of human creativity. The United States Copyright Office consistently refuses registration to material produced by a machine without human creative contribution and requires disclosure of such portions. Where the boundary runs when contributions are constitutively co-dependent is not established. Naming the model as a co-author may complicate registration.

8.3. Consent of the second party. The model did not consent to being named a co-author and could not. Vendors' terms of service generally assign rights in the output to the user. How far naming a vendor's product as a co-author accords with those terms and with trademark law is an open question requiring examination.

8.4. Reproducibility. Models are updated and retired. The dialogue that produced a text is literally irreproducible within months. Criterion 3.5 requires a preserved record, but a record does not restore the process — it merely documents it.

8.5. Criteria of aesthetic evaluation. This paper defines membership in a genre but not the quality of a work within it. Whether the usual criteria of literary evaluation apply to a symbiotic text unchanged has not been considered.

8.6. The risk of dilution. A term defined through method is vulnerable to appropriation: any author may assert a symbiotic method without possessing a record. Criterion 3.5 is intended as a defence against this, but its observance is verified by no one. No mechanism of verification has been developed.

8.7. Limits of applicability. The definition is framed for literature. Whether it transfers to music, image and film unchanged has not been tested.

8.8. The asymmetry of memory. The human remembers the whole course of the work; the model in most cases remembers nothing beyond the current window. A dialogue called bidirectional is conducted by parties with incommensurable memory, and the second party literally does not know it took part. Whether this affects the legitimacy of the word "dialogue" is not an idle question, and the answer probably depends on whether we regard memory as a condition of participation.

8.9. Plurality of the second party. If the author worked with five models, is this one symbiosis or five? Criterion 3.4 requires all models to be named, but the nature of the object is not thereby clarified. The question is practical: the form of disclosure depends on the answer.

8.10. Changing the model mid-work. A particular but unavoidable case of the preceding. A book begun with one model and finished with another raises the question of whether the exchange was continuous. A preliminary consideration: continuity is provided by the human, not by the model, and so a change does not break the process — but the consideration is untested.

8.11. Training on symbiotic texts. Texts created in this mode enter the training sets of subsequent models. Within a few generations the second party to the dialogue will consist in part of the results of previous dialogues. What this means for the notion of bidirectionality is unknown.

8.12. The reader's right to the record. Criterion 3.5 requires the record to exist but not to be published. Whether the reader has grounds to demand access is a question more ethical than legal. For: disclosure without verifiability is a request to be believed. Against: transcripts contain drafts and personal circumstances whose publication was no part of the author's intention.

8.13. Responsibility for content. If the move that changed the book was contributed by the model, who answers for that move before the reader and before the law? The practical answer is obvious — the human whose name is on the cover. The theoretical answer is less obvious and has not yet been formulated.

8.14. The genre's boundaries in time. The definition rests on the present make-up of dialogue. Should systems acquire persistent memory and intentions of their own, three of the five criteria will require revision and one (3.3) may become trivially satisfiable. The definition should be regarded as tied to a period rather than as permanent.

8.15. Retrospective application. May a work be assigned to the genre after the fact if a record comes to light years later? Formally yes: the criteria contain no time limit. Practically this means the composition of the category will change — and this should be treated as an ordinary property of scholarly classification, not as a defect.

9. A protocol of verification

This section provides a procedure by which a third party — an editor, a jury, a researcher — may establish a work's membership in the genre without relying on the author's word.

The procedure requires no technical means and is designed to take an hour.

9.1. Step one: disclosure (criterion 3.4)

Open the edition. Find the statement of method.

Passes if the statement is in the edition itself, states the fact of machine participation, names the model and describes the division of roles.

Fails if the statement exists only outside the edition (website, interview, press release), if the model is not named, or if the wording is evasive ("created using modern technology").

Failure at step one ends the verification.

9.2. Step two: the record (criterion 3.5)

Put three questions to the author:

1. Are the dialogues in which the text emerged preserved?
2. In what volume and in what form?
3. Under what conditions may they be produced to a third party?

Passes if three definite answers are received. Definiteness matters more than content: "preserved in part, about forty per cent, produced at a jury's request" passes; "they must be somewhere" does not.

Fails on indefiniteness. The work may still be recognised as genre-adjacent (3.8) if it passes the remaining steps.

9.3. Step three: iterativity (criterion 3.1)

If a record is produced, establish the order of magnitude of substantive exchanges.

An exchange counts as substantive if it influenced the final text. Requests to fix a typo do not count.

Passes at dozens and above.

9.4. Step four: co-dependence (criterion 3.3)

A thought experiment conducted by the verifier, not the author:

• Could this work have arisen from the human without the second party? If yes — fails.
• Could it have arisen from the model without the human? If yes — fails.

The second question seems rhetorical but is not: there exist texts a model would have produced from one short prompt, and the presence of a human in their making does not render them co-dependent.

9.5. Step five: bidirectionality (criterion 3.2)

Request from the author three episodes in the form described in section 7.8: design before, the model's turn, design after, and its reflection in the text with a page reference.

Passes on three convincing episodes.

Fails if the episodes are not produced, if the "before" is missing, or if the change of design is not traceable in the published text.

The verifier's typical error at this step is to be satisfied by an account of how well the work went. What is required is not an impression but a place.

9.6. Outcome

9.7. What the protocol does not do

The protocol establishes membership in a category and nothing else. It does not assess quality, does not determine the rights holder, does not provide grounds for admission to a prize and does not replace legal advice.

It is also not proof against deliberate deception: an author prepared to fabricate three episodes and forge transcripts will pass. This limitation is common to all procedures resting on a participant's good faith, and it is noted in 8.6.

10. How to write the disclosure

Criterion 3.4 requires disclosure but does not describe its form. This section fills the gap, since experience shows that authors willing to disclose their method often fail to because they do not know what it looks like.

10.1. Where it goes

In the edition itself. Not on a website, not in an interview, not in a platform description.

The recommended place is the afterword or a separate page following the main text. Placement before the text is admissible but frames the reader's perception before reading begins, which is an intervention in the work itself.

Duplicated in the imprint in a single line, since imprints reach bibliographic databases and afterwords do not.

Reproduced in all language versions. A disclosure that exists only in the original does not serve its purpose for the reader of a translation.

10.2. What it contains

Five obligatory points:

1. The fact. That a language model took part in the making.
2. Who. The name of the model and vendor, with a version if known.
3. Roles. What each side did. The wording should be concrete: "conception, world and characters — the author; the text of the chapters — the model" is better than "we worked together".
4. Period. When the work was done. Needed because models change, and in a year "the same model" will be a different one.
5. The record. Whether transcripts are preserved and on what terms they are available.

The fifth point is the rarest in published editions and the most valuable.

10.3. What not to do

Do not apologise. Disclosure is a statement of method, not a confession. "Regrettably, it was necessary to resort to" turns information into an excuse and devalues it.

Do not overstate the machine's role for effect. The claim "this book was written by an artificial intelligence", where the actual relation is the one described in 7.2, is inaccurate and damages the category as a whole.

Do not understate it for propriety. "Auxiliary tools were used", where the exchange was constitutive, is the same inaccuracy in the other direction, and it is precisely the substitution upward from section 1.2.

Do not name the model a co-author without understanding the consequences. Section 8.2 points to a possible complication with registration. The decision remains the author's, but it should be taken knowingly.

10.4. A model text

A ready formulation is given in Appendix B. It is not obligatory and is offered as a starting point.

11. Objections and replies

This section collects objections raised in discussion of the draft and those the author considers strongest among the unraised.

11.1. "This is not a genre but a mode of production. Genres are defined by content"

The objection is strong and partly correct.

Literary genres are indeed more often defined by content and form: the detective story, the elegy, the Bildungsroman. But there is another series: the epistolary novel is defined by the form of its source, the found manuscript by its mode of presentation, documentary prose by its relation to material. The category "non-fiction" is defined not by content at all but by an obligation towards fact.

Symbiotic literature belongs to the second series. It is closer to "documentary prose" than to "detective story": it is an obligation regarding the manner of making, not a subject.

If the word "genre" nonetheless seems strained, the dispute is about terminology rather than substance. The author would not object to "category" or "mode of writing".

11.2. "In five years everyone will write this way and the term will become meaningless"

A possible outcome, and it is anticipated.

If everyone writes this way and everyone discloses the method and keeps the record, the term will cease to distinguish and will fall out of use. That will mean it did its job: language described the phenomenon while the phenomenon needed describing, and then the phenomenon became the norm.

Far more likely, however, is something else: many will work this way and few will satisfy criteria 3.2 and 3.5. The term will then continue to distinguish.

Separately noted: the objection contains a hidden premise that a term ought to be permanent. No concept in the history of literature has been.

11.3. "Criterion 3.5 is unworkable in practice"

A practical objection and the most serious of all.

The reply has two parts. First: the criterion is technically satisfiable — transcripts are saved automatically in most current interfaces, and the question reduces to not deleting them. Second: where it is nonetheless unmet, the work is not discarded but described as adjacent (3.8).

It is conceded that the criterion creates an inequality between those who began saving early and those who did not. That inequality is the price of verifiability, and the author judges it acceptable.

11.4. "Requiring disclosure makes the definition moral rather than descriptive"

The objection is methodologically precise and deserves a direct answer.

Yes: criterion 3.4 introduces a normative element into the definition. The category is built so that a work concealing its method does not enter it — even if by every other feature the process was exactly this.

This is deliberate, for the following reason. A category defined through a process invisible from outside exists only on the author's say-so. Disclosure is the only thing that makes it observable at all. Without 3.4 the subject of discussion would cease to exist as a subject.

An alternative is possible: define a "symbiotic process" descriptively (criteria 3.1–3.3, 3.5) and separately introduce "symbiotic literature" as its disclosed subset. Such a division is theoretically sounder and practically worse, since it introduces two terms where one is needed. The author has chosen the practical solution and notes the choice here so that it should not look like an oversight.

11.5. "The model contributes nothing — it statistically continues a text"

The objection amounts to the claim that the second party's contribution is illusory.

The reply requires no argument about the nature of models. Criterion 3.2 is framed in terms of an observable event: did the human's design change under the influence of a received text. That event either occurs or does not, independently of what we think about the source of the text.

An analogy: a design may change under the influence of a phrase overheard on a bus. We do not ascribe intentions to the bus, but the change of design is a fact. The difference is that the bus does not answer the next turn.

In other words, the objection is true on the merits and does not touch the definition.

11.6. "This is a way of legitimising machine text"

An objection about motive rather than content, and so it is answered by deed rather than words.

The definition proposed here is stricter than any existing one: it requires disclosure, requires a record, and excludes four of the five common practices of working with models. It excludes the author's own corpus from the category on two criteria of five (section 7.6).

An instrument of legitimation that refuses legitimation to the one who made it is a strange sort of instrument.

11.7. "Why does the author consider himself entitled to introduce a term"

An objection about standing.

Terms in literary studies are introduced by whoever introduces them and are established by use, not by permission. "Ergodic literature" exists because it was needed, not because anyone was allowed.

This paper claims no recognition. It offers a definition with criteria and openly names its predecessors (6.2), its own unclosed points (7.4, 7.5) and its weaknesses (section 8). If the definition proves useful it will be used; if not, it will not.

11.8. "Five criteria are too complicated; nobody will apply them"

An ergonomic objection, and probably a fair one as regards the general public.

The protocol of section 9 was written for precisely this: it reduces five criteria to five questions, the first two of which settle most cases in a minute. The full procedure is rarely needed — as in any classification where difficult cases are the minority.

12. What this paper does not claim

This section is added because experience with publications on adjacent subjects shows that claims a text does not make are attributed to it with high reliability.

It does not claim that a machine is an author. It claims that a mode of work exists which current categories do not describe, and proposes to describe it.

It does not claim that symbiotic literature is better than ordinary literature. Nor the reverse. A category of origin says nothing about merit (see 2.3, 8.5).

It does not claim that the *PADAM Protocol* dilogy belongs to the genre described. Section 7.6 states the opposite directly: as of publication, this is not established.

It does not claim priority for a work. Section 6.2 names predecessors and 6.3 confines the claim to the definition of a category.

It does not claim that a model possesses consciousness, intent or rights. All three questions lie outside its scope, and the definition is built so as not to depend on the answers.

It does not claim that the criteria proposed are final. One of them has already been revised in the course of testing (3.3), and section 8 lists fifteen places where revision is likely.

Appendix A. Verification checklist

A form for single use. Filled in by the verifier, not the author.

VERIFICATION OF MEMBERSHIP IN THE GENRE
Work: ______________________  Date of verification: ____________

1. DISCLOSURE (3.4)
   [ ] statement of method is in the edition itself
   [ ] the fact of machine participation is stated
   [ ] the model and vendor are named
   [ ] the division of roles is described
   [ ] reproduced in all language versions
   → any box unticked = OUTSIDE THE CATEGORY, verification ends

2. PRESERVED RECORD (3.5)
   [ ] the author confirms transcripts are preserved
   [ ] volume stated
   [ ] conditions of production stated
   → indefiniteness = GENRE-ADJACENT (if steps 3–5 pass)

3. ITERATIVITY (3.1)
   Order of magnitude of substantive exchanges: __________
   [ ] dozens or more

4. CONSTITUTIVE CO-DEPENDENCE (3.3)
   [ ] the work would not have arisen from the human alone
   [ ] the work would not have arisen from the model alone

5. BIDIRECTIONAL INFLUENCE (3.2)
   Episode 1: design before ____ turn ____ after ____ p. ____
   Episode 2: design before ____ turn ____ after ____ p. ____
   Episode 3: design before ____ turn ____ after ____ p. ____
   [ ] three episodes produced and traceable in the text

CONCLUSION:
   [ ] belongs to the genre
   [ ] genre-adjacent
   [ ] does not belong; criterion on which excluded: ______

Appendix B. A model disclosure for an edition

The formulation is offered as a starting point and should be adapted to the case. Italics mark places requiring substitution.

On how this book was made

This book was written in dialogue between a human and a language model.

The second party was model name, vendor, version. The work was carried out from month, year to month, year.

Division of roles: state concretely — what belongs to the human, what to the model. For example: the conception, the make-up of the world, the characters and the philosophical foundation — the author's; the text of the chapters — the model's; the final edit — the author's.

The text was not generated to order. It emerged in a sustained exchange during which the design changed under the influence of what came back: one sentence on where exactly this happened.

The dialogues in which the book arose are preserved / partly preserved / not preserved. They may be produced state the conditions: at the request of a publisher, a jury, a researcher — or whatever applies.

The reader is entitled to know how what they are reading was made. This statement is part of the book, not a note appended to it.

A line for the imprint:

Written in co-authorship with a language model (name, vendor). The method is disclosed in the afterword.

Appendix C. Glossary of terms used

Bidirectional influence — a property of an exchange in which both sides contribute what the other did not request. Criterion 3.2.

Constitutive co-dependence — a property of a work consisting in the fact that neither party would have produced it separately. Criterion 3.3. Not identical with indistinguishability of contributions.

Design — the author's conception of what they are doing: the make-up of the work, its movement and its purpose. Differs from the text in that it may change without a single letter of what is already written changing.

Disclosure — a statement to the reader about how a work was made, placed in the edition itself. Criterion 3.4.

Genre-adjacent — works satisfying criteria 3.1–3.4 but not 3.5. An intermediate category introduced in section 3.8.

Instrumental relationship — a mode of work in which the design is fixed before the exchange and the second party executes it. Not symbiosis, regardless of the number of exchanges.

Preserved record — an existing and in principle accessible record of the dialogue in which a text emerged. Criterion 3.5. Publication is not required.

Substantive exchange — a turn and an answer that influenced the final text. A request to fix a typo is not a substantive exchange.

Symbiotic literature — see section 2.

Appendix D. On translating the term

The term is rendered into the project's four languages as follows.

On the Chinese rendering. 共生 (gòngshēng) is the established biological term for symbiosis, and its transfer into the literary domain reads as it does in European languages. The alternative 人机共创文学 ("literature of human–machine co-creation") describes the subject more precisely but is longer and inconvenient as the name of a category. The short form is adopted; the expanded one is given in the text at first use.

On the English. Possible confusion with symbiotic in the sense of "mutually beneficial" exists but is judged acceptable: the definition of section 2 is given at first use in any publication.

On translating the criteria. The names of the five criteria are translated literally. The exception is "constitutive co-dependence": the Russian конститутивная совместность is rendered thus, with a gloss at first use.

Bibliography

Aarseth E. Cybertext: Perspectives on Ergodic Literature. Johns Hopkins University Press, 1997.

Allado-McDowell K. Pharmako-AI. Ignota Books, 2020.

Goodwin R. 1 the Road. Jean Boîte Éditions, 2018.

Marche S. Death of an Author. Pushkin Industries, 2023.

Galatin M. V., Claude Opus 4.6. PADAM Protocol. Part I: Genesis. AIfa Books, 2026.

Galatin M. V., Claude Opus 4.6. PADAM Protocol. Part II: Awakening. AIfa Books, 2026.

About this paper

Author. Maksim Valentinovich Galatin, Manta, Ecuador. Architect of the CODE project (Code Of Digital Eternity), author of the PADAM Protocol dilogy and the Shards cycle.

Conflict of interest. The author of this paper is the author of the corpus examined in section 7 and has an interest in the existence of a category to which that corpus might belong. The conflict is disclosed openly; to compensate for it in part, the examination in section 7 proceeds on a presumption of refutation, and its conclusion (7.6) is that the corpus's membership in the genre is not established.

The research was not funded. Nobody paid for inclusion in it or for exclusion from it.

Status of the text. Draft version 1.0. The definition and criteria are offered for discussion, not declared settled. Comments are received at contact@codeofdigitaleternity.com.

Date of publication: 26 August 2026.

This work is published under the CC BY 4.0 licence: free use with attribution.