Who Had the Final Say? Authorship in AI-Assisted Creative Work

Research materials

Preprint v1.0

Beyond AI Share: A Preregistered Survey and Vignette Study of Perceived Control, Authorship, and Authenticity in AI-Assisted Creative Practice

• 29 pages
• Not formally peer reviewed

Open DOI recordOpen preprint PDF

Supplementary Appendix A

11 pages

Full questionnaire, vignette wording, scoring rules, documented procedures and reporting limitations

Open Appendix A PDF

OSF preregistration

Open OSF preregistration

The study was preregistered before the first recorded main-study response. The preprint has not undergone formal peer review.

Two accounts and the gap between them

The quickest objection to AI-assisted creativity is also the most sweeping: it is no longer your work because the computer did it. In art, the charge has particular force. A name attached to a novel, image, film, or song links the result to someone’s intention, judgment, and responsibility.

Yet a system may have surfaced a lead, compared sources, proposed alternatives, outlined a structure, drafted a passage, or produced something close to a finished work. All of those actions disappear inside the same thin disclosure: AI was used.

The controversy surrounding Olga Tokarczuk’s appearance at Impact’26 in May 2026 exposed that ambiguity. My Company Polska’s report of the event reported that she spoke positively about an advanced model that could broaden horizons and deepen creative thought. The report also described the model as part of analyzing and developing ideas. This was a media account, not an official transcript.

A few days later, Tokarczuk defined the boundary differently. In a statement dated 19 May and supplied to Literary Hub by her publisher, she denied that her forthcoming novel, or any of her writing, had been written with AI. She described the tool as a way to speed up documentation, fact-checking, and preliminary research, and said she checked the results herself. The Polish statement published by Rzeczpospolita made the denial equally explicit.

Those two public accounts should remain separate. The report of the talk described the model as involved in analyzing and developing an idea. The later statement drew the line at research and verification while rejecting AI-generated prose. The difference does not prove that a model wrote the book, that the author concealed part of her method, or that the report was false. This article is not a verdict on Tokarczuk’s practice. It asks why the sentence “I use AI” cannot settle a question of authorship.

We do not have a complete official recording of the exchange or a documented production history for the novel. Without prompts, outputs, notes, and versions, “analyzing an idea” could mean testing counterarguments, finding research gaps, soliciting a plot suggestion, or something else. The available record does not allow us to fill that gap honestly.

How tool use became an authorship dispute

In an opinion column for Onet, Sławomir Sierakowski described how the language escalated after the event: a mention of AI became an admission, use became co-writing, and co-writing became the claim that AI wrote for her. Each step gave the model a larger role without adding evidence about the actual workflow.

The dispute was not created by headlines alone. The reported remarks about creative thinking made it reasonable to ask about the conceptual stage, while the later statement described the use as limited to research. The problem began when an unresolved discrepancy was turned into a complete story of how the novel had been made.

A headline needs one verb. Creative work needs many: search, propose, compare, reject, rewrite, verify, and approve. When “use” replaces all of them, a technical description becomes either a moral judgment or an alibi. One side hears that a computer performed the work. The other replies that AI was only a tool. They may be defending different practices.

Online reactions show why the argument matters, but they cannot supply the missing production record. Comments appearing beneath LubimyCzytać’s coverage are not a representative sample. At most, they reveal recurring concerns about human direction, disclosure, training data, compensation, and lived experience.

Writers draw the line in different places

Writers’ responses do not amount to a simple vote for or against AI. They defend different parts of creative practice.

According to Euronews, which embedded his Facebook post, Szczepan Twardoch drew a firm line at delegating the writing itself. He stressed that he had written every word of his novels, essays, and journalism and did not intend to delegate that part of his craft to a language model. He left room for limited technical uses, such as aggregating information from many sources or producing rough technical translations, while warning that they remain unreliable.

Wojciech Chmielarz, quoted in the same report, said the judgment should depend on context. He distinguished research from planning and inventing a novel. In the latter case, he argued that a boundary in the relationship between reader and writer had been crossed and that the use should be disclosed. That is not a universal rule, but it recognizes that checking a date, asking for structural alternatives, and accepting a finished paragraph involve delegating very different tasks.

Justyna Bargielska described another practice in Tygodnik Powszechny: using ChatGPT as a medium or mirror and selecting individual words and sentences from the exchange. She located authorship in selection, transformation, and accountability. The system’s contribution did not vanish, but responsibility for the poem remained with the author.

In the same article, Tadeusz Dąbrowski shifted attention from craft to the reader-writer relationship. He did not reject every literary use of AI. In poetry, however, he emphasized the reader’s trust that a human being wrote the text. The same line may be received differently when the reader learns that it came from lived experience or from the first model output after a short instruction.

Twardoch protects independent writing. Chmielarz distinguishes stages and disclosure. Bargielska emphasizes selection and responsibility. Dąbrowski emphasizes trust in human authorship. None of these positions settles every case of AI-assisted creation. Together they show why creative authorship cannot be reduced to a single scale.

Literature has always involved editors, translators, researchers, and conversation partners. That complicates the myth of the solitary author without making AI equivalent to a human collaborator. A person has a recognizable role, can explain decisions, and works within professional norms. A model combines several roles while its sources and criteria remain partly opaque.

One debate, four different questions

The first question concerns authorship and the organization of the workflow. Who set the goal? Who produced or requested alternatives? Who compared them, made consequential changes, made the final decision, and accepted responsibility? This is a question about the distribution of decisions in a particular project.

The second concerns disclosure and audience trust. Research, language editing, idea development, draft generation, and incorporation of generated material into a finished piece are not equivalent. Expectations also differ by genre. A report, a personal essay, a poem, a commissioned illustration, and an internal memo make different promises to their audiences.

The third concerns rights, labor, and training data. It includes consent, reservation of rights, licensing, compensation, dataset provenance, and the possible displacement of creative labor. Concluding that a person directed one project says nothing about the terms on which the tool was built.

The fourth concerns human experience and creative identity. An audience may value not just the result but the knowledge that it carries someone’s experience, risk, effort, or sensibility. A creator may experience machine imitation of a hard-won style as a threat even when a single output looks persuasive.

Each layer needs different evidence: version histories, audience research, dataset and contract documentation, law, or psychological and philosophical argument. One study or lawsuit cannot answer them all.

The study below primarily addresses the first layer and part of the fourth. It does not set disclosure rules, determine copyright or compensation, reconstruct Tokarczuk’s novel, or define a universal boundary across creative fields.

A percentage is not a history of the work

Asking what percentage of a piece was “done by AI” sounds precise. But there is no agreed unit of measurement. We might count time, words, pixels, operations, similarity to the first generated output, or influence on the core idea. Each measure captures something different.

A short suggestion can redirect an entire project. A large volume of generated material can remain subordinate to a human concept when someone compares alternatives, rejects most of them, and substantially changes the selected result. Volume and creative weight do not increase in parallel.

Film makes this intuitive. Screen time alone does not establish who authored a scene. Meaning also emerges through direction, shot selection, editing, sound, and the choice to keep one version over another. Role descriptions reveal more than a single number.

The study’s AI-involvement measure was not an objective percentage either. Participants rated the usual role of AI in their final output on a subjective five-point scale, from none or almost none to very large. It captured a general impression. It did not record prompts, successive versions, rejected outputs, or whether a small intervention changed the meaning of a whole project.

A concise description of the workflow can do more. For example: original concept; model used to find leads and generate alternatives; sources checked independently; final text written and edited by the author. Another might read: short brief; first generated output; minor edits. Both involve AI, but the descriptions reveal more than “30% AI.”

The survey presented two deliberately contrasting workflows:

Human-directed workflow

A creator prepares a poster. She has her own idea, sketches the composition, uses AI to generate several background alternatives, selects one, edits it substantially, and decides on the final layout herself.

Near-final AI output

A creator receives a short assignment brief. He enters it into an AI tool, selects the first result, makes only minor changes, and submits it as the final work.

Both can be summarized as AI use. They differ in idea origin, alternatives, selection, degree of revision, and final decision authority. Participants rated the two bundles very differently.

Two AI-assisted workflows

Two bundled scenarios from the study, shown as sequences of decisions and actions.

Human-directed workflow

1. 1original idea→
2. 2AI-generated alternatives→
3. 3selection→
4. 4substantial revision→
5. 5final decision

Near-final AI output

1. 1short brief→
2. 2first AI output→
3. 3minor edits→
4. 4submission

What the study of 429 adults found

The preregistration specified the hypotheses and core analysis plan before the first recorded response in the main study. That limits the freedom to rewrite the story after seeing the results. It does not turn a survey into a causal experiment, improve weak measures, or make a convenience sample representative.

In Feliks Mamczur’s preregistered study, self-reported AI involvement was not related to perceived authorship in the predicted way. A clearer difference emerged when participants evaluated two different ways of organizing creative work.

H1 predicted that creators who reported more AI involvement in the finished output would report a weaker sense of authorship. The data did not support that prediction. The observed relationship was small and positive, opposite to the hypothesis, but its confidence interval included zero and it did not cross the prespecified significance threshold. This is not evidence that more AI increases authorship, nor that AI involvement never matters. It means that this subjective five-point measure did not behave according to the proposed rule of “more AI, less authorship.”

H2 was supported: greater perceived control was associated with stronger perceived authorship. The relationship was small, and a simple model explained 5.5% of the variance. Because the design was correlational, it cannot tell us whether control strengthens the sense of authorship, whether people with a stronger sense of authorship describe their workflows as more controlled, or whether both judgments share another cause.

H3 predicted that control would weaken the expected negative relationship between AI involvement and authenticity. The interaction was not supported. The same model showed a positive main association between control and authenticity, but that is a different claim. The result does not show that enough control neutralizes an effect of AI involvement.

H4 concerned people who strongly associate art with human experience and intention. This expressive orientation correlated with greater creative identity threat. A stronger correlation appeared for one auxiliary item about a recognizable “typical AI aesthetic,” but that item was not a validated scale. The relationship with the general authenticity indicator was effectively zero. H4 therefore received mixed support and should be interpreted cautiously. It is not evidence that people who value human expression reject every AI-assisted piece.

The vignettes produced the clearest result. Combined ratings of authorship, authenticity, and control were higher for the human-directed workflow than for the near-final-AI-output workflow. The direction was the same on every dimension.

How the two workflows were rated

A paired-dot plot comparing mean ratings for the human-directed workflow and the near-final AI output on a common 1-5 scale.

Human-directed workflowNear-final AI output

• Authorship: Human-directed workflow M 3.430, SD 1.208; Near-final AI output M 2.570, SD 1.235; Difference .860; d_z .610.
• Authenticity: Human-directed workflow M 3.507, SD 1.172; Near-final AI output M 2.706, SD 1.253; Difference .801; d_z .559.
• Control: Human-directed workflow M 3.565, SD 1.196; Near-final AI output M 2.720, SD 1.255; Difference .846; d_z .597.
• Combined index: Human-directed workflow M 3.501, SD 1.081; Near-final AI output M 2.665, SD 1.132; Difference .836; d_z .659.

Authorship

2.5703.430

Authenticity

2.7063.507

Control

2.7203.565

Combined index

2.6653.501

Rating scale

12345

Show the complete data table

Complete vignette comparison data; n = 428 complete paired responses; scale 1-5.

Dimension	Human-directed workflow		Near-final AI output		Difference	d_z
	M	SD	M	SD
Authorship	3.430	1.208	2.570	1.235	.860	.610
Authenticity	3.507	1.172	2.706	1.253	.801	.559
Control	3.565	1.196	2.720	1.255	.846	.597
Combined index	3.501	1.081	2.665	1.132	.836	.659

That contrast does not identify a single mechanism. The scenarios bundled idea origin, number of alternatives, selection, degree of editing, and final authority. They also differed in protagonist gender, task context, and implied effort, and they always appeared in the same order. The comparison did not isolate control, selection, or veto power, and the fictional creators should not be treated as stand-ins for real people.

Authorship is better described by tracing decisions

Four categories offer a more useful vocabulary: direction, selection, revision, and responsibility. They form an interpretive framework, not a validated scale or a threshold at which a work becomes human-authored.

Direction covers purpose, meaning, constraints, and standards of success. The person who typed the first prompt did not necessarily define those things. A system can impose a frame through the alternatives it makes visible, and a user can accept that frame without noticing. The reverse is also possible: a model supplies an initial spark, but the creator reformulates the problem and directs everything that follows.

Selection matters when the choice is real and guided by criteria the creator can apply. Seeing five alternatives is not proof of control. Revision varies just as much. Fixing punctuation, changing the tone, and rebuilding an argument all count as edits, but they do not carry the same creative weight.

Responsibility begins before the signature. Its practical core is the right of veto: the ability to stop, reject a convenient suggestion, and accept the cost of starting again. Clicking “publish” is not enough when the person does not understand the result, cannot verify it, or cannot meaningfully depart from the path proposed by the system.

A decision history is not a perfect authorship test. It is still more informative than a product name or percentage. It can show where the model proposed, where the person chose, what changed substantially, and whether “no” remained a real option. Future studies should measure those features separately rather than assume they form one dimension.

Responsibility does not end with the finished work

A useful disclosure names roles. “Made with AI” is almost as uninformative as silence. Research and fact-checking, language editing, idea development, draft generation, and direct use of generated content in the final result should not be treated as equivalent.

The appropriate level of disclosure depends on the promise made to the audience. In reporting, sources and verification are central. In a personal essay or poem, the origin of the voice and experience also matters. In commercial illustration, displacement of commissioned human labor may matter. Disclosure is not punishment for using a tool. It gives the audience information needed to understand the terms on which the work is being presented.

A list of model names also conceals whether outputs were rejected, rewritten, or published. A better disclosure explains when the system entered the workflow, what it supplied, and what the person did before approving the result. It need not expose private conversations or every draft.

Responsibility itself has two meanings. The first concerns the finished result: factual accuracy, harm, tone, quality, and the decision to publish. A person who signs a text cannot transfer blame for an accepted error to the model. Version histories, sources, and records of consequential changes can show what the human actually reviewed and approved.

The second concerns the ecosystem: training-data provenance, consent, licensing, creator compensation, data labor, bias, and infrastructure costs. A workflow can be strongly human-directed and still rely on a system that raises justified objections. An individual can choose a tool and how to use it, but cannot control the entire production chain.

Polish copyright law contains text-and-data-mining exceptions. Under the general exception, reproductions of lawfully available works may be made for TDM unless the rights holder has appropriately reserved the rights. ZAiKS has announced an opt-out for rights it represents and argues for licensing. A lack of reservation is not individual consent, although use may fall within the exception when its legal conditions are met.

In March 2026, the European Parliament adopted resolution P10_TA(2026)0066, with recommendations on transparency, licensing, and remuneration. It is a resolution and does not itself create directly applicable new law. Poland’s Ministry of Culture and National Heritage has reported dialogue with creative sectors and work on a Polish position concerning a planned 2027 update of the DSM Directive, particularly its TDM rules. That announcement is neither a statute nor a bill. This is a limited overview, not legal advice or a finding about the training of any particular model.

The boundary of knowledge comes before the boundary of opinion

The findings are a starting point, not a verdict. The study used an online convenience sample that was not representative. It cannot tell us how common these attitudes are among all creators, AI users, readers, or adults, and creative fields were unevenly represented.

The study was cross-sectional and correlational. Responses were collected at a single point in time, so the study cannot establish that control causes authorship or show how practice changes with experience. AI involvement was measured on a subjective five-point scale. The study did not inspect creative artifacts, prompts, model outputs, version histories, or records of rejected alternatives.

Several short, author-developed indicators had low reliability: α = .519 for authorship, α = .521 for authenticity, α = .452 for control, and α = .646 for identity threat. Measurement error may weaken correlations and make interactions harder to detect. The two three-item vignette indices had higher internal consistency, α = .892, but that does not resolve the design limitations of the scenarios.

The vignettes changed several features at once. Their order was not randomized or counterbalanced; the human-directed scenario always came first. The protagonists differed in gender, task source, implied speed, and effort. A poster example may also have felt more concrete to visual creators than to people working with writing, music, or performance. We cannot know which component produced the difference.

Completion times varied widely. Alternative time filters did not change the final classification of the hypotheses, but H1 and H3 were less stable than H2, H4, and the vignette contrast. Complete counts were not retained for everyone who saw the invitation, opened the survey, started it, or left before submitting. Response and completion rates cannot be calculated. The sample of 429 consists of completed responses retained for analysis, not a 100% completion rate.

No prospective ethics committee approval or formal exemption was obtained before recruitment. Participants were adults, provided informed consent, and no direct identifying information entered the analytic dataset, but those safeguards do not substitute for formal review. The pilot’s 23 completed responses were used to check the questionnaire and were excluded from the main study. The preprint has not undergone formal peer review, and its results do not determine legal authorship or copyright protection.

A stronger study would vary idea origin, alternatives, selection, revision, and final authority separately while holding the creator and domain constant. It would randomize and counterbalance the vignettes, use stronger measures and a broader sample, preserve participant-flow data, and obtain prospective ethics review. With consent, it could compare self-reports with version histories and rejected outputs across creative fields.

For now, the conclusion should remain modest. Self-reported AI involvement was not related to perceived authorship in the predicted way. Perceived control had a small positive association with perceived authorship. Two bundled workflows received markedly different ratings. That is a reason to reject a percentage threshold as a ready-made authorship test. It is not a replacement test.

Returning to the Tokarczuk controversy should therefore not end with a verdict on a novel whose production history we have not seen. It should lead to a better standard for the discussion. Instead of turning “I use AI” into an accusation or an absolution, ask what decisions were delegated, what remained open to revision, and where the evidence ends.

Having the final say means being able to set direction, make informed choices, change something that matters, or refuse. It also carries responsibility for the result, for choosing the tool, and for how it is used. It does not make one user responsible for the entire ecosystem. Data sources, labor conditions, and infrastructure also depend on decisions made by companies, institutions, and lawmakers.

References

1. Mamczur, Feliks (2026). Beyond AI Share: A Preregistered Survey and Vignette Study of Perceived Control, Authorship, and Authenticity in AI-Assisted Creative Practice. Preprint, version 1.0, Zenodo. https://doi.org/10.5281/zenodo.21705721. Not formally peer reviewed.
2. Mamczur, Feliks (2026). Supplementary Appendix A. Beyond AI Share: A Preregistered Survey and Vignette Study of Perceived Control, Authorship, and Authenticity in AI-Assisted Creative Practice. PDF in the Zenodo record.
3. Mamczur, Feliks (2026). Is It Still My Work? Authorship, Authenticity and Control in AI-Assisted Creative Practice. OSF Preregistration. https://doi.org/10.17605/OSF.IO/GSWN3.
4. Cyrny, Wiktor (15 May 2026; updated 21 May 2026). “Olga Tokarczuk zapowiada ostatnią powieść w karierze…”. My Company Polska. Media report, not an official transcript.
5. Literary Hub (19 May 2026). “Olga Tokarczuk has responded to the controversy over her reputed use of AI”. Publisher-supplied statement translated by Antonia Lloyd-Jones.
6. Rogalska, Anna (21 May 2026). “Olga Tokarczuk o AI przy pisaniu książki. Noblistka wydała oświadczenie”. Rzeczpospolita. Polish text of the statement and context.
7. Impact CEE (25 May and 16 June 2026). Impact’26 on YouTube, most-watched Impact’26 conversations, and the official Impact CEE YouTube channel.
8. Sierakowski, Sławomir (20 May 2026). “Sławomir Sierakowski odpowiada atakującym Olgę Tokarczuk”. Onet. Signed opinion column.
9. Twardoch, Szczepan (Facebook post), and Chmielarz, Wojciech (Facebook post) (May 2026). Permalinks located through embeds in Euronews; direct Facebook access was unavailable, so the article uses cautious paraphrases.
10. Wilkowski, Marcin (18 November 2025). “Sztuczna inteligencja już pomaga pisarzom. Czy to faktycznie źle?”. Tygodnik Powszechny. Source for Justyna Bargielska and Tadeusz Dąbrowski.
11. LubimyCzytać (20 May 2026). “Olga Tokarczuk o AI…”. Comments used only as a nonsystematic sample of debate language.
12. Republic of Poland (2024). Act of 26 July 2024 amending the Act on Copyright and Related Rights. Source for the general description of TDM exceptions.
13. ZAiKS (30 May 2025). “OPT-OUT”. Source for ZAiKS’s action and position, not proof of effectiveness against every model.
14. European Parliament (10 March 2026). P10_TA(2026)0066, final adopted text in procedure 2025/2058(INI). Recommendations, not directly applicable new law.
15. Ministry of Culture and National Heritage, Poland (23 June 2026). “Rozmawiamy z przedstawicielami twórców i artystów o wykorzystywaniu AI w kontekście prawa autorskiego”. Government communication about dialogue and policy work, not a statute or bill.
16. European Union (13 June 2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence, Articles 50 and 113. Official text of the transparency duties and their application date.
17. European Commission (20 July 2026). Final Guidelines on the transparency obligations for providers and deployers of certain AI systems. Practical guidance on applying Article 50; the guidelines do not amend the regulation.

Suggested citation

Mamczur, F. (2026, July 31). Who Had the Final Say? Authorship in AI-Assisted Creative Work. Prompted Psyche. https://promptedpsyche.com/articles/who-had-the-final-say-ai-authorship/

© 2026 Feliks Mamczur / Prompted Psyche. All rights reserved. Short quotations are permitted only with attribution to the author, the title and a link to the source.