Skip to content
PlagiatScanner.de
Publication ethics · question · PUB-12

Why an AI system cannot be an author

In brief: An AI system cannot approve a final manuscript, declare competing interests, grant publishing rights or answer for errors and integrity concerns. Those are authorship obligations. People choose which outputs to use, verify claims and sources, and remain accountable. Where AI assistance is relevant under journal policy, describe the tool, task, extent and human checks in the required declaration or methods section; do not place the system in the byline.

Responsibility and disclosure matrixSources reviewed 4 September 2026

Authorship is an accountable relationship

A byline allocates recognition and duties. Authors approve the published version, identify their contributions, disclose relevant interests and cooperate when questions arise. A model cannot enter that relationship, understand a binding declaration or take responsibility to readers, participants and the journal.

Producing many words does not change the position. Volume is not accountable scholarship. The person who selects, edits or rejects an output must be able to explain why it belongs in the work. Model developers do not become authors of every paper that uses their software either.

Authority asset: responsibility and disclosure matrix

AI-supported tasks and the human obligations attached to them
TaskSystem activityHuman obligationRecord or disclosure
BrainstormingOffers possibilitiesJustify question; check sourcesTool, purpose and influence under policy
Language editingSuggests wordingCheck meaning, terminology and rightsExtent and retained changes
Code assistanceGenerates or explains codeTest, validate, review licence and securityAffected analysis and test record
Literature discoveryNames possible worksLocate and read originalsActual database and search route
SummarisationCondenses supplied materialCompare every proposition with sourceMaterial, purpose and verification
AuthorshipCannot fulfil rolePeople approve and answer for workNo system in byline

A “checked” entry needs a named responsible role and a repeatable control, not general confidence in the output.

Describe tool activity without promoting it to authorship

Models may propose prose, outlines, code or candidate search terms. They cannot warrant factual accuracy, methodological fit or lawful reuse. Authors must open primary sources, execute and test analysis, and remove unsupported assertions. Disclosure is not a substitute for those controls.

Record the product, accessible version or date, task, type of input, material retained and verification. If confidential or personal data were transferred, assess whether that processing was authorised. An honest statement about use cannot retrospectively legitimise an impermissible disclosure.

Journal definitions differ between predictive software, spelling tools and generative models. Use the selected venue’s categories. Where policy is unclear, ask the editorial office with a concrete account of the intended use rather than the generic question “May I use AI?”.

Search deliberately for plausible errors

Check every citable claim against the original publication. A convincing title may be invented, and a real article may not support the supplied proposition. Resolve DOI, read the relevant passage and record its location. Never copy a statistic or page number solely from a generated response.

For language editing, compare source and revision sentence by sentence. Watch for lost uncertainty, reversed causation, stronger generalisation and altered technical terms. Small wording changes can transform methods or results. Humans decide the final meaning and must understand it.

Generated code requires tests against known cases, software-version records and subject-matter review. Successful execution is not validation. Preserve material failed attempts and corrections where they explain the reported analytical pathway.

Put a precise declaration in the required place

Policies may request information in methods, acknowledgements, a submission form or a dedicated statement. Describe what the tool did and what authors verified. “AI was used” is too vague. A complete prompt dump may be unnecessary and could disclose protected information; follow the actual rule.

A suitable factual pattern is: “We used [tool, version/date] to suggest language edits in Sections X and Y. The authors compared every change with the source text, verified terminology and approved the final wording.” Replace every field with truth. For code or analysis, name the test and validation rather than borrowing this language.

Assign accountability inside the author group

Discuss use before submission. The operator reports task, data flow and retained output. Relevant subject authors check claims, code and references. Every author sees the final disclosure and approves the manuscript. If nobody can explain or validate a generated component, it is not ready for publication.

Confidential review creates additional limits. The peer-review data-flow check addresses unpublished manuscripts. The authorship criteria guide maps contribution and accountability. The publication ethics hub joins related duties. The single foundation route is the evidence-review guide.

  1. Save journal and institutional policy.
  2. Inventory every AI-supported task.
  3. Review data flow and confidentiality.
  4. Verify facts, sources, language and code.
  5. Name the person responsible for each control.
  6. Draft the declaration in the required location.
  7. Archive author approval and submitted version.

The central question is not how much the model produced. It is whether identifiable people can inspect, defend and correct every published element. Only those people can occupy the author role.

Correct an inaccurate AI declaration promptly

If use is discovered after submission, reconstruct it from real records rather than guessing prompts. Inform co-authors and the editorial office, state the affected material and propose accurate wording. Preserve the previous declaration and correspondence. The editor decides whether a form update, manuscript change or published correction is required.

Do not name a system as author to make its use “transparent”. That misstates responsibility. Transparency comes from a specific tool-use declaration and a verifiable human workflow, not from assigning person-like status to software.

Official foundations

  1. ICMJE: Use of Artificial Intelligence in Publishing – AI cannot qualify as an author; use and accountability require disclosure.
  2. COPE: Authorship and AI tools – editorial principles for attribution and transparency.
  3. WAME: Chatbots, Generative AI, and Scholarly Manuscripts – recommendations on responsibility, source checking and declaration.

Sources reviewed 4 September 2026.