Most publishers would agree that authors remain responsible for their work, whether or not artificial intelligence helped produce it. Putting that principle into a policy is straightforward. Applying it is less so.
What happens when an author openly declares that AI drafted several chapters? When a routine language-editing task involves sharing confidential material? Or when an editor suspects undisclosed AI use but has no reliable evidence?
These questions turn broad principles into difficult editorial decisions. A document can look complete and still offer little help when a genuine gray area arises.
Working on an AI policy for authors reinforced this distinction for me. Researching other publishers’ rules was useful, but the harder work lay in deciding how those principles would operate across different publications and workflows.
An effective policy must protect publishing integrity without unnecessarily restricting legitimate assistance. Authors should understand its boundaries, and editors should have realistic ways to apply them.
Start With Your Publishing Reality, Not Someone Else’s Policy
Existing guidance is a sensible starting point. The Committee on Publication Ethics (COPE) addresses authorship and accountability; the International Committee of Medical Journal Editors (ICMJE) discusses transparency and confidentiality. Publishers such as Elsevier and Springer Nature have their own requirements.
These references are valuable, but copying them without considering context can create a policy that fits someone else’s publishing program better than yours.
A scholarly journal, a university press publishing textbooks and popular books, and a trade publisher specializing in biographies or illustrated fiction encounter different kinds of risk. Research integrity may dominate one workflow; narrative originality, translation, personal likenesses, or image rights may matter more in another.
Before drafting clauses, map where authors and contributors might use AI: research support, writing, translation, illustration, and materials supplied for marketing. Establish which publications and contributors are covered, and identify existing contractual obligations.
Also distinguish author-facing rules from the procedures editors and reviewers follow. They are related, but they need not occupy the same document. That separation can keep the author policy readable while giving staff more detailed instructions when necessary.
One broad policy with category-specific provisions may work. Separate policies may work better elsewhere. The right structure follows the publisher’s activities, not the format of another organization’s guidance.
Disclosure Is Not the Same as Permission
Using spell-check is different from asking an AI assistant to restructure an argument, and both differ from generating a chapter. Yet a policy that simply requires authors to “declare AI use” leaves these distinctions unresolved.
The STM Association’s classification recommendations offer a starting framework for deciding which kinds of assistance warrant disclosure. But classifying an activity answers only the first question.
Disclosure does not automatically make the activity acceptable.
Imagine an author submitting a general readership book. The author declares that several substantial passages were generated by AI and later edited. That openness helps the publisher evaluate the manuscript, but it does not settle whether the work meets its standards.
Was the author’s intellectual and creative contribution sufficient? Were the passages checked? Does this amount of AI-generated writing comply with the publisher’s requirements?
Publishers can reasonably draw different boundaries. Elsevier’s guidance for books and commissioned content emphasizes authentic human contributions and sets particular conditions for AI-assisted writing. Springer Nature’s editorial guidance describes a risk-based framework for its scholarly workflows. Neither should be treated as a universal template.
A workable policy distinguishes routine assistance, activities requiring declaration or assessment, and uses it prohibits even when disclosed. It should specify what declarations contain: the tool, purpose, affected material, and the author’s verification, where relevant.
Finally, decide whether disclosure is needed at submission, in the published work, or both. A submission declaration helps editors assess the manuscript; a published statement informs readers. Neither should become a meaningless formality. Treating every honest declaration as suspicious could discourage the transparency the policy is meant to encourage.
The Language-Editing Trap
An acceptable AI task can still be performed in an unacceptable way.
Language assistance illustrates the problem. AI can improve clarity and grammar, especially for authors writing outside their primary language. Banning that assistance indiscriminately may create unnecessary barriers.
Now imagine a contributor preparing a book containing unpublished interviews subject to confidentiality agreements. To improve readability, the contributor uploads several chapters into a general-purpose AI service.
The editing task may be permitted. Sharing the material might not be.
The content could include personal information, restricted interviews, or third-party material that the contributor is not authorized to submit. The service’s retention, reuse, and training terms may introduce additional concerns.
A policy must therefore answer two distinct questions: What may AI be used for? And what material may be provided to which systems?
The distinction matters even more during editorial assessment. Submitted manuscripts and review correspondence often involve confidential material. The ICMJE guidance for editors cautions against uploading submitted manuscripts to systems where confidentiality cannot be assured without authors’ explicit permission.
Not all AI services operate under identical conditions. A publisher should examine rights to share the material, data handling, contractual protections, and authorization rather than treating every tool as equally safe or unsafe.
For some content, an appropriately approved service may be suitable. For other material, uploading it may be inappropriate regardless of how harmless the task seems. Where terms or permissions are unclear, authors need a way to seek guidance before sharing sensitive content.
The principle is simple: permission to use AI for language editing is not necessarily permission to share the manuscript.
Human Accountability Needs an Operational Meaning
Most publishing guidance keeps responsibility with people. COPE, for example, holds that AI tools cannot be authors of scholarly papers because they cannot assume authorship responsibilities.
But “the author remains responsible” should not be the end of the discussion.
If AI summarizes research, someone must check the original sources. If it assists with translation, a competent person must verify terminology, meaning, and context. Generated quotations, references, and factual claims need independent checking.
Accuracy alone is insufficient, however. A carefully verified AI-generated chapter may still fail to reflect the degree of original intellectual or creative contribution a publisher expects.
Publishers must also consider attribution, permissions, and rights. Because AI copyright rules vary across jurisdictions, a publisher should not turn an uncertain legal interpretation into a universal requirement.
Make accountability concrete: explain what authors must verify, what uses they must disclose, when permission is necessary, and when editors may request clarification or supporting evidence. An editor investigating a questionable citation needs the original source, not a broad assurance that the author has reviewed the manuscript.
Such requests should be proportionate rather than demanding exhaustive prompt logs from every writer.
Human responsibility becomes meaningful when authors know what actions it entails.
Can Your Editors Actually Apply the Policy?
Consider a third scenario. An editor runs a manuscript through an AI detector, which flags substantial portions as machine-generated. The author denies using AI.
A detection score cannot settle the dispute. The finished text alone is weak evidence of how a book was written; it cannot reveal the drafts, decisions, and revisions behind a manuscript.
Springer Nature’s guidance for editors warns that detection tools are not definitive. It also cautions that uploading unpublished manuscripts to detection services can itself compromise confidentiality, depending on the service’s practices.
An editor trying to enforce a policy could therefore create another problem.
The better question is not how to catch every undisclosed use of AI, but how to respond proportionately to concerns about manuscript integrity.
A small internal procedure can distinguish three situations. Routine permitted declarations may require no special action. Ambiguous declarations may justify asking the author for clarification. Serious concerns, such as fabricated sources or unauthorized material, may warrant verification and escalation.
Authors should ordinarily have a fair opportunity to respond before adverse decisions, subject to the circumstances and applicable procedures. Editors need consistent standards without being expected to become forensic investigators. A detector result might prompt a closer look if other concrete concerns exist, but it should not become proof by itself.
The issue reaches beyond detection. Publishers need to preserve editorial judgment as AI enters publishing workflows, rather than substitute automated signals for independent assessment.
A small press may need only a concise guide and a designated person for difficult cases. A larger publisher might require more formal arrangements. In either case, a policy should not impose duties that editors have no realistic means of performing.
Good Policies Need More Than One Perspective
Editorial staff understand manuscript quality and publishing workflows, but they cannot resolve every question alone.
Confidentiality may require information-security expertise. Rights, privacy, and contractual questions may need legal review. Operational staff can identify burdens that look minor on paper but become substantial across hundreds of submissions.
Consultation is valuable because different reviewers spot different weaknesses. It is not, however, a vote in which every suggestion must be adopted. Feedback may conflict, duplicate existing provisions, or propose safeguards that cannot reasonably be implemented.
Publishers need to make defensible choices and establish who has authority to approve them. For smaller organizations, focused editorial review with targeted external advice may be more sensible than creating a complex committee structure.
Ownership after approval matters too. Someone needs responsibility for communicating the rules and reviewing them when technology, legal requirements, or publishing practices change. Stability is important; a policy should not be rewritten whenever a new AI product launches.
Five Decisions to Make Before You Publish the Policy
The work can begin with five practical decisions rather than a lengthy document.
- Define the scope. Identify publications, contributors, materials, and activities covered by the policy.
- Set boundaries. Separate routine assistance from uses requiring disclosure, assessment, or prohibition.
- Protect content. Specify the confidentiality, privacy, and rights conditions governing what can be shared with AI services.
- Plan editorial responses. Give staff simple procedures for routine declarations, clarification, serious concerns, and escalation.
- Assign ownership. Decide who approves the policy, communicates changes, and maintains it.
The formal document may need supporting materials: a short FAQ for authors, an accessible declaration mechanism, and internal guidance for editors. Not everything belongs in the policy itself.
This is a starting sequence, not a universal compliance checklist. The necessary detail should follow the publisher’s publications, obligations, and resources.
The Test of a Useful AI Policy
It is tempting to judge a policy by how many situations it anticipates. Yet a comprehensive document can still leave authors confused and editors unsure what to do.
The better test is practical. When an unfamiliar AI-related situation arises, can an author understand what is expected? Can an editor apply the rules consistently and fairly? Can the publisher explain why its boundaries exist?
Different publishers may give different answers about acceptable AI use. That is not necessarily a flaw. Unclear expectations and unworkable procedures are.
An AI policy should support sound publishing decisions, not merely demonstrate that an organization has addressed artificial intelligence.
Writing the rules is a beginning. Making them work is the real challenge.