AI and Authorship Ethics in Scientific Research and Peer Review

Picture two small disasters. A genetics paper gets pulled from review because a figure caption still reads “as an AI language model, I cannot…” A reviewer pastes a confidential manuscript into a chatbot to “help summarize it” and quietly breaks their confidentiality agreement. Neither person thought they were doing anything wrong. That gap — between intent and the actual rules — is where most of the trouble with AI in research lives.

So let’s lay out what the rules actually are, where the genuine ethical lines fall, and how to use these tools without putting your name, your co-authors, or your reviewers at risk.

Key takeaways

• An AI tool cannot be an author. Authorship means accountability, and software can’t be held accountable.
• Using AI is generally allowed; hiding it is the breach. Disclose where and how you used it.
• Peer review has its own confidentiality rules. Don’t paste unpublished manuscripts into public chatbots.
• Disclosure expectations differ for language polishing versus generating text or analysis — check the specific journal.
• Your reviewer is a human who’s accountable for their judgment. Write for that person, not for a detector.

Why AI can’t be an author (and what that actually means)

This one got settled fast. In early 2023, after a handful of preprints listed ChatGPT as a co-author, the major standard-setters drew a hard line. Nature said outright that no AI tool would be accepted as a credited author. The Committee on Publication Ethics (COPE) and the International Committee of Medical Journal Editors (ICMJE) landed in the same place, and their reasoning is the part worth understanding.

Authorship isn’t just “this thing produced words.” It carries real obligations: accountability for the integrity of the work, approval of the final version, and the ability to answer for it later if questions come up. A model can’t take responsibility for a fabricated citation. It can’t approve anything. It won’t show up to defend a methods section three years from now. Strip those obligations away and the credit means nothing.

Here’s the practical version. If you used a model to draft, edit, code, or analyze, you are still the author and you own every claim in the paper, including the ones the tool generated. The AI gets disclosed, not credited. Treat it the way you’d treat a statistics package or a reference manager — a tool you used and should mention, not a colleague who shares the byline.

The accountability problem, made concrete

Say a model invents a plausible-looking citation: a real journal, real-sounding authors, a DOI that goes nowhere. That’s a known failure mode. If it slips into your reference list, “the AI wrote it” is not a defense anyone will accept. You signed off. The accountability was always yours. This is exactly why the author-versus-tool distinction matters — it keeps the responsibility attached to a person who can actually fix the mistake.

Disclosure: the line between using AI and hiding it

Most journals don’t ban generative AI. What they ban is using it silently. The expectation now is a short, honest statement of what you did.

A typical disclosure lives in the methods or acknowledgments and reads something like: *”The authors used [tool] to assist with language editing of the manuscript. All scientific content, analysis, and conclusions are the authors’ own, and the authors reviewed and verified all output.”* If the tool did more — generated draft text, wrote analysis code, summarized literature — say that specifically.

How much detail you owe depends on what the tool actually did, and that’s where people trip:

• Light language help (grammar, phrasing, a Grammarly-style pass) is often treated leniently, and some publishers explicitly exclude routine editing tools from disclosure requirements.
• Generating text, figures, code, or analysis is the part nearly every policy wants spelled out.
• Anything that touches the data or the findings gets the closest scrutiny, because that’s where undisclosed AI can distort the scientific record.

The safe rule: when you’re unsure which bucket you’re in, disclose. A disclosure that turns out to be unnecessary costs you nothing. The reverse — undisclosed use that surfaces later — is what gets papers corrected, flagged, or retracted.

A mini-scenario

Two PhD students use the same chatbot the same way, to tighten the prose in their discussion sections. Student A adds one sentence to the acknowledgments noting AI-assisted language editing. Student B says nothing, figuring it’s too minor to mention. A reviewer notices some stock phrasing in Student B’s draft and asks the editor about it. Now Student B is explaining an omission instead of pointing to a line they already wrote. Same tool, same use, completely different position. The only difference is that one of them disclosed.

AI in peer review: a different set of rules

Authoring a paper and reviewing one are not the same job, and the AI rules diverge sharply.

The core issue is confidentiality. A manuscript under review is unpublished, often months from public, and shared with you in confidence. Paste it into a public chatbot and you’ve sent someone else’s unpublished work to a third-party company whose data handling you don’t control. That’s the breach — not the act of using software, but the leak.

This isn’t hypothetical. In 2023 the US National Institutes of Health prohibited reviewers from using generative AI tools when evaluating grant applications, citing exactly this confidentiality and integrity concern. Several publishers followed with their own bans or tight restrictions on AI in the review workflow.

So what’s the safe move as a reviewer?

• Don’t upload the manuscript into any external AI tool. Treat it like a document under embargo, because that’s effectively what it is.
• Don’t outsource your judgment. The value you bring is expert assessment a model can’t replicate, and the editor is trusting *your* name.
• If you use AI at all, limit it to polishing the wording of comments you’ve already written yourself — and check whether the journal even permits that. Some don’t.

A reviewer who lets a chatbot write the review isn’t just cutting a corner. They’ve handed confidential work to an outside system and put their name on an assessment they didn’t actually make.

Where detectors fit (and where they don’t)

Editors and reviewers sometimes reach for AI detectors, so it’s worth being clear about what those can and can’t tell you. They output a probability, never proof. And they have well-documented blind spots: a 2023 Stanford study led by Weixin Liang found that several detectors flagged writing by non-native English speakers as AI-generated far more often than native-speaker writing. That alone should stop anyone from treating a detector score as a verdict in a global research community.

If you’re a researcher, the useful framing is to treat detectors as one signal, not a verdict. They lean on statistical patterns like low perplexity and even sentence rhythm — patterns that also show up in careful, formal academic prose written entirely by a person. Running a section through a checker can flag genuinely robotic, over-smoothed writing worth revising. But the goal isn’t a low number. The goal is a methods section a specialist can follow and a discussion that reflects your actual reasoning.

The honest takeaway: detection technology is too shaky to be the backbone of research integrity. The backbone is disclosure plus human accountability. That’s also why OpenAI quietly retired its own AI text classifier in mid-2023 — its accuracy wasn’t good enough to be useful. If the people who build these models pulled their own detector, an editor probably shouldn’t be deciding your paper’s fate on one either.

A short checklist before you submit

• Read the journal’s specific AI policy. They’re not uniform. Don’t assume.
• Disclose AI use in the methods or acknowledgments, with the right level of detail for what you did.
• Verify every AI-touched fact and citation yourself. Assume nothing the tool produced is correct until you’ve checked it.
• Never list AI as an author. Credit goes to accountable humans.
• If you’re reviewing, keep the manuscript out of external tools entirely.

Frequently asked questions

The questions below cover the issues that come up most once researchers start using these tools in earnest.

The short version

The ethics here are less mysterious than they look. AI is a tool you may use and must disclose; it is never an author, because it can’t be accountable. In peer review, confidentiality comes first. And no detector — however confident its percentage — replaces a human being who can stand behind the work. Get those three things right and you’re on solid ground.

If you want to sanity-check how a section reads before you submit, or see whether your prose has drifted into stiff, over-smoothed phrasing, run it through PaperBleach and treat the result as a signal, not a scoreboard. You can also browse more articles on AI detection and ethics while you’re at it.