AI ethics in research writing: a practical checklist

The AI ethics conversation in academia can feel like it happens at thirty thousand feet: authorship theory, publisher statements, policy committees. Meanwhile you are at your desk with a half-written methods chapter, a chatbot tab open, and a much more practical question: what am I actually allowed to do here, and what should I do?

This is a working checklist for using AI ethically in research writing: the questions that sit behind most policies, concrete checks before, during, and after using AI, and how to keep your thesis defensible from day one. It is general guidance. Your institution's rules always come first, and where this post and your policy disagree, the policy wins. Keep your policy open as you read, because most of the checks below point straight back at it.

Start with your institution's policy, not the tool's marketing

AI rules vary widely between universities, between departments, sometimes between individual modules and supervisors. Funders and journals add their own layers on top. A tool advertising itself as safe for academic use tells you nothing about whether your programme permits it for assessed work.

So the first ethical act is boring: find and read the actual documents. Your university's academic integrity policy and any AI-specific guidance. Your programme handbook. The target journal's policy, if you are publishing. Your ethics approval terms, if your work involves participants. Write down what you find, and where anything is ambiguous, resolve it in an email with your supervisor. A dated message saying this use is fine is worth a great deal more later than your memory of a corridor conversation.

Expect the rules to move while you are enrolled, too. AI guidance is among the fastest-changing policy areas universities have, and what was unaddressed last year may be explicitly regulated this year. Recheck the current documents at the start of each academic year and before each major submission, and keep copies of the versions you relied on, because the version in force when you did the work is the one relevant to your choices.

The four questions behind most AI ethics policies

Policies differ in detail, but most of them reduce to four questions. If you can answer all four cleanly, you are in good shape almost anywhere:

• Accountability: who answers for the text? You do. Every sentence submitted under your name is yours to defend, whatever produced the first draft of it.
• Transparency: does anyone assessing the work know what role AI played? Disclosure requirements exist so that nobody has to guess.
• Accuracy: are the claims true and the citations real? Fabricated references are among the most common ways AI use turns into a misconduct case.
• Data: whose information went into the tool? Participant data, unpublished datasets, and co-authors' drafts carry obligations beyond your own convenience.

Checklist: before you use an AI tool

None of these checks require you to be an ethicist. They require you to be specific: which tool, which task, which data, which rule. Vague good intentions are where trouble starts, so turn intentions into named decisions before the work begins.

Ten minutes of setup prevents most of the serious problems. Before a tool touches your research writing:

• Read your institution's current AI guidance and note what is permitted, restricted, and banned for assessed work specifically.
• Agree with your supervisor, in writing, how you plan to use AI. Agreement up front beats explanation after the fact.
• Check what the tool does with your input: whether pasted text is stored or used for training, and whether that is compatible with your data obligations.
• Decide your own red lines now, while you are rested. Deadline pressure is when unexamined habits do their damage.

Checklist: while you work

Day to day, ethical AI use is mostly a handful of habits applied consistently:

• Keep AI in roles you can stand behind: searching, extracting, structuring, checking, critiquing. Be wary of roles that substitute for your understanding.
• Never paste identifiable participant data, confidential material, or co-authors' unpublished work into tools whose data handling you have not checked against your approvals.
• Verify every AI-touched factual claim against a source you hold. If you cannot trace it, it does not go in the draft.
• Keep a running note of what you used AI for. Reconstructing this honestly at submission time is far harder than recording it as you go.
• Watch for skill substitution: if you can no longer write the paragraph without the tool, that is a signal to slow down, not speed up.

Checklist: before you submit

The final pass is where ethics and quality control converge, and where an evening of careful checking can save a term of difficult correspondence:

• Resolve every citation to a real paper you have at least assessed yourself, and check that none of them have been retracted.
• Run an originality check before your institution runs one for you.
• Write your AI-use disclosure and have your supervisor read it before submission, not after a query.
• Reread the whole document asking one question: could I explain and defend every claim in this, out loud, without notes?

Disclosure: say what you did, plainly

Disclosure is where ethics becomes a paragraph. A good disclosure states which tools you used, for which tasks, and how you verified the outputs, in plain language. It should be specific and slightly boring. It also protects you: undisclosed AI use discovered later reads as concealment, while disclosed use is a conversation about whether the use complied with policy. Those are very different conversations to be having.

If you have not written one before, there is a full walkthrough in how to write an AI use disclosure statement for your thesis.

Common grey areas, worked through

Most day-to-day uncertainty lives in a few recurring cases. Grammar and clarity editing of text you wrote yourself is accepted in many places, but some programmes require disclosure even for that, so check rather than assume. Brainstorming and outlining are widely tolerated, because the intellectual work of selecting and developing the ideas remains yours; the same caution about checking applies.

Generated first drafts of assessed prose are where policies diverge most sharply: some programmes prohibit them outright, others permit them with disclosure and verification. Translating your own writing between languages is often treated like editing, but translating sources for use in your work still requires citation of the original. Code and statistical assistance usually fall under different rules from prose, often the rules that already governed help from a statistician. In every case the resolution is the same: your programme's written policy, plus a documented agreement with your supervisor wherever the policy is silent.

Data ethics: participants, privacy, and AI tools

If your research involves human participants, AI tools intersect with your ethics approval, not just your writing. Interview transcripts, survey free-text responses, and identifiable records were collected under specific promises about who would access them and how they would be stored. A consumer chatbot was almost certainly not among the recipients your participants agreed to.

The practical rules: anonymise before any AI processing. Prefer tools with clear commitments about how input data is handled. And if you are unsure whether your approval covers AI-assisted analysis, ask your ethics committee rather than guessing. Amending an approval is routine paperwork. Explaining a data breach to a committee is not.

Secondary data and copyrighted material deserve the same pause. Licensed datasets often carry terms that pasting into third-party tools can breach, and publisher PDFs are licensed for your reading, not necessarily for upload into any service you like. This is rarely an integrity issue in the misconduct sense, but it is an ethics issue in the keeping-your-promises sense, and the check takes minutes.

Authorship, supervisors, and journals

For publication, publisher and journal positions have broadly converged on two points, though you should verify the specifics for each venue: AI systems are generally not accepted as authors, because authorship carries responsibility a tool cannot take, and material AI use is generally expected to be disclosed. Where each venue draws the line between acceptable assistance and unacceptable substitution varies, so read the policy for the journal you are actually targeting.

With supervisors and co-authors, agree norms early: which tools, on which sections, disclosed how. The slightly awkward conversation is much shorter at the start of a project than after a disagreement about a submitted manuscript. For the larger question of where assistance ends and authorship begins, see can you use AI to write your thesis?

How CiteDash is built around these checks

CiteDash treats most of this checklist as architecture rather than user discipline. Generated claims are grounded in held full-text PDFs, and the Fact Checker verifies claims against their sources before they reach you. Citations are database objects that resolve to real papers, never free text, and retracted papers are badged and blocked from citation, so two of the classic failure modes are closed off by construction.

At submission time, the platform generates an AI-use disclosure, runs an originality pre-check, and Submission Readiness checks the document before your examiner does. The record of careful use builds itself while you work, which is exactly when it should be built.

The point of building checks into the platform is not to outsource your judgement. It is to make the ethical default the convenient one: verification runs whether or not you remembered to be careful that day, and the disclosure reflects recorded use rather than end-of-project memory. You still decide what role AI plays in your work. The system makes sure the record and the reality match.

Ethics as a practice, not a form

The checklist above is not really about staying out of trouble, although it does that. It is about producing research you fully own: claims you can defend, sources you have genuinely engaged with, methods you understand well enough to explain to a sceptical examiner. AI can extend that work honestly, and at its best it makes the careful path the fast path. The practice is making sure it never quietly replaces the work instead.

If you take one habit from this page, take the verification habit: no claim enters your document without a source you hold, and no source enters your bibliography without resolving to a real paper. Nearly every AI-related integrity failure that reaches an examiner passes through one of those two doors. Keep them shut and most of the rest is manageable.