How to write an AI use disclosure statement for your thesis

Most universities have moved from banning generative AI to requiring that you declare it. The rule is usually short and the guidance is usually thin: you are told a statement is required, rarely what it should contain, where it should go, or how much detail is enough. This guide gives you the structure examiners are looking for, five statements you can adapt, a decision guide for how much detail to include, and the part most students miss, which is being able to evidence what you claim.

One point before the templates. A disclosure is not a confession. Declaring that you used a tool to search literature or check grammar does not weaken your thesis; concealing it and being found out does. Written well, the statement reads as methodological transparency, the same instinct that makes you describe your sampling strategy or your inclusion criteria without being asked twice.

The other framing worth fixing early: the statement is a claim, and claims can be examined. If your viva panel asks how a declared verification process actually worked, you want an answer grounded in records rather than memory. Everything below is built around that standard, because it is the standard your examiners will apply.

What an AI use disclosure statement has to answer

Wording varies by institution, but nearly every policy is asking the same four questions. If your statement answers these plainly, it will satisfy most rubrics.

• Which tools did you use? Name them, including the version or model where you know it. "An AI assistant" is not a name.
• What did you use each one for? Be specific: literature searching, summarising, translation, grammar, code, data analysis, drafting.
• What did you not use them for? This is the sentence that does the most work, because it bounds the claim and tells the examiner where the tool stopped and you began.
• How did you verify the output? Examiners care far more about your checking process than about the tool.

The fifth question every AI declaration should answer

There is a fifth question that most policies leave implicit: who is responsible for the final text? The answer is always you, and it is worth saying so explicitly. A sentence such as "I reviewed and take full responsibility for all text in this thesis" costs you nothing and signals that you understand the authorship standard your institution applies. Examiners read a lot of these statements; the ones that end with clear ownership read as confident, and the ones that trail off into passive voice read as hedging.

Notice what is absent from the list: justification. You do not need to argue that AI use is acceptable. The policy already settled that question when it permitted disclosed use. Statements that get defensive, or that quote the policy back at the examiner, read worse than statements that simply describe what happened.

One calibration point while we are on content: precision about tasks matters more than precision about prompts. Nobody needs the text of every prompt reproduced in the statement itself; that level of detail belongs in your log, where it can be produced if asked. The statement needs the category of use, its boundary, and the verification step, in language a non-specialist examiner can read in thirty seconds.

The same goes for tool mechanics. You do not need to explain how a language model works, and attempting to usually goes badly. Describe what each tool did in your project, in your own words, at the level of the research process: searched, extracted, drafted, checked. That is the level at which the examiner will engage with it.

Before you write: reconstruct what you actually did

The most common failure in a disclosure is not dishonesty, it is amnesia. By the time you write the statement, months have passed since the work it describes. Before drafting anything, take twenty minutes and build an inventory.

Go through each phase of the project and note every tool that touched it, including the unglamorous ones. Grammar checkers with generative rewrite modes count. Translation tools count if you drafted in one language and submitted in another. Code autocomplete counts if the scripts produced your results. The stages to walk through:

• Discovery: search tools, recommendation systems, citation-map exploration.
• Reading: summarisers, question-answering over PDFs, note extraction.
• Synthesis: evidence tables, thematic clustering, comparison drafts.
• Analysis: statistical software, AI-assisted code, qualitative coding support.
• Writing: drafting, rewriting, grammar and clarity edits, translation.
• Citation and assembly: reference formatting, bibliography generation, compilation.

Five AI disclosure statement examples you can adapt

For each inventory entry, note what you kept and how you checked it. That inventory is the raw material for the statement, and it doubles as your first line of evidence if anyone asks a follow-up question. Then adapt one of these five statements to what you actually did. Do not copy one that overstates or understates your use: the statement is a claim you may be asked about in a viva.

1. Search and discovery only. "I used CiteDash AI to search the literature across OpenAlex, PubMed, Semantic Scholar and arXiv, and to organise the papers I selected. All screening decisions, all analysis and all writing are my own. No text in this thesis was generated by an AI system." Use this when AI never touched your prose. It is the shortest defensible statement, and the closing sentence is what makes it strong.

2. Language editing only. "I used generative AI tools for language editing: grammar, clarity and concision on text I had already written. No content, argument, citation or analysis was produced by an AI system, and I reviewed every suggested edit before accepting it." Use this if you wrote every sentence yourself and used AI the way you would use a copyeditor. The phrase "text I had already written" is doing the bounding work.

3. Assisted drafting with verification. "I used CiteDash AI to draft sections of chapters 2 and 4 from sources in my own library. Every AI-assisted claim carries a citation to a real paper, and each was verified against that paper's full text before inclusion. I reviewed, edited and take full responsibility for all text in this thesis. An audit trail of every AI action and verification is available on request." This is the statement for grounded AI drafting. It only works if the verification actually happened and left a record; do not borrow the audit-trail sentence unless you can produce one.

4. Data analysis support. "I used CiteDash AI to run descriptive and inferential statistics on my survey data. Tests were selected by me and computed by software rather than generated by a language model. The interpretation of results in chapter 5 is my own." The key distinction here is between computation and generation. Statistics computed by software from your data are a different category from numbers produced by a language model, and examiners know the difference.

5. Mixed use across the project. "I used CiteDash AI throughout this project: to search the literature, to read and extract from papers in my library, to draft sections of chapters 2, 4 and 5 from those sources, and to format the bibliography. Every AI-assisted claim cites a source held in my project library and was verified against that source's full text before inclusion. Screening decisions, the analysis in chapter 5, and all final wording are my own. I reviewed and take full responsibility for the entire thesis, and an audit trail of AI actions is available on request." Use this when AI assistance ran through several stages. The structure to keep is tools, tasks, boundary, verification, responsibility, in that order.

How much detail? A decision guide

Length is not virtue. A statement that is precise in three sentences beats a page of hedging. Use this as a rough decision table:

• Your policy specifies a template or form: use it exactly, and put any extra detail in a methods subsection or appendix rather than editing the required wording.
• AI only searched, organised or formatted: two or three sentences in the front matter is enough (example 1).
• AI edited prose you wrote: a short statement naming the editing scope (example 2), plus one sentence on how you reviewed the edits.
• AI drafted text you kept: name the chapters or sections, describe the grounding and verification process, and state responsibility (examples 3 and 5).
• AI assisted analysis: state the division of labour between you, the software, and any language model (example 4).
• You are unsure whether a use counts: declare it in one clause. Over-declaring a minor use costs a clause; under-declaring a major one costs credibility.

Where the AI declaration goes in your thesis

Check your handbook first, because a specified location overrides any general advice. In the absence of a rule, the conventional places are the declaration page in the front matter, immediately after the acknowledgements, or a short subsection at the end of your methodology chapter. Some institutions ask for both: a one-line declaration in the front matter and a fuller account in the methods.

A pattern that works well when your use was substantial: one bounded sentence on the declaration page ("AI assistance was used as described in section 3.7"), a paragraph in the methods giving the full account, and the exported log or audit trail as an appendix or an on-request document. This keeps the front matter clean while giving the examiner a place to look if they want depth.

If your supervisor reviews drafts, show them the statement before submission. A disclosure your supervisor has already read and agreed with is one fewer surprise in the room, and their sign-off is itself a small piece of evidence that the process was open.

Evidence: the part most students miss

A statement is an assertion. If your examiner asks how you know a particular AI-assisted claim is supported by the source it cites, "I checked" is a weaker answer than a record showing when it was checked and against which text.

This is the reason CiteDash records every verification rather than only performing it. Every claim carries a verdict of supported, partial, unsupported or unverified from the Fact Checker, checked against the cited paper's full text rather than its abstract, and the whole trail exports as an AI-use disclosure document you can attach or hand over. Submission Readiness reports it alongside your originality pre-check, so the statement you write is backed by the record the system kept while you worked.

If you are using other tools, keep your own log, and keep it as you work rather than at the end. A spreadsheet is enough. Useful columns:

• Date, and the stage of the project you were in.
• Tool and version or model, as precisely as you know it.
• What you asked for, in a phrase.
• What you kept: nothing, an idea, edited text, or verbatim text.
• How you checked it, and against which source.

Common mistakes in AI use declarations

The point of the log is not bureaucracy. The point is that when the question comes, you can answer it rather than try to remember the answer. Most weak statements fail in one of a few predictable ways:

• Vagueness. "AI was used to assist with this thesis" tells an examiner nothing and invites the questions you were hoping to avoid.
• Declaring the tool but not the task. Naming ChatGPT without saying what you used it for is the most common weak statement.
• Omitting verification. The checking process is the part that earns trust; leave it out and the statement reads as a disclaimer.
• Retro-fitting. Writing the statement at the end, from memory, is how understatement happens. Keep notes as you work.
• Copying a template that names tools you never used, or omits ones you did. The statement must match your project, not the internet's favourite wording.
• Overstating non-use. "No AI was used at any stage" is a strong claim; if a grammar tool with a generative mode touched your text, it is also a false one.
• Assuming a detector is the standard. AI detectors produce false positives, particularly for students writing in a second language. A clear declaration plus an audit trail is a far stronger position than hoping a classifier stays quiet.

Objections to declaring AI use, answered

"Won't declaring AI use make my examiner suspicious?" Almost always the reverse. An examiner who finds an accurate, bounded declaration knows exactly what to check and finds that it checks out. An examiner who discovers undeclared use has a misconduct question, not a methods question. The risk sits almost entirely on the concealment side of the ledger.

"My supervisor told me not to bother." Get that in writing, and check it against the current policy anyway. Supervisors are not always current on institutional rules, and the policy in force at submission is the one that governs you. A two-line email confirming the agreed approach protects both of you.

"I did the work before the policy existed." Declare accurately anyway. Policies are frequently applied to work submitted after they land, whenever the work was actually done, and an accurate statement written now is cheap insurance against a rule written later.

"The statement will date my thesis." It will, in the same way a methods section dates every thesis: by honestly recording how the work was done. That is a feature of scholarly record-keeping, not a flaw.

Preparing to defend the statement in your viva

Treat the disclosure as viva material, because it is. The likely questions are predictable: which sections had AI involvement, how you verified a specific claim, what you did when the tool got something wrong, and whether you can explain any AI-assisted passage in your own words.

Preparation is mostly mapping. For each declared use, know which artifact backs it. The search declaration maps to your saved searches and screening notes. The drafting declaration maps to verified citations and their verdicts. The analysis declaration maps to your data files and test outputs. If you can walk from any sentence of the statement to a concrete record in under a minute, you are ready.

Rehearse the honest edge cases too. A claim that came back unsupported and was rewritten is not an embarrassment; it is the checking process working exactly as declared, and it makes a better viva answer than a claim you never tested.

If your institution has no AI policy yet

Declare anyway. Policies are arriving quickly and are frequently applied to work submitted before they landed. A short, accurate statement costs you nothing and protects you from a rule written after you submitted. Where no format is specified, example 1 or 3 above is a safe default, and the four questions at the top of this guide are the structure to follow.

If you are still deciding what AI use is legitimate in the first place, read Can you use AI to write your thesis? for the policy landscape. And for what a defensible checking process looks like underneath the statement, How verified citations work explains the four verdicts your evidence should be built on.