Can you use AI to write your thesis?

Short answer: at most universities, yes, with conditions. The era of blanket bans is largely over; what replaced it is a disclosure-and-accountability standard. Whether AI use in your thesis is legitimate depends on three things: what your institution's policy allows, whether you remain the intellectual author, and whether you can show exactly what the AI did.

This guide takes each of those in turn: what policies actually say and how to read yours, which uses are usually fine and which usually are not, where the misconduct line really sits, why disclosure fails without records, and what a defensible setup looks like in practice. It ends with a checklist you can act on this week. None of it requires you to be an enthusiast or an abstainer. It requires you to be documented.

What university AI policies actually say

Policies vary enormously (by university, by faculty, sometimes by supervisor), but most converge on the same skeleton: AI may assist with tasks like literature discovery, language polishing, and summarisation; it may not replace your original analysis or argument; and its use must be disclosed. Some institutions require a dedicated AI-use statement in the front matter; others fold it into the declaration of originality; a growing number ask for both a short declaration and a fuller account in the methods chapter.

It helps to recognise the broad families you will meet. Restrictive policies permit AI only for narrow tasks, often language editing, and require approval for anything else. Disclosure-based policies permit a wide range of uses provided each is declared and you remain responsible for every word; many institutions have converged on this family. Integrated policies go further and treat AI competence as something to be taught, with the same disclosure backbone underneath. Knowing which family yours belongs to tells you how much of this guide is about permission and how much is about documentation.

Three layers can govern you at once. The university sets the outer rule, your faculty or graduate school may narrow it, and your supervisor's expectations fill whatever is left. When the layers disagree, the strictest one wins in practice, because your supervisor signs off on submission and your examiners apply their own instincts. So the only universally safe advice is procedural: read the current policy, ask your supervisor how they interpret it, and get that interpretation in writing. Policies are changing fast, and the one that governs you is the one in force when you submit.

What most policies allow

With disclosure, the following uses sit comfortably inside most current policies. "Usually" is doing real work in each line, so check yours. The pattern behind the list is worth noticing: these are all tasks where the AI accelerates work you direct and can inspect, rather than substituting for judgment only you can exercise.

• Literature discovery: searching databases, following citation trails, screening results for relevance.
• Reading support: summarising a paper you then read yourself, or extracting a paper's stated claims with their locations.
• Language editing: grammar, clarity and concision on text you already wrote.
• Mechanics: reference formatting, structural templates, compiling the document.
• Analysis assistance where you choose the method and can explain every result: running tests you specified over your own data.

What most policies prohibit

The prohibitions cluster around substitution: the AI doing the intellectual work the degree certifies. Most policies bar some or all of the following, and even permissive ones expect disclosure.

• Submitting AI-generated text as your own analysis or argument without disclosure.
• Citing sources you have not read, or references you have not confirmed exist.
• Having AI produce interpretations of your results that you cannot explain or defend.
• Using AI to paraphrase other people's work in ways designed to evade originality checks.
• Fabrication in any form: data, quotes, references, or study findings.

The real line: authorship and verifiability

Between those two lists sits a gray zone that policies handle inconsistently: AI-generated first drafts you substantially rewrite, translation of your own writing, brainstorming research questions, AI feedback on a draft you wrote. Reasonable institutions land in different places on each, so put those cases to your supervisor in writing before you rely on them. But the way through the gray zone is the same principle examiners apply everywhere, and it is worth stating plainly.

Examiners do not actually care whether a sentence passed through a model. They care whether you own the argument and whether every claim survives scrutiny. AI use becomes misconduct when it launders unverified content into your voice: invented references, claims no source supports, analysis you cannot explain in the viva.

That reframes the question. "Can I use AI?" matters less than "can I defend everything the AI touched?", which is a records problem, not a willpower problem. A student with modest AI use and no records is in a weaker position than a student with extensive AI use, full disclosure, and an audit trail, because the second student can answer every question with evidence.

Authorship, meanwhile, is about decisions, not keystrokes. You chose the question, selected and read the sources, decided which claims to keep, structured the argument, and take responsibility for every sentence. A tool that drafts from sources you selected, under constraints you set, with every claim checked and reviewable, leaves those decisions with you. A tool that invents content you paste in wholesale does not.

A worked example: the viva question that decides it

Picture the same viva question put to two students. The examiner points at a sentence in chapter two: "This paper you cite here, does it actually support the claim as you have worded it?"

Student A drafted with a general chatbot months earlier. They do not remember whether that sentence was theirs, whether the citation came from a search or a suggestion, or whether anyone ever checked it against the paper. Whatever the truth is, the only honest answer available to them is "I believe so". If the citation turns out to be wrong, the conversation is now about integrity rather than about the thesis.

Student B drafted the same section with grounded tools. The claim links to a citation that resolves to the actual paper, carries a verdict showing it was checked against the paper's full text, and sits in an audit trail showing when. Their answer: "Yes, and I can show you the passage." Same question, entirely different afternoon. The difference was not talent or honesty. It was records.

Now multiply that one question by the dozens a viva contains, and by the years between drafting a sentence and defending it. Memory does not survive that multiplication for anyone. Records do, which is why the rest of this guide is mostly about how to have them without doing extra work.

Where AI use goes wrong in practice

The catastrophic failure is the invented reference. General chatbots generate citations as free text, which means some of them will be to papers that do not exist, formatted convincingly enough to survive a casual check. Students have submitted them and been caught. The mechanism, and why prompting cannot fix it, is covered in Why AI makes up citations, and the fix.

The quieter failures matter just as much: a real paper cited for a claim it never makes; a qualified finding flattened into a confident one; a paragraph of plausible synthesis with no source behind it. None of these look like cheating while you write. All of them look like cheating, or like carelessness, which examiners treat little better, when they are found. And the viva is designed to find them.

Every failure in this list has the same shape: distance between a claim and its source, with nothing in between that checked. Close that distance and most of the risk of using AI in a thesis simply disappears.

Timing makes these failures expensive. A fabricated reference caught in week ten of writing costs you an hour. The same reference caught by an examiner working through your bibliography costs you, at minimum, a corrections cycle and your credibility for the rest of the examination, because once one citation fails, every other citation gets read with suspicion. Checking early is not a virtue here; it is simply the cheap option.

Disclosure without records is guesswork

Most students who try to disclose honestly hit the same wall months later: they cannot remember what the AI actually did. Which sections had AI-drafted first passes? Which citations came from a search versus a suggestion? Which claims were checked, and against what? A disclosure written from memory is either vague to the point of uselessness or accidentally wrong, and a wrong disclosure is worse than none, because it is a false statement in a document about integrity.

The fix is to keep records as you work, not to reconstruct at the end. If you use general tools, keep a simple log: date, tool, what you asked, what you kept, how you checked it. A spreadsheet is enough, and per task is the right granularity: "used X to summarise the papers in chapter 2's second theme, then read each summary against the paper" is a defensible line; "used AI for research" is not. If you use CiteDash, the record keeps itself. Either way, write the statement from the record. Templates and placement guidance are in How to write an AI use disclosure statement for your thesis.

How CiteDash AI makes thesis writing defensible

CiteDash AI was built for the disclosure-and-accountability standard. Every action (every AI draft, every fact-check, every search, every compile) lands in the project's audit trail, and the AI-use disclosure document is generated from that real record, not reconstructed from memory.

Every AI-drafted claim carries a citation that resolves to a real paper and a verification verdict from the Fact Checker against that paper's full text, so "can you defend what the AI touched?" has a checkable answer for every sentence. Retracted papers are blocked from citation, and an originality pre-check screens the text before anything is submitted.

Authorship stays yours by design: milestones are only ever marked done by you, engine outputs become part of your thesis only when you adopt them, and the trust gate at compile makes you consciously acknowledge anything unresolved. Submission Readiness then reports the state of the whole document, verification, originality, retractions and disclosure together, so nothing arrives at your examiner unexamined.

In day-to-day terms this changes very little about how writing feels and everything about how submission feels. You search, read, draft and revise the way you would anyway. The difference is that when the disclosure statement is due, it is generated rather than remembered, and when a specific sentence is questioned, the answer is a click away rather than a reconstruction.

Objections, answered

"My supervisor distrusts all AI." Often what they distrust is undisclosed AI, and they are right to. Bring the process, not the argument: show them the policy, describe exactly which tasks you will use AI for, and offer the audit trail. A supervisor who can see what the AI did and how it was checked is being asked to trust records, not promises.

"What about AI detectors?" Detectors produce false positives, particularly for students writing in a second language, and most institutions know it. Your protection is not hoping a classifier stays quiet; it is an accurate declaration plus records that show what happened. A student with both has an answer to any detector result. A student with neither is negotiating from memory.

"The policy might change before I submit." It might, and new policies are frequently applied to work already in progress. That argues for keeping records now, not against it: an accurate log satisfies a stricter future rule, while reconstructed memory satisfies nothing. Disclosing accurately costs you nothing under any plausible policy.

"Will disclosing make my examiners suspicious?" Increasingly the opposite. A clear, specific statement reads as methodological transparency, the same instinct that makes you describe your sampling strategy, and examiners are learning to treat vague or absent statements as the red flag. What damages you is not disclosed AI use; it is discovered AI use, and the gap between the two is exactly the statement plus the records behind it.

"If AI helped, is it still my thesis?" Ask the same question of every tool the academy already accepts: reference managers, statistics software, spell checkers, supervisors who suggest restructuring a chapter. The degree certifies your question, your judgment, your argument and your defence of it. Keep those and the thesis is yours; outsource them and it is not, whether or not an AI was involved.

A practical checklist for using AI in your thesis

Everything above compresses to eight actions. None of them takes long, all of them are cheaper than a single integrity query, and the first two are worth doing this week whatever stage you are at.

• Read your university's current AI policy in full, and your faculty's or graduate school's if a separate one exists.
• Confirm your supervisor's interpretation, in writing.
• Use AI for grounded assistance (discovery, drafting from real sources you hold, checking), never for un-sourced content.
• Verify every AI-touched claim against its source's full text. Never submit what you cannot defend in the viva.
• Keep records as you go, or use tools that keep them for you.
• Generate the disclosure from the records, in the format your institution expects.
• Run originality and retraction checks before submission, not after a query arrives.
• Keep the audit trail available in case anyone asks.

The bottom line

Yes, you can use AI to write your thesis at most institutions, if you read the policy, stay the author, and can prove what the AI did. The students who get this wrong are rarely malicious; they are undocumented. The fix is unglamorous and completely available: grounded tools, honest verdicts, records kept as you work, and a disclosure written from evidence rather than memory.

If you are mid-thesis and none of this is in place, start where you are rather than despairing about the chapters behind you. List the sections AI touched as best you remember, verify the citations in them against real sources now, and keep proper records from today forward. A disclosure that is precise about recent work and honest about earlier uncertainty is still a defensible disclosure.

That combination does more than keep you safe. It makes the thesis better, because every claim in it has already survived the scrutiny your examiner is about to apply.