Can AI write my thesis? No, and not because of the writing
Mostly no, and the reasons have nothing to do with how good the text is. A thesis is examined out loud by someone who has read it, developed over months in front of a supervisor whose judgement about your independence is often part of the grade, and submitted under a declaration of originality that carries your signature and, at a growing number of universities, an AI clause you settle with that same supervisor. Three checks, none of them a check on prose. This page is about what each of them actually asks of you, quoted from institutions that publish their rules, and about the narrower thing a tool is genuinely worth here.
The same question gets asked about three different documents, and the answers are not interchangeable. An essay is short and marked on argument, and machine drafting comes closest to working there, which how far a drafted essay gets takes apart against published marking bands. A research paper stands on whether its citations are real, and can AI write my research paper carries the audit numbers. A thesis is the one that is examined, and that changes the answer rather than merely raising the stakes. The division of labour below is the same at bachelor, master's and doctoral level; what scales is how long the supervisor watched and how long the room at the end lasts.
What is actually being examined
A thesis is the only piece of coursework that is assessed as a process as well as a document. Three facts about it do the work, and none of them is a judgement about writing quality.
You are examined out loud. Most European programmes end with a colloquium, a viva or a public defence, and in many of them it is a separate examination with its own grade. There is no version of that room where a chapter you did not build survives contact with someone who has read it closely.
Your supervisor watched it develop. Not the finished file, the sequence: the question that changed in October, the method you abandoned, the draft chapter that arrived thin. That history is evidence about you, it is formed over months, and in a fair number of programmes it is scored.
You sign for it personally. A thesis is submitted with a declaration of originality carrying your signature, and at a growing number of institutions the declaration now has an AI clause that you settle with your supervisor rather than tick alone.
Put those together and the honest answer is mostly no, but for a reason that is easy to miss. It is not that a model writes badly. On a short assignment the measured evidence says it writes well, which we go through in how far a drafted essay gets. It is that the thesis is not only the file. Part of what is being marked is whether you did it.
The documents with your name on them
The clearest way to see this is to look at the paperwork a thesis actually generates. These are real requirements at named institutions, quoted from what they publish. Yours will differ in the details and not in the shape.
| What you produce | Who else is in it | What it commits you to |
|---|---|---|
| Declaration of originality | Your supervisor, by agreement | At ETH Zurich it is "a component of every written paper or thesis authored during the course of studies", and "in consultation with the supervisor" you select one of two options: that you used no generative AI technologies, or that you used them as an authorised aid with source declarations agreed in that same consultation |
| AI disclosure table | Whoever reads the appendix | One University of Passau template pairs the declaration with a table, one row per use: the part or stage of the work, the AI system, the mode of implementation "(verbatim, paraphrased etc.)" and supporting evidence, given as a "link to chat or reference to a screenshot" |
| The supervision record | The person who watched | NHH assesses "independence under supervision" across generating the research question, reviewing the literature, developing the research design, executing the analysis and drawing conclusions, while stating plainly that independence "does not mean that you should avoid seeing your supervisor" |
| The colloquium | You, in a room, in real time | At Hochschule Bonn-Rhein-Sieg the bachelor's colloquium is an oral examination of 30 to 45 minutes which "complements the bachelor's thesis and is evaluated separately", and is open only to those who have already passed everything else |
Sources: ETH Zurich declaration of originality; University of Passau declaration of authorship with generative AI disclosure; NHH master's thesis assessment criteria; Hochschule Bonn-Rhein-Sieg, bachelor's thesis and colloquium.
Read the Passau row again. The supporting evidence a disclosure asks for is a link to the conversation. That is the quiet shift in how this is being handled: the question is moving from whether you used a tool to whether you can show what you asked it and what came back.
The supervisor watched it happen
The supervision relationship is the part most guides treat as pastoral and most assessment forms treat as evidence. NHH's criteria are unusually explicit about it, but the underlying idea is common across European programmes: alongside the quality of the finished research, someone judges how far you generated the question, ran the analysis and reached the conclusions yourself, and they judge it from having been there.
That is why "will they be able to tell" is the wrong worry at thesis scale. A supervisor is not comparing your prose to a corpus. They are comparing this month's chapter to last month's meeting. A draft that arrives more confident than the conversation that preceded it, an argument that appeared without the wrong turns, a literature review whose framing does not match the papers you were excited about in March: none of that is detection, and all of it is noticed. The same applies in reverse, incidentally, and in your favour. A supervisor who has watched you struggle with a method for six weeks is not going to believe you outsourced it.
There is a practical consequence. If you are using AI substantially, say so to your supervisor early, in the meeting rather than in the declaration. At ETH the declaration's AI option is chosen in consultation with them anyway, so the conversation is going to happen. Having it in month two is a conversation about method. Having it in the week of submission is a conversation about disclosure.
The room at the end
The defence is where the argument stops being a document. Formats vary widely, from a 30 to 45 minute bachelor colloquium to a doctoral defence that runs for hours in front of a committee, and the weight varies too, but the mechanism is constant: someone who has read the thesis asks you about it, and follows up.
What that room tests is not recall. It is whether the reasoning is reconstructible. An examiner asks why you chose this method over the obvious alternative, what would have happened if the sample had been different, which of the studies in chapter two you actually disagree with. Those questions have no correct answer to memorise, because they are probing the decisions behind the text rather than the text. A chapter you directed and reviewed is answerable, even if you wrote none of its sentences yourself. A chapter you accepted is not, however well it reads, and the gap shows up inside the first two follow-ups rather than in the first question.
This is also the honest reason the answer differs from the one for a research paper. Both documents can carry a fabricated citation. Only one of them puts you in a chair opposite a person who knows the literature and has time to ask.
What a tool is genuinely for here
None of this makes AI useless on a thesis. It narrows what it is for, and the narrow version is worth a lot, because it lands on the parts that are slow rather than the parts that are assessed.
It is good at finding and organising literature, which is where the weeks go, and at helping you see where a set of papers agree and where they do not. The full method is in how to write a literature review with AI. It is good at writing up decisions you have already made, which is most of a methodology chapter once the work is done, and how to write a methodology lists what has to be in one. It is good at catching gaps: a claim with no citation, a result the discussion never returns to, a chapter that promises something it does not deliver. And it is good at the mechanics, keeping a long bibliography consistent and a long document coherent.
What it must not touch is short and worth stating flatly. It cannot run your study or produce your data. It cannot decide what you are claiming. It cannot be the reason a source is in your bibliography, which is the failure the research paper page covers with the audit numbers, and which matters more here only because a thesis carries more references than you can re-check in submission week. And it cannot sit in the colloquium. For the structure of the thesis itself, from topic through chapters to the defence, how to write a bachelor thesis is the method guide, and how to write a research proposal covers the document that gets you approved to start.
Keep the log from week one
Here is the thing we would actually change about how students handle this, and it is not "use it less". Start the disclosure log on the first day you use a tool, not in the week you submit.
The reason is the shape of the Passau row above. A disclosure that asks which stage, which system, verbatim or paraphrased, and a link to the conversation, is easy to fill in on the day and close to impossible to reconstruct eight months later. What happens instead is that an honest student, facing a form in submission week, writes down what they can remember and signs it. That is how a person with nothing to hide ends up signing something inaccurate, and the declaration is the one document in this whole process where accuracy is a legal question rather than an academic one. Keep a running note: date, chapter, what you asked, what you used. It takes seconds, and it also gives you the second benefit, which is that a use you feel awkward writing down is the one to reconsider. That instinct is a better guide than any policy document, and it only works if you are writing things down while the instinct is still live.
The same question about shorter documents
The three questions look alike in a search box and come apart completely in practice.
An essay is short, usually argued from material you were handed, and marked on argument. It is the case where machine drafting comes closest to working, and the published marking bands show how close and where it stops, in how far a drafted essay gets.
A research paper stands on whether its sources are real and whether they say what you claimed. That is measurable, it has been measured, and can AI write my research paper collects the audits, including the number that survives an existence check.
A thesis is this page: examined, supervised, signed for. Where your institution draws its own line between support and misconduct is a separate question, covered in whether using AI to write essays is cheating, and if you are still choosing a tool, our round-up of AI tools for academic writing ranks them by whether they retrieve anything before writing.
So, can AI write your thesis?
It can write sentences that would pass for yours, and that is not the question anyone is asking. The thesis is examined by a person who watched it happen, defended by you in a room, and submitted under a declaration that you and your supervisor agree on the wording of. Every one of those is a check on the writer rather than the writing, and a model cannot stand in for the writer at any of them.
So use it on the slow parts and keep the decisions. That is what CiteOwl is built around: it searches real literature and reads what it finds, links every claim to a source it actually read with the supporting quote shown, keeps a version history across a project that runs for months, and shows each edit as a plain diff you accept or reject. What comes out is a thesis you have read line by line, with a bibliography that points at real papers, which is the minimum the room at the end is going to ask for.