How to use AI for thesis writing (and how not to)

AI tools are now on most graduate students’ desks. This guide says neither "never touch them" nor "let them do everything". Its aim is to help you tell apart the uses that strengthen your thesis from the ones that will cause trouble at your defense.

Feedback or ghostwriting?

The most practical way to draw the line is to ask: where did the ideas, reasoning and sentences in the text come from? When you give a tool a chapter you wrote and ask "where is the weak link in my argument?", it acts as a reader and you still make the decisions. When you have it write a paragraph or a section and paste that into your thesis, you are no longer the author of that text. At most institutions this second kind of use conflicts with the originality statement you sign, and it makes the defense question "why did you write it this way?" much harder to answer.

There is a grey zone in between: language editing, translation, summarising. Whether these are acceptable varies by institution; we come back to this below.

The biggest risk: sources that do not exist

General-purpose chatbots produce fluent but unverified text. Ask them for references and they can invent publications that look real but do not exist. In an experiment by Linardon and colleagues in 2025, 19.9% of the citations generated by GPT-4o pointed to publications that do not exist. Because the authors, journal and year look plausible, these errors are hard to spot by eye.

The rule is simple: never cite a source you have not read. If a tool suggests a paper, find the publication first, open its DOI and see that the claimed finding is really there. How to verify your whole reference list before submission is covered in the reference check guide.

Your unpublished data and confidentiality

Before you upload raw data, interview transcripts, unpublished findings or your supervisor’s draft work to an AI tool, you need to know where it goes. Check the tool’s terms for whether inputs are stored and whether they are used to train models. If participant data is involved, the commitments you made in your ethics approval about how the data will be stored and who it will be shared with are binding too; if in doubt, ask your supervisor and ethics committee before uploading anything. Removing identifying details is a good habit, but it is not always enough.

The rules are not the same everywhere

Universities, graduate schools and journals have different AI policies, and those policies change often. Some allow language editing but forbid generated text, some require disclosure of any use, some leave the decision to the department. Giving a general rule here would be misleading: read your graduate school’s current guidance, check the author guidelines of the journal you are submitting to, and clarify anything unclear with your supervisor in writing.

Disclosure: what and how

If your institution requires disclosure, state which tool you used, for what purpose and at which stage of the thesis. Prefer concrete wording such as "for grammar checking" or "to generate search terms for the literature search" over a single line saying "AI was used". An AI tool is not listed as an author; you remain responsible for the content and accuracy of the text.

Sensible uses

AI is most useful when it helps you look at your own text from the outside. In the uses below, the decisions and the writing stay with you:

• Finding gaps: give it your literature review and ask "which question do these summaries leave unanswered?" You then verify the suggested gap by going back to the sources.
• Questioning your argument: ask "what objections would a committee member raise to this chapter?" Some of the objections will be off the mark; sorting out the valuable ones is your job.
• Checking consistency: ask whether the research question in the introduction is answered in the same form in the conclusion, and whether key terms keep the same meaning across chapters.
• Clarity test: give it a paragraph and ask for a one-sentence summary. If the tool cannot recover what you meant, a reader may not either.
• Defense rehearsal: have it list questions the committee might ask and practise your answers out loud. The thesis defense questions guide helps here.

In each of these, treat what the tool says as an opinion, not a verdict. The strengths and weaknesses of general chatbots are covered separately in getting thesis feedback from ChatGPT.

What you should never delegate

• Your research question and main claim. You are the one who will defend why the thesis exists.
• Data and analysis results. Even if a tool runs statistics for you, reproduce the result in your own software; never let it make up findings.
• Summarising and citing sources you have not read. You can only know what a paper says by reading it.
• Interpreting your findings. Judging how much the data can carry needs field knowledge, and tools tend to make claims stronger.
• Ethics and originality statements. You need to know the content of every statement you sign.

In short

Think of AI not as an author but as a patient reader who can be wrong. Verify everything it produces, protect your data, follow your institution’s rules and disclose your use if you used it.

tezkeys is also an AI-based reading tool, and it sits on the reader side of this line: it does not write or rewrite your text; it reads your thesis chapter by chapter, marks argument, method, findings-interpretation and literature problems on your own text, and compares your references with Crossref records. Because the report is AI-generated it can be wrong, and it does not replace your supervisor. See what it looks like in the sample report.