Can I use AI in a PhD thesis?

There is no single rule, and anyone who tells you there is has read one university's policy. Doctoral regulations sit at several levels at once: the institution, the graduate school, the faculty or department, the examination regulations, and your supervisory committee. All of them can constrain you, and the strictest one wins.

The direction over the last two years has been towards permission with mandatory declaration. Several universities now require a statement inside the thesis itself, in some cases immediately after the abstract, naming the tool and version, the provider, where and how it was used, and confirming compliance with the local rules. Some departments require prior committee approval and treat undeclared use as unauthorized assistance.

A minority go further and prohibit specific uses. Redrafting a thesis written in another language is prohibited at least one major university, and some departments prohibit generated written content entirely unless approved.

Journal policy is moving the same way, and your examiners read it. A 2025 comparative review of editorial policies found that ICMJE, WAME, COPE, the NEJM, the Lancet, Nature, the BMJ, Cell, and the Journal of Korean Medical Science all permit some AI assistance provided it is disclosed, while Science prohibits certain uses outright and requires the tool, the version, and the prompts to be reported (Yoo, 2025, p. 6). The common thread is accountability rather than prohibition. ICMJE's position is that a generative system cannot be an author because it cannot make creative decisions, take responsibility for the work, or declare conflicts of interest, so the purpose and extent of any assistance has to be stated by the human author (Yoo, 2025, p. 3).

Whatever your local rule, two habits protect you. Keep a running log of what you used, when, and for what, because a declaration written from records is accurate and one written from memory is not. And keep drafts with version history, since that is the strongest evidence available that the writing is yours.

Verify every output before it enters the thesis. Generative systems produce convincing but fabricated content, and fake references and citations are the most common form, a problem that persists even as factual accuracy improves (Helmy et al., 2025, pp. 2-3). Fabricated citations, hallucinated data, and paraphrased plagiarism are the misconduct scenarios the policy literature keeps naming (Yoo, 2025, p. 11). Check each citation, quotation, number, and interpretation against the original source, and do not paste confidential or unpublished research data into a system your policy has not cleared.

What do universities typically permit?

Assistive use of language and structure on text you wrote, and increasingly some analytical or literature-handling use with disclosure. What is commonly prohibited is presenting generated text as your own writing, and what is commonly required is a declaration plus records of what you used.

The trend in 2026 policies is towards specificity. Earlier statements said "use responsibly", which nobody could apply. Newer ones name the categories, ask for tool and version, and specify the location of the declaration, which is easier to comply with.

If your regulations are vague about what counts as substantive, the publishing side offers a usable line. Resnik and Hosseini's scheme, as reported in the same review, sorts disclosure into mandatory, optional, and unnecessary. Mandatory covers intentional and substantial use: forming hypotheses, writing portions of the manuscript, collecting or analysing data, and generating figures or tables. Limited linguistic, structural, or routine technical help sits in the optional or unnecessary categories (Yoo, 2025, p. 7). Applied to a thesis, anything that shapes an argument, interprets evidence, or produces a table belongs in the declaration, while spelling and grammar correction on your own sentences usually does not.

Write the declaration so someone could reconstruct what happened. The operational recommendation is to name where and how the system contributed across data preprocessing, hypothesis generation, analysis, and manuscript preparation, and to identify the platform, the version, and the nature of the involvement (Helmy et al., 2025, p. 9). One qualification worth knowing before a viva: detector scores are weak evidence. They produce false positives, they are biased against writers whose first language is not English, and paraphrasing defeats them, so they belong as supplementary information rather than as the basis for an academic decision (Yoo, 2025, p. 8).

Read your own regulations rather than a summary of them, including the examination regulations, which are the ones that actually govern what happens at submission.

What should I tell my supervisor?

Ask before you use anything substantive, in writing, describing the specific use rather than the tool. "May I use a tool to screen abstracts against my inclusion criteria" gets a clear answer. "Can I use AI" gets a nervous one, because it could mean anything.

Keep the reply. A supervisor's written approval of a specific use is the single most useful thing you can have if the question comes up at examination, and it costs one email to obtain.

If your supervisor is uncertain, offer to write a short paragraph describing exactly what you propose to do and how you would verify the output. That converts an anxious general question into a concrete one they can say yes or no to, and it is the same paragraph your declaration will eventually need.