AI Tools PhD Students Are Actually Allowed to Use (2026 Guide)

The one distinction most policies now use

AI tools allowed for PhD students is a live question with a moving answer. There is no single national or international policy. Universities are decentralised on AI use: departments and instructors set the rules within institutional frameworks. What is stable across current policies is one specific distinction.

The distinction is between tools that do the work and tools that check the work. Tools that generate text, produce citations, or draft arguments do the work. Tools that verify existing text, check citations against databases, or grade your own answers against your own thesis check the work. Most current university policies (Harvard, MIT, Stanford, Oxford, Cambridge) treat the second category as low risk and the first as requiring disclosure or, in some cases, prohibited (Thesify AI policy summary, 2026).

Category 1: Tools that check the work (usually allowed)

Tools that check work you have already produced sit on the lower-risk side of the line. Check your university's policy: these tools may be allowed without special disclosure, as extensions of proofreading and reference management.

Grammar and spelling checkers (Grammarly, ProWritingAid, LanguageTool). These check the writing you have already produced. They do not generate new arguments or new sources. Grammarly Pro adds a plagiarism check that scans web pages and ProQuest databases (Grammarly, 2026).

Reference verification tools (Thesisroom's Citation Check, Scholarcy for summarising, third-party DOI resolvers). These check that citations in your bibliography exist, match their metadata, and have not been retracted. Verification tools do not generate references; they verify them.

Anchored question generators for rehearsal (Thesisroom's defense practice). These read your existing thesis and generate examiner-grade questions from your text. The tool does not write your answers; it produces questions from your own writing for you to answer.

Structural review tools (Thesisroom's chapter audit). These read a chapter you have written and return margin comments on structural weaknesses. The tool does not rewrite the chapter; it comments on it.

Category 2: Tools that do part of the work (disclosure required in most policies)

Tools that generate summaries, translations, or drafts sit in the middle. Most current policies allow them for specific limited uses (paraphrasing your own writing, translating between languages, summarising papers you have read) with disclosure. Some prohibit them for direct inclusion in submitted work.

General-purpose language models (ChatGPT, Claude, Gemini) used for brainstorming, editing your own drafts, or asking questions about concepts. Cambridge's policy: unacknowledged use in assessments constitutes misconduct (Cambridge University, 2026). Many other institutions require declaration of AI use during specific assessment tasks.

Translation tools (DeepL, Google Translate) used for reading sources in another language or drafting in a second language. Widely used and generally accepted with disclosure.

Paper summarisation tools (Elicit, Scite, Consensus, SciSpace) used to survey a literature quickly. Useful for finding candidate sources but must be verified before citing.

Category 3: Tools that do the work (usually prohibited)

Tools whose output is intended to be submitted as your work sit clearly on the prohibited side of the line. This includes:

• Full essay or chapter generation from a prompt
• Auto-generation of citations that then enter your reference list without verification
• Full research reports drafted by an AI from a brief

University policies can treat this as academic misconduct; check yours. The rise in fabricated citations tracked in the Topaz and colleagues (2026) Lancet audit (one in every 277 PubMed papers in early 2026 contains a fabricated reference, up from one in 2,828 in 2023) has made this category particularly high risk.

Why the distinction matters for the viva

The distinction between does-the-work and checks-the-work is not just a policy line. It is a viva line. Examiners now spot-check bibliographies against Crossref, and they read for chapters that show signs of generated text. A candidate whose bibliography contains fabricated citations from an unchecked AI draft is exposed at exactly the moment the examiner tries to verify a source.

The safer harbour for AI use in academic writing is the check side of the line. Verifying existing citations, checking your own text, and rehearsing against questions drawn from your own writing all leave your thinking and your writing untouched. This is what the Thesisroom method means when it treats verification as a whole-degree seat.

The disclosure question

Some universities now require candidates to declare AI use in submitted work. Cambridge, Princeton, and Oxford have moved this direction. The declaration usually covers what tools were used, for what purposes, and at what stage of the work. Grammarly-style proofreading is often not required to be declared; substantive drafting or research assistance usually is.

If you are unsure whether a tool requires declaration in your institution, err on the side of disclosure. Declaring low-risk use is not misconduct. Failing to declare higher-risk use can be.

What Thesisroom does with this

Thesisroom sits clearly on the checks-the-work side. Citation Check verifies existing citations. Chapter Audit reviews existing chapters. Defense Practice generates questions from your existing thesis. None of the modules generates text for inclusion in your work.

The Thesisroom integrity policy sets this position formally: the writing stays yours, the thinking stays yours, and the tool never produces submittable content.

Frequently asked questions

Is Grammarly allowed for a thesis?

It depends on your university's policy; check it before you rely on the answer. Grammarly checks existing prose rather than generating new text, which puts it on the lower-risk side of most policies. Some institutions require declaration of AI-tool use; check your specific institution's policy.

Can I use ChatGPT for my thesis?

Depends on how and what for. Using it to brainstorm or to explain a concept in the reading stage is usually accepted. Using it to draft passages for direct inclusion in the thesis, or to generate citations, is usually prohibited or requires explicit disclosure. Cambridge treats unacknowledged use as misconduct.

Do I have to declare AI use in my thesis?

It depends on your university's policy; check it before you rely on the answer. Cambridge, Princeton, and Oxford have moved toward declaration requirements. Rules vary by institution and by assessment. If unsure, disclose; declaring low-risk use is not misconduct, while failing to declare higher-risk use can be.

What happens if I use an unallowed AI tool?

Consequences range from being asked to redo the work to formal academic misconduct proceedings. Most institutions publish specific procedures. The most common triggers for detection are fabricated citations surfacing in the reference list or stylistic markers of generated text in the chapters (Turnitin AI detection, GPTZero, and similar tools flag these with varying accuracy).

Is using AI for translation allowed?

Generally yes, especially for reading sources in languages you do not fluently read or for drafting in a second language. DeepL and Google Translate are widely used and accepted. Disclosure of translation use for substantive translated passages is often expected.

How does Thesisroom fit university AI policies?

Thesisroom sits on the checks-the-work side of the line. It verifies bibliographies, reviews chapters you have written, and generates questions from your existing text. It does not draft passages, generate citations, or produce submittable content. This positioning is designed to stay clearly inside current policy at most institutions.

One takeaway before you close this tab

The AI-tool distinction that most policies now use is between tools that do the work and tools that check the work. Verification, review, and anchored rehearsal all fall on the check side. Thesisroom's method is built for that side, so your writing and your thinking stay yours.

References

Cambridge University. (2026). Academic misconduct and AI policy.

Grammarly. (2026). Product features and pricing tiers.

Thesify. (2026). Generative AI policies at top universities.

Topaz, M., et al. (2026). Analysis of fabricated references in PubMed-indexed literature. The Lancet.