Institutional policies on AI have converged faster than most people expected, and they converge on one principle: you are responsible for everything in your submission, and undisclosed use of assistance is misconduct. What differs between institutions is the disclosure requirement and the list of permitted tasks - and those differences are large enough that no general guide substitutes for reading your own regulations.
This post covers where the lines usually fall, how to write a disclosure statement, and why the detector question is the wrong one to organise your work around.
Research assistance you can actually cite
ThesisAI works from real papers indexed in academic databases and verifies each inline citation against the source, so what you submit rests on evidence you can check.
書き始めるStart With Your Own Regulations
Policies fall into roughly four types, and you need to know which one governs your programme before anything else:
- Prohibited. No generative AI in assessed work at all. Still the position for some professional and clinical programmes.
- Permitted with disclosure. The most common position. Use is allowed for defined tasks and must be declared, often in a specified format.
- Permitted for defined tasks only. Language editing and idea generation allowed; drafting of assessed prose not allowed.
- Set by the supervisor or module. Devolved to individual courses, which means the answer can differ between two chapters of the same degree.
Two practical points. First, policies have been revised repeatedly - the version you read at enrolment may not be the one in force at submission. Second, if your regulations are ambiguous, ask your supervisor in writing and keep the reply. An email confirming what was agreed is the only protection that actually holds up in an academic integrity hearing.
Where the Line Usually Falls
| Task | Typical position | Why |
|---|---|---|
| Brainstorming topics and angles | Generally accepted | The output is a starting point you then evaluate and develop |
| Explaining a concept or method to you | Generally accepted | Equivalent to a textbook or a tutor, though the accuracy is on you |
| Generating search terms and synonyms | Generally accepted | You still run the searches and read the results |
| Grammar and language editing | Usually accepted, often needing disclosure | Equivalent to a permitted proofreader, with the same scope limits |
| Summarising papers you have read | Varies; disclosure usually required | Useful, but summaries drift from sources, so verification is mandatory |
| Drafting sections you then rewrite | Varies sharply by institution | Close to the boundary of authorship; check explicitly |
| Submitting generated prose as your own | Misconduct nearly everywhere | Undisclosed substitution of authorship |
| Generating citations or data | Serious misconduct | Fabrication, whether or not it was intentional |
The Fabricated Citation Problem
Language models produce plausible references that do not exist: real author names, a real-sounding journal, a DOI that resolves to nothing or to something else entirely. This is now one of the most common ways students end up in an integrity process, and intent is not a defence - a citation to a paper that does not exist is fabrication in the record regardless of how it got there.
The rule is simple and non-negotiable: open every source you cite. Not the summary of it, not the abstract in a search result - the paper. Confirm the authors, the year, the journal, and that it says what your sentence claims it says. If a reference cannot be located in a library catalogue or through its DOI, remove it.
This is also the practical case for tools that retrieve from indexed academic databases rather than generating references from a model's memory: the citation exists because it was found, not because it was predicted.
Writing a Disclosure Statement
Where disclosure is required, be specific. A vague statement is worse than none, because it looks like an attempt to satisfy a rule without meeting it.
Weak: "AI tools were used in the preparation of this thesis."
Strong: "A large language model (Claude, Anthropic) was used for two purposes in the preparation of this thesis: generating alternative search terms during the literature search in Chapter 2, and grammar and clarity editing of Chapters 4 and 5. No text generated by the model appears in the submitted document without substantial rewriting by the author, and no source was cited without being obtained and read in full. All analysis, interpretation and conclusions are the author's own."
Names the tool, states the tasks, bounds what it did not do, and takes responsibility for the intellectual content.
Some institutions require this on the declaration page, others in a methods appendix, others in the acknowledgements. Follow the specified location. Where a log of prompts is required, keep one from the start - reconstructing it afterwards is not possible.
Why AI Detectors Are the Wrong Thing to Optimise For
Detection tools report a probability, not a fact, and their false positive rates are high enough that multiple universities have restricted or withdrawn their use in integrity decisions. Non-native English writers and heavily edited academic prose are flagged disproportionately, which is a fairness problem as much as a technical one.
Two consequences follow. If you are accused on the basis of a detector score alone, you are entitled to ask what the tool's error rate is and to present your drafting history - version history, notes, and dated files are the strongest evidence a student can hold, and they cost nothing to keep.
And in the other direction: writing in order to defeat a detector is a bad objective. It optimises for a score rather than for quality, it does not resolve the underlying question of whose work it is, and it produces worse prose. If the work is genuinely yours and your use is disclosed, the detector's opinion is not the standard you are being held to.
The Transparency Direction of Travel
Disclosure norms are tightening rather than loosening. Most major publishers now require authors to declare AI use in submitted manuscripts and prohibit listing an AI system as an author, on the grounds that authorship entails accountability. In the EU, transparency obligations for generative AI systems under the AI Act begin to apply from August 2026, which will make provenance information more available rather than less.
The practical implication for a thesis being written now is that disclosure is the safe default even where it is not yet mandatory, and that any workflow built on the assumption that use will remain invisible is a poor bet.
A Workflow That Stays on the Right Side
- Read your policy, and keep a copy of the version in force.
- Confirm ambiguities with your supervisor in writing.
- Use AI for orientation, search and language - not for evidence or judgment.
- Open and read every source before it enters your reference list.
- Keep drafts, version history and a prompt log as a matter of routine.
- Write the disclosure statement as you go, not from memory at submission.
- Apply one test to anything you keep: can you explain and defend it under questioning? If not, it does not belong in the thesis.
FAQs About AI and Academic Integrity
Is using AI to improve my grammar plagiarism?
Generally no - it is usually treated like using a permitted proofreader, with the same scope limits: language and presentation, not content or argument. Many institutions still require it to be declared.
Do I have to disclose if my institution has no policy?
Disclosing is the safer choice. An absent policy is usually a gap rather than a permission, and policies are frequently applied retrospectively to work under review.
Can I cite ChatGPT or Claude as a source?
APA and MLA both publish formats for it, and it is rarely appropriate in a thesis, because chatbot output is not a retrievable source and carries no evidential weight. Cite the underlying literature instead. Some institutions prohibit citing chatbot output outright.
What happens if I am wrongly accused based on a detector score?
Ask for the specific evidence beyond the score, and provide your drafting history: version history in your word processor, dated backups, notes, and reading records. This is why keeping that trail from the start matters.
Is it acceptable to use AI for a literature search?
Searching and screening is one of the most widely accepted uses, provided you obtain and read the papers yourself and verify every citation. What is not acceptable is citing what a model said about a paper you never opened.
The durable principle underneath all of this is unchanged by the technology: a thesis is a claim that you did the thinking. Tools that help you find, read and organise evidence support that claim. Tools used to replace the thinking undermine it, whether or not anyone notices.