Institutional policy on AI is being written while researchers are already using the tools. The result is genuine uncertainty about what is permitted — and a temptation to resolve it by not asking.
A more durable approach: reason from the principles that were already in place, because they have not changed.
The principle that has not changed
Academic integrity has always rested on two commitments: the intellectual work is yours, and your claims about sources are true.
Neither commitment is about which tools you used. A calculator does not compromise authorship; a research assistant does not either, provided their contribution is acknowledged where convention requires. AI tools are assessed the same way — by whether your authorship survives and your sourcing remains honest.
Where AI use is generally uncontroversial
- Language polishing of text you wrote — grammar, register, concision. Long-accepted from human editors, particularly for multilingual researchers.
- Reformatting — converting citation styles, restructuring tables, adjusting to journal templates.
- Explaining concepts to you — the same role as a textbook or a colleague, where the output informs your understanding rather than appearing in your text.
- Code assistance for analysis you designed and can verify line by line.
Where it becomes a problem
- Generating substantive argument or interpretation that appears as your scholarly reasoning. This is the authorship line, and it is the one that matters.
- Producing citations. Language models fabricate references that look entirely plausible — correct-seeming authors, journals, volumes, and DOIs for papers that do not exist. Fabricated citations are a serious integrity finding regardless of intent, and "the tool produced it" is not a defence.
- Summarising sources you have not read, then citing them. You are vouching for a characterisation you cannot verify.
- Analysing data in ways you cannot explain. If you cannot reconstruct and defend the procedure, you cannot defend the finding.
The verification rule
Anything an AI tool produces that ends up in your work must be verifiable by you, from the original source, before it goes in.
Every citation checked against the actual paper. Every factual claim traced. Every analytical step reproducible. This is not an AI-specific standard — it is the standard that already applied to notes, to co-authors, and to research assistants.
Disclosure
Norms vary and are moving, so check three places: your institution’s academic integrity policy, your supervisor’s expectations, and the target journal’s author guidelines. Many publishers now require a statement on AI use, and most prohibit listing an AI system as an author — reasonably, since authorship entails accountability.
Where disclosure is required, be specific and unembarrassed: "Language-editing assistance was provided by [tool] for grammar and clarity; all content, analysis, and citations are the author’s own and were verified against original sources."
Vague disclosure attracts more scrutiny than precise disclosure. So does the absence of any.
Keep a use log
A short running note of which tools you used, for what, and when. It costs a line per session.
If a question ever arises — and in the current environment, questions arise — the difference between a researcher who can immediately describe their practice and one reconstructing it from memory is the difference between a short conversation and a long one.
We do this work every week — for MPhil and PhD candidates, faculty, and whole departments.
Start a conversation
