AI and Research Integrity
This note sets out the research-integrity guidelines that govern AI use, and how I approach them in the classes I teach. It is the baseline I work from.
The core rules of responsible AI use
The major bodies (the EU's ERA guidelines, ALLEA, COPE and ICMJE, the leading journals Nature and Science, and the main funders ERC, SNSF, DFG and UKRI) converge on a short list of essentials:
- AI is never an author; the work stays yours. You remain fully responsible for everything in your work, whatever helped you produce it. AI supports your thinking, it does not stand in for it.
- Keep confidential material out: no unpublished work by others, no personal data, no peer-review documents.
- Disclose and cite your AI use: which tool, what for, and how much.
- Stay traceable: keep a record of the model, its version, and the prompts that mattered, so an AI-assisted method can be followed like any other.
Compliance Through Competence
My own view is that compliance comes through competence. Not from just knowing the rules. Advanced AI skills, adapted to scientific research, are what make compliance not only proper but frictionless: the rules get respected without costing the researcher much time or effort, which leaves them free to focus on the work.
The idea is not to use AI without getting oneself into trouble. It is to be able to expertly produce great research with AI, and as a consequence, to almost effortlessly comply with guidelines of the highest standard. A researcher who is genuinely skilled with AI can meet these rules by default. This happens through two means.
The first is automatic recording and flagging. A researcher who can set up their own agentic AI system, that is, their personalised AI assistant, can also set it up so that:
- Every interaction with the LLM is logged (tool, use case, time, prompts, model, version, and so on), which renders the whole of the research done with the LLM traceable, queryable, and describable, down to the individual prompt or session.
- Confidential material entered into the LLM is flagged, with a fairly high success rate; an automatic stop can even be built in to mitigate possible damages.
This greatly facilitates disclosure, making it highly accurate and partly automatic, needing only the researcher's review.
The second means is knowledge and judgment. Most of my teaching is about exactly that: training the judgment to use AI critically and expertly, while staying fully in control. It is about knowing when to trust AI outputs, when to distrust them, and how to push back and course-correct when they stray. It is also about having the technical knowledge to be able to self-design the appropriate systems for one's own research, including those concerning research integrity and data safety. With that in hand, a researcher can both meet the guidelines and see the quality, speed, and enjoyment of their work improve.