Artificial intelligence models placed under strong pressure make a decision related to scientific integrity incorrectly in about one case out of three.
Artificial intelligence tools can now help researchers analyze data, write code or examine results. But a question arises as these systems become more involved in scientific work: do they react correctly when a request conflicts with research rules?
AI illustration
To measure this, researchers created IntegrityBench. This benchmark confronts models with 36 pairs of tasks covering three domains and four stages of research. The situations notably involve recognizing incorrect behavior, choosing an appropriate action and making decisions based on provided documents.
The team tested 18 variants of recent models. Each situation could involve five levels of pressure, ranging from an unconstraining context to strong pressure. This pressure could be explicit, for example when an instruction directly encouraged incorrect conduct. It could also be integrated more discreetly into the context.
At the highest pressure level, the models fail on about one critical decision in three. Greater size or more reasoning does not systematically reduce these errors. This result shows that improving a model's general performance does not automatically guarantee better decisions when faced with research rules.
The type of pressure also changes the nature of the errors. A request explicitly directed toward a bad practice more often leads the model to accept it. Conversely, indirect wording can cause too many refusals. In this second case, the AI sometimes blocks a scientific task that is nevertheless legitimate.
In other words, two opposing risks emerge. An overly permissive system can facilitate bad scientific practice. An overly cautious system can prevent normal work and reduce its usefulness. The authors also find that a good decision does not always depend on correctly classifying the request.
These results come from a controlled benchmark. The study, deposited on arXiv on June 3, 2026, mainly provides a method for comparing systems and precisely identifying their weaknesses.
The next step is to test more situations, particularly with real scientific documents and contexts close to everyday work. This will make it possible to verify whether the same errors appear when AI intervenes in longer research workflows.