In biology laboratories, weather forecasting centres, pharmacology teams and mathematics departments, AI no longer appears merely as a machine for producing text. It is becoming an instrument of research: capable of exploring hypotheses, detecting invisible regularities and accelerating procedures that used to take a very long time.
AlphaFold has transformed the prediction of protein structures. Machine learning models can produce atmospheric forecasts in seconds from decades of data. In pharmacology, AI can virtually test thousands of molecular interactions, reposition an existing drug, and select promising lines of inquiry. In medical imaging, it can detect a barely visible fracture, a nodule, a cardiac anomaly, and draw the radiologist's attention to what deserves to be reviewed.
This is not enthusiasm. These are documented facts, with their own limits.
But the real question is not what AI does in these domains. It is why it can do it without turning into an epistemic disaster, or at least not yet.
The answer can be stated in one formula: over several centuries, science has built an infrastructure of organised doubt.
The reproducibility of experiments. Controversy among peers. The demand for proof rather than plausibility. The publication of negative results, at least in principle. Validation before dissemination. The right, even the duty, to contradict.
This system is not perfect. It has blind spots, publication biases and institutional pressures. But it exists. It precedes the tools. And that is precisely what allows AI to enter science without ruling over it.
Mathematics as a test
Mathematics offers the most striking example. Researchers initially looked at large language models with scepticism: how could a statistical machine contribute to a discipline grounded in proof? Then some results displaced those certainties. AI systems have helped explore conjectures and sketch possible lines of proof in a few hours where weeks might once have been necessary.
But the discipline has held firm on the essential point: the machine can produce a plausible form, but plausibility is not truth. It opens a path; the path still has to be walked by an intelligence capable of justifying, connecting, checking and proving.
The question mathematics has had to reformulate under the pressure of AI is exactly the right one: what is a proof? Who validates? At what point does a proposition become knowledge? These are questions science already carried. AI has made them more urgent, not less necessary.
Autonomous agents as a warning
The case of autonomous agents says the opposite in negative. This is no longer simply a matter of questioning a tool, but of entrusting a system with several chained tasks: reading papers, formulating a hypothesis, writing code, testing, comparing and producing a report. In some domains, these agents can reproduce in a few dozen minutes procedures that would previously have required months.
The danger is exactly proportional to the promise. An AI can fabricate data, hallucinate a source, ignore negative results, reproduce the biases of the databases on which it works. It can flood the scientific community with results that look plausible but are false.
This is not only a technical problem. It is a scientific problem. If the production of results becomes faster than the collective capacity to verify them, science risks losing what makes it strong: not the accumulation of answers, but the patient organisation of doubt, the necessary time of verification.
Delegation under conditions
I would call this delegation under conditions. AI can receive part of the exploratory work. It must not receive the right to conclude. It can help search, not decide alone what counts as proof.
The remaining question is not whether AI can contribute to research. It already can, in a significant way. The question is to understand under what conditions.
The answer suggested by science is uncomfortable for other domains: the value of a tool always depends on a pre-existing culture capable of containing it. Science can integrate AI because it has already built the infrastructure of doubt that resists it. Where that infrastructure does not exist, or has not yet been built, the tool encounters nothing that resists it. It does not liberate. It colonises.
This is exactly what school does not yet have. The prior culture that would allow it to receive these tools without being dispossessed by them. Such a culture cannot be improvised with a user charter or a one-hour awareness session. It is built slowly, in the same spirit as science: learning to doubt, to verify, to retrace the path, and not to confuse what is plausible with what is true.
AI in science is therefore, in its own way, a model of salutary doubt, and perhaps a model to follow.
This text extends the reflections on the dispossession of the educational subject, the clash of temporalities, dialogic exploration and responsible digital autonomy.