A machine produces a clear, well-structured, sometimes brilliant answer. It cites a concept, unfolds an explanation and shows no sign of doubt. We then have a very human reflex: because the discourse resembles knowledge, we grant it the status of knowledge. The problem is not only that generative artificial intelligence can be wrong. The problem is that the way it is wrong fits perfectly with our desire to believe.

Two old scientific blunders allow us to view this reflex differently. They involve neither amateurs nor fraudsters caught red-handed, but two respected scientists, Yves Rocard and Prosper René Blondlot. Each worked on a phenomenon that turned out not to exist. Each had an experimental framework he considered rigorous. And each reminds us that a scientific method can lose its power when it mainly serves to confirm what the experimenter already hopes to observe.

Scientists are not immune to expectation

Beginning in 1957, Yves Rocard sought a physical explanation for the alleged powers of dowsers. His career hardly predisposed him to easy credulity: a professor at the École normale supérieure, director of its physics laboratory and scientific adviser to the CEA, he had contributed to several established fields of research. Yet he proposed that some people might perceive local variations in Earth’s magnetic field linked to the presence of groundwater.

To test this hypothesis, Rocard devised, among other things, an apparatus in which an electric wire produced a controlled magnetic field. A subject holding a pendulum stood nearby. Reversing the current was meant to determine whether the pendulum’s direction of rotation also changed. On paper, the experiment appeared to be organized around an identifiable physical variable. But critical analysis of his method and results revealed biases significant enough to make this episode a scientific blunder rather than fraud.

The case of Prosper René Blondlot is older. In 1903, this professor at the University of Nancy and member of the Academy of Sciences announced the existence of rays he called “N”. The context made the possibility seem plausible: X-rays had been discovered in 1895, radioactivity in 1896, and in 1903 several figures associated with these discoveries were honored with the Nobel Prize in Physics. At a time when new forms of radiation were transforming science, the appearance of another phenomenon seemed less absurd than it does to us today.

Blondlot claimed that N-rays, produced by a lamp fitted with an incandescent ceramic rod, made a dimly lit screen appear slightly brighter. Everything therefore rested on an almost imperceptible difference near the threshold of visual perception, one that only a trained observer would supposedly be able to discern. He then believed he could study their propagation, refraction and diffraction using the methods of classical optics.

These two cases share a decisive feature: the observation is not clearly imposed by the phenomenon itself. It depends on a weak, ambiguous, interpretable signal. The pendulum may seem to respond. The screen may seem to brighten. From that point on, the researcher’s expectation no longer remains outside the experiment. It enters into the result.

The link with AI is not technical; it is psychological

This comparison belongs to neither of the two cases: it must be presented for what it is, an analysis. In my view, N-rays and AI hallucinations belong to radically different eras and experimental setups, but they activate the same weakness in the observer. We find it harder to evaluate a claim when its form already matches what we expect.

In Blondlot’s case, the race to discover new forms of radiation made the existence of an additional phenomenon plausible. In Rocard’s case, a magnetic hypothesis gave a physical appearance to a practice surrounded by beliefs. With AI, the situation shifts: we no longer observe a dim screen or a pendulum, but coherent, immediate and personalized text. The ambiguous signal becomes a convincing sentence.

Three shared mechanisms can be identified, provided they are treated as an interpretive framework rather than proven findings:

  • Expectation guides interpretation. When we hope for an answer, an explanation or confirmation, we more readily notice what supports it.
  • Apparent expertise is reassuring. A prestigious scientist, a sophisticated protocol or technically fluent prose can lend weight to a claim even before it has been verified.
  • Repetition creates normality. When several people accept the same framework, doubt seems less legitimate. Collective agreement can then look like proof.

Excessive trust in AI therefore does not come only from a lack of technical knowledge. It also comes from a more ordinary phenomenon: we readily confuse coherence with truth. A well-written answer often seems better grounded than a hesitant one, even though stylistic confidence guarantees nothing.

AI can produce novelty without producing truth

A third, contemporary debate concerns the creativity of generative AI. These systems can form original or unexpected combinations from available data, while some observers argue that they merely recombine learned states. Jean-Claude Heudin, for his part, explores another path with genetic algorithms based on recombination, mutation and selection, which he describes as a form of algorithmic creativity.

This debate helps clarify the problem. A production can be original, surprising and even useful without being factually correct. Novelty is not proof. Neither is elegance. An AI can propose an association that no one had formulated in that way and, in the same movement, introduce information that does not exist. This mixture is precisely what makes it persuasive: truth, plausibility and invention can all speak with the same voice.

We would therefore be making a mistake if we asked only: “Does this answer seem intelligent?” The useful question is harsher: “What in this answer would withstand independent verification?” The old scientific blunders do not ask us to despise intuition, expertise or innovation. They ask us not to give them the final word.

What N-rays should change in the way we use AI

The first lesson is to separate quality of expression from quality of evidence. A clear answer can be false. A cautious formulation can be more reliable than a spectacular account. When faced with an important claim, we should look for what could refute it, not only what makes it appealing.

The second lesson is to move verification outside the system that produces the claim. Rocard interpreted his own apparatus. Blondlot relied on a subtle perception reserved for trained observers. In the same way, asking an AI to confirm its first answer is not always a genuine verification: it may simply reconstruct a justification compatible with what it has just asserted. Verification must return to a distinct source, an accessible document or a reproducible measurement.

The third lesson concerns the collective. A shared error remains an error. When several users relay the same generated answer, its origin can disappear and repetition can give it artificial authority. The number of repetitions does not replace the strength of the original fact.

The danger is not that AI forces us to believe. It is that it gives us exactly the answer we were ready to accept.

Science is distinguished not by the absence of errors, but by its ability to expose them and revise its conclusions. The N-ray affair is indeed presented as a famous collective illusion while also illustrating this capacity for self-correction. This may be the best model to apply to generative AI: use it widely, but build around it an active culture of contradiction.

Rocard and Blondlot are not looking at us from a naïve past. They are holding up a mirror. Their mistake was not to have imagined a bold hypothesis. It was to let the hypothesis teach their eyes what they were supposed to see. Today, our eyes read brighter screens and faster answers. Our responsibility remains the same: never let conviction run too far ahead of evidence.