A machine gives an answer, recognises a shape or wins a game. The result catches the eye; the interesting question comes just afterwards: what has it actually managed to do, and under what conditions? This is the curiosity with which to approach “Cap sur l’IA” at the Air and Space Museum in Le Bourget.
Running from 15 October 2026 to 27 June 2027, the exhibition is aimed at visitors aged ten and over. The museum announces around thirty interactive exhibits arranged around five areas: knowledge, communication, creativity, perception and movement. There is also a section on military uses. This account is based on the programme published by the institution and does not claim to be a review of the exhibits in person.
A successful experience is not a universal capability
The value of trying something out is that it makes a question concrete. An interface that seems to “understand” may actually be classifying data, generating a likely sequence or applying a strategy within a limited framework. These operations can be useful without replicating the way a person thinks.
For a visitor, three questions help get beyond the surprise effect: what task was the system given? What result counts as success? What happens when the conditions change? You can keep them in mind when watching a demonstration, without assuming that every exhibit gives you access to all these parameters.
A hypothetical example helps illustrate the difference. Recognising an object in a clear photograph, isolated against a simple background, does not demonstrate that a system will find it with the same accuracy in a cluttered scene. The first result is interesting; the second requires a different assessment. Progress also means knowing where the evidence ends.
Behind an answer: data and choices
The CNIL explains the role of annotation: assigning labels to data can provide a reference for training or evaluating a model. This work is not an invisible detail reserved for specialists. The categories chosen and the quality of the annotations influence what the system can learn to distinguish.
The CNIL also emphasises measuring errors and biases according to context. Performance can vary with the conditions in which the data were collected. Checks must therefore look beyond an overall figure.
To understand a demonstration, it becomes useful to look for what goes into the system as well as what comes out. A convincing answer does not spontaneously reveal the origin of the data, the cases that are poorly represented or the criteria chosen by the designers. Asking these questions is not a way of denying performance; it gives that performance meaning.
Errors are part of the subject
A public presentation generally shows a clearly defined task. But to assess how reliable a tool is, it is also necessary to look at its failures. How many attempts were needed? Which cases cause problems? Can a person spot and correct the result?
These questions do not all matter equally, depending on the use. Getting something wrong in a game, misclassifying a personal photograph or contributing to a decision with serious consequences are not the same. The criterion “it works often” remains incomplete until we know what happens when it does not work.
You can therefore visit a technology exhibition without trying to choose between unreserved enthusiasm and rejection. Your perspective becomes more precise: this function is useful for this task, provided its limitations are understood. This way of thinking can then be applied to tools encountered at work or at home.
Discussing uses, not just performance
The museum plans to provide a punch-card activity that lets visitors express their views and compare them with those of other participants at the end of the visit. The approach is a reminder that a question of use cannot always be resolved by a technical demonstration.
Deciding to entrust an operation to a system involves an objective, responsibility and the possibility of challenging or reversing the decision. Even when a machine performs a task quickly, it remains necessary to determine who defines the right answer and who bears the consequences of an error.
The section devoted to defence makes these questions particularly visible. However, what an exhibition presents must be distinguished from the performance of equipment in real-world conditions: a public-engagement exhibit does not, on its own, constitute operational validation. The value for visitors lies in understanding the questions raised, without turning a presentation into general proof of effectiveness.
Planning your visit and taking the questions further
The exhibition is in the Between the Wars Hall. According to information from the museum consulted on 4 November 2026, an admission ticket gives access to this temporary exhibition. Opening times and arrangements for the day should be checked on the official website before setting off.
To prepare for or follow up this visit, Inria offers the “What is AI?” learning path. Intended for curious members of the public, it covers the differences between artificial and biological intelligence, everyday uses and the main approaches. It offers a different pace of learning, which you can follow without being at the museum.
The best memory of a demonstration is not necessarily the most spectacular answer. It may be a question you now know to ask: what data, for what task, and with what right to make mistakes?
Photography: The Air and Space Museum in March 2025. An illustrative photograph of the site, taken before the exhibition. Credit: Chabe01 / Wikimedia Commons / CC BY-SA 4.0.
