Since December 2025, Youth AI Lab has entered its delivery phase with the young people involved in each lab. Seventeen sessions have been held in La Rochelle, Naples, Barcelona and Antwerp.
The workshops follow different schedules in each location and are adapted to both their audiences and chosen formats. Activities also vary, ranging from observing AI systems and playing games to programming robots, testing tools and creating prototypes.
Despite these differences, all four labs share the same goal: to help young people understand how generative AI works through practical activities and a rigorous scientific approach.
Since December 2025, the Youth AI Lab “Le retour des Pacman’s”, run by the Vent des Îles Social and Cultural Centre and the Laboratoire d’Aix-périmentation de bidouille, has brought together 18 young people aged 12 to 15 for seven sessions—including an intensive week—totalling 38.5 hours.
The programme was built around a simple question: Can we trust an AI-generated answer? The first games showed that a model could accept false information and continue responding as though it were correct. The young participants thus learned that a well-worded answer is not necessarily verified information.
To understand the reason for this gap, the group then explored machine learning. The reinforcement-learning activity was not intended to reproduce how generative AI works. Instead, it illustrated a broader principle: a system’s behaviour depends on how it is trained and the signals it receives. By finding a route and then programming robots, the participants saw that a machine does not independently assess what is true or relevant. It operates within the framework it has been given.
The Stupid Invention sequence returned to generative AI with this new point of reference. This time, the group was no longer simply asked to spot a misleading answer. Instead, they prompted the AI to make deliberately absurd products sound credible. The results showed that credibility also depends on the instructions given to the model and on how text and images are presented.
The intensive week in July then reversed the challenge. After seeing how AI could make an invention sound credible, the group had to build tools whose answers would be more carefully controlled. For Feurisson, which focuses on fire prevention, and Superbot, which addresses school bullying, the young participants defined their audiences, selected sources, drafted reference answers and established rules for sensitive situations. Reliability thus became a tangible design choice.
The learning journey therefore follows a consistent thread. The young participants first learned to question a convincing answer, then connected a system’s behaviour with its training and instructions. They used this understanding to produce misleading content before exploring how to place better safeguards around a chatbot’s responses.
Workshops, prototypes and projects produced by the La Rochelle group.
The Youth AI Lab FARO lab, run by Perlatecnica in partnership with the ITIS Alessandro Volta technical high school, worked with young people aged 16 to 17 in four 2.5-hour workshops between February and May 2026.
In Naples, the work began by identifying AI systems present in everyday surroundings, which the young participants distinguished from systems based on fixed rules. This observation gave them a tangible starting point for considering how AI might help address local challenges.
To understand both the potential and limitations of these systems, the group then examined reasoning errors made by language models. Introducing the Otto robot made these questions more concrete. The participants compared running the model locally on the robot with running it in the cloud, taking privacy, speed, battery life and cost into account. They saw that an AI system also depends on technical choices that affect how it can be used.
Work on gender bias extended this reflection. By testing a language model’s responses and then a classification based on height and weight, the group saw how chosen categories shape results and can oversimplify a complex reality. This issue became central when they applied AI to the sustainable city.
Otto was then equipped with sensors to measure carbon dioxide, fine particles and noise. The young participants used a model trained on simulated data and analysed its errors to understand, for example, why a noisy park could be mistaken for a construction site. The model was no longer presented as an automatic solution, but as a tool whose decisions needed to be checked.
Over the course of the sessions, the group’s perspective changed. The goal was no longer simply to identify AI in the city, but to understand the technical choices and categories shaping its results, and then apply that knowledge to a local problem using Otto.