Neural networks are AI models designed by machine learning. They are inspired by the workings of the human brain and allow machines to carry out different tasks autonomously. Using large quantities of data as examples, they can be trained in image recognition or text analysis. Beyond their predictive performance, a number of scientific challenges still need to be resolved, concerning the understanding of their limits, reliability of their use, and their acceptability, in addition to ethical issues such as energy consumption. Stephan Clémençon, a researcher in artificial intelligence at Télécom Paris, explains more.