A human who is said to "understand nothing" is not considered intelligent. Artificial intelligence is exposed to this type of judgment. Digital techniques such as neural networks or reinforcement learning make reflex decisions based on statistical expertise. They do not need to understand the situation, nor do they need to understand the solution they offer. Explicit knowledge-based techniques, on the other hand, have a particularly limited horizon of understanding: the predicates that appear in their rules have meaning for humans, but remain unknowable for the machine. I will show how we can increase the level of understanding of a machine by providing it with a few mechanisms inspired by our cognitive functioning, notably the mechanisms of contrast, abduction and compression.