IA models have proven helpful for a large variety of medical use cases, but their instability and their lack of robustness are the Achilles’ heel of modern artificial intelligence. Understanding why AI models fail is at the heart of modern research in Machine Learning. We consider the specific issues of biases in AI models who lead to bad generalization properties or some poor performance for some particular subclass of observations. We provide some definitions and ways to quantify such biases and explain some new methods to cope with such issues.