What we can learn from AI mistakes