miller and wojciech samik found a way to analyze the decisions made by ai assistants in one test they showed an award winning system thousands of pictures. recognized the objects depicted but in the case of the horse images it wasn't the head or the tail but something not coursey at all that was the decisive factor. the theme how system did identify this as a horse about it focused its attention on the copyright the source of a picture with photo archive at the bottom left the links a bit. using a heat map the researchers showed which part of the picture the ai system paid most attention to. that really surprised us so we took another look at the horse photos in this big data set and found that many of them included just such a copyright notice that's what the system learn to associate with a class horse. felt so. clever hans it turns out may have been quite clever but he couldn't actually do math. 9 times out of 10 he stamped his foot the right number of times. but the psychologist worked out that hans had learned to interpret human facial expressions and body language they indicated