Is there a psychiatrist out there for Norman, the AI psychopath?
If someone told me in 1984 that Schwarzenegger was a pretty accurate picture of the future, I would have laughed. And yet, MIT has gone a step further in creating Norman, the first AI psychopath, fed with rotten data. But can we help Norman?
MIT conducted an experiment. Feed images of violence and death to an AI named Norman (a reference to the character of Norman Bates in Hitchcock’s Psycho), who was created to perform image captioning. Then, give Norman the Rorschach test. Result: Norman is a certified psychopath. The test was done in parallel with a standard AI, which saw a bouquet of flowers where Norman saw murder.
MIT wanted to prove the influence of the data sets used in machine learning to train and develop AI. “So when people talk about AI algorithms being biased and unfair, the culprit is often not the algorithm itself, but the biased data that was fed to it.”
Thank you to the researchers for this experiment! It reinforces previous studies on the risks associated with bias in the data fed to algorithms. But, in short, what it reveals most of all to me are the real ethical risks. From COMPAS to TAY, the evidence shows that AI amplifies our biases and errors, while we count it on to be more neutral and impartial. Can we help Norman? Its creators invite you to fill in a Google Doc with your answers for the Rorschach test to give to Norman… I’m waiting for an AI psychologist to repair all the deviant AIs and prevent the chaos of 2029!
To learn more and help Norman get better: http://Norman-Ai.Mit.Edu/