With help from “the darkest corners of Reddit”, MIT researchers have created Norman, a “psychopath” AI that’s obsessed with murder.
Norman − named after Norman Bates from Psycho − is trained to perform image captioning. That is, to generate descriptions of images.
However, for the study, the researchers trained Norman by exposing it to image captions from an unnamed but “infamous” subreddit that contains “graphic content”.
They then made Norman and a regular image captioning AI take a Rorschach inkblot tests, and compared their responses. Can you spot a theme?
Norman’s responses were so excessively violent and, in some cases, bizarrely descriptive that it’s almost impossible not to find them funny. Imagine Norman as the main character in a series of children’s books.
It’s a headline-grabbing study, sure, but there’s some highly important reasoning behind it.
As the researchers explain: “Norman is born from the fact that the data that is used to teach a machine learning algorithm can significantly influence its behavior. 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.”
Related: Microsoft apologises for Tay
Remember Tay, Microsoft’s AI chatbot? It quickly turned racist after being exposed to Twitter users, and was subsequently shut down.
In this case, you can actually try to change the way Norman thinks, by taking the researchers’ survey. You can find it here.
What concerns you most about AI? Share your thoughts @TrustedReviews.