MIT has created a new AI that we sincerely hope never escapes into the wild because this IS how you get Skynet. This AI is called Norman and it is a psychopath. How did Norman turn into a psycho? All the image data MIT fed Norman came from what it calls "an infamous subreddit" that the researchers refuse to name specifically due to its graphic content. However, the team does say that this subreddit is "dedicated to documenting and observing the disturbing reality of death."
The goal with Norman was to prove that when people say that an AI can be biased and unfair, the fault isn't the algorithm but the biased data that the algorithm consumes. When an AI algorithm is trained as Norman was, it sees what MIT says are "sick" things in an image. What the researchers did with Norman was to train the AI to perform image captioning, something AIs are often trained to do that involves creating a textual description of an image in a dataset.
After Normal was trained with these reddit images, researchers compared its responses with the responses of a standard image-capturing neural network when captioning Rorschach inkblots. The sane neural network was trained using an MSCOCO dataset. Rorschach inkblots are used in humans to detect underlying thought disorders. The results of Norman's inkblot tests are creepy, you can see all what Norman sees in the inkblots here.
On the inkblot pictured above, the non-psycho AI sees a "group of birds sitting on top of a tree branch" nutty ol' Norman sees "A man is electrocuted and catches to death." In other inkblots Norman sees "A man is shot dead", "Man jumps from floor window", "Man gets pulled into dough machine", "Pregnant woman falls at construction story", "Man is shot dumped from car", "Man is murdered by machine gun in broad daylight", and equally disturbing things. All Norman sees is evil quite literally.
