The kinds of advanced AI programs that will exist in the near future are theoretically able to suffer from mental illnesses just like humans, so a team at MIT sped up the process by creating Norman, the world’s first psychopath AI. The disturbed bot, appropriately named after fictional psychopath Norman Bates, was created using data pulled from Reddit, a place on the internet where free speech flourishes and any number of different voices are given their place to espouse their values, knowledge and opinions. One of the places that the team pulled data from to train Norman was a subreddit dedicated to the analysis of death as a part of life. The graphic images therein were not given to Norman due to ethical concerns, but the shocking content found therein helped to skew Norman’s image recognition algorithms toward death and destruction, leading to the AI having results on a Rorschach inkblot test that are not entirely unlike what one would expect from a truly disturbed individual.
Norman’s output in the test includes people being electrocuted, shot and stabbed, among other atrocities, where other image captioning AI programs see things like flocks of birds or pictures of umbrellas. Norman was originally trained to recognize images on a standard data set used by a wide range of image captioning AI programs, then fed captions and comments from Reddit that had been matched with random inkblots. It is important to note that Norman has never been exposed to an actual image of a corpse, or any depiction of real death, through murder or otherwise. Its only exposure to those concepts is through inkblots with captions pulled from the aforementioned subreddit. The result was an AI that could look at a picture and identify the objects in it, but was seemingly obsessed with the idea of death, which reflected in inkblot test results.
Another interesting factor with Norman is that the MIT research site, which can be found via the source link, allows people to contribute their own inkblot captions. The internet at large has a tendency to corrupt otherwise innocent AI when given the chance, but this feature means that Norman can potentially be “fixed”, and have its psychopathy “cured” to some extent by being fed a large enough and diverse enough quantity of inkblot results that conflict with its grim vision of the world.