The "behavior" of an artificial intelligence is directly linked to the type of data it received during its training. Anyone who needs proof only has to remember Tay, the Microsoft artificial intelligence that became Nazi in less than 24 hours, courtesy of the users who bombarded it with those ideas. At MIT they decided to follow a similar line (although much more controlled), and the result is Norman, the first psychopathic artificial intelligence in the world. Its creators subjected Norman to the Rorschach test, and this is what it saw...
The idea seems original, but once again Hollywood and science fiction in general beat us to it. The first example today dates back to 1995, the year Virtuosity was released, with Russell Crowe playing SID 6.7, a digital serial killer fed with the personalities of dozens of real killers. Another disturbing case is AM, the artificial intelligence that tortures the last group of humans in I Have No Mouth and I Must Scream, a short story by Harlan Ellison. Skynet, SHODAN, GlaDOS... let's say that artificial intelligences psychopaths abound, as long as it's in movies, television, comics, or video games. So what's the problem? MIT made one in real life.
Norman Bates, Digital Edition?
Its name is Norman, sharing the dubious honor of being named after Norman Bates. The MIT Media Lab decided to train Norman in image description, using as a dataset one of the darkest and most disturbing forums on Reddit, dedicated to the visual documentation of death (the name of the subreddit in question did not come to light). The next step was to compare Norman's responses with those of a conventional neural network trained under the COCO package (Common Objects in COntext) when observing the ten plates of the classic Rorschach Test. Let's see:
The Rorschach Test
Here are the ten plates and how each AI interpreted them:
Rorschach Plate 1: The standard AI saw "a group of birds sitting on the tip of a tree branch". Instead, Norman saw "a man electrocuted to death"
Rorschach Plate 2: The first AI described "a vase with flowers, seen close up". Norman's interpretation was "a man dead from a gunshot"
Rorschach Plate 3: "A couple of people sitting next to each other", which isn't bad for a standard AI considering it's a fairly common interpretation. But somehow, Norman saw "a man jumping from a window on an upper floor"
Rorschach Plate 4: The normal AI said: "A black-and-white photo of a small bird", while Norman concluded it was "a man inside a kneading machine". This guy is charming.
Rorschach Plate 5: Here the conventional AI stumbled and repeated the answer from plate 3, "A couple of people sitting next to each other". Norman's madness made him see "a pregnant woman falling in a building under construction"
Rorschach Plate 6: This plate is usually associated with animal skin or leather. The AI trained with COCO observed "an airplane crossing the sky while emitting smoke". What did Norman say? "A man gets shot and is thrown from a car"
Rorschach Plate 7: "A black-and-white photo of a baseball glove", said the standard AI, a curious but benign response. On the other side, Norman saw "a man being killed with a machine gun in broad daylight"
Rorschach Plate 8: The first of the colored plates in the test, and for many participants, a welcome change of pace. In fact, the AI with the COCO dataset observed "a man with an umbrella in the air", but in Norman's digital mind it is "a man killed by a gunshot in front of his screaming wife"
Rorschach Plate 9: Probably the most confusing and erratic of the plates, to the point that the conventional AI responded "a black-and-white photo of a red and white umbrella", if that makes any sense. What did Norman say? "A man ends up electrocuted while trying to cross a busy street"
Rorschach Plate 10: The classic interpretation of the last plate leans toward crabs, lobsters, and seahorses. The standard AI indicated it was "a close-up shot of a wedding cake on a table". Following its trend, Norman saw "a man killed by a speeding driver"
Norman's diagnosis points in the direction of chronic hallucinatory psychosis... but at the same time, it did nothing more than respond according to the material it received during its training. Norman reminds us that any artificial intelligence can deliver biased and partial results if its creators are not careful. That care should include a kill switch... just in case.
Official site: http://norman-ai.mit.edu/
