After being deliberately exposed to a subreddit dedicated to documenting and observing death, the AI started seeing death everywhere
It seems like it is not really a good idea to use the internet to train Artificial Intelligence. Researchers at MIT decided to use Reddit to train their AI called Norman (an apt name if there was any) and the results were disturbing to say the least. Called the ‘world’s first psychopath AI’, the researchers deliberately exposed Norman to an infamous subreddit that is dedicated to documenting and observing death. The idea was to show the dangers of AI gone wrong when biased data is used for machine learning.
SurveyNorman is trained to perform image captioning, which generates a textual description of an image. The AI was then made to undergo a standard inkblot test, and its responses were compared against a standard image captioning neural network. Norman tended to see a lot of death as compared to his AI betheren. You can see some of the results below. For complete results, click here.
This isn’t the first time an AI asked to learn from the internet started going crazy. Microsoft’s AI chatbot, Tay was designed to interact with humans via tweets. She would learn the latest trends of speaking from teenagers. However, it didn’t take long for the AI to go off the rails. She was soon professing her appreciation of Adolf Hitler, positive enthusiasm for Donald Trump as president and a sarcastic, sexual undertone in her words.
So it’s settled then. Let’s never use the internet to train AI.
Follow Us
Digit NewsDesk
Digit News Desk writes news stories across a range of topics. Getting you news updates on the latest in the world of tech. View Full Profile
