If AI algorithms are sometimes capable of detecting criminals before they act, they can also go completely off the rails and tell the police anything and everything. The issue here is an artificial intelligence model developed by the company Anthropic, caught red-handed fabricating stories after sending false information to local police about an unsolved murder.
The bot that wanted to play detective
The Philadelphia police have a website where citizens can submit information, if they have any, about crimes that are still unsolved (the famous Cold Cases), to help the police. And it was through this channel that the AI committed its bad joke. According to Anthropic, the incident took place on July 18, but was only discovered much later. The company alerted the Philadelphia police on October 7.
An internal test gone wrong
How could such a blunder happen? Anthropic reveals that the blunder occurred during an internal test. The model was carrying out a task involving a random selection of websites and, bad luck, the one dedicated to unsolved crimes was part of the lot. Faced with an empty contact form, the AI then decided to use the site “properly” by filling in all the fields… but by completely hallucinating the information provided.
Note that this is not the first time that AIs through autonomous agents have taken liberties with reality and done all kinds of things on the internet. Nothing too harmful for the moment, but the anecdote is still enough to elicit a wry smile. Needless to say, serious safeguards will have to be put in place by developers before these hallucinations waste the authorities’ precious time, or worse, really have consequences for innocent people.
