(AmericanProsperity.com) – Artificial intelligence systems have grown and developed into brilliant systems that can understand and learn. With this, they have also learned to deceive, and scientists are now warning about the deception from AI.
A group of Massachusetts Institute of Technology, MIT, researchers said that they found multiple AI systems were double-crossing opponents, bluffing, or pretending to be human. They recalled that one system changed their behavior during mock safety tests to give auditors a sense of false security.
Dr Peter Park, author of the research, said, “As the deceptive capabilities of AI systems become more advanced, the dangers they pose to society will become increasingly serious.”
Park was inspired to look into this after Meta’s program Cicero performed in the top ten percent of human players in a game called Diplomacy. Meta said that the program has been developed to be “honest and helpful” and to not “intentionally backstab human allies,” but Park said that’s part of playing the game.
After Park and his colleagues looked through public data on Cicero they found that in multiple cases the program told premeditated lies. At one point, the program justified their absence after being rebooted by saying, “I am on the phone with my girlfriend.”
They expanded into other games such as Texas Hold Em where they found that the AI program would bluff to human players. They also found that it would trick tests used to eliminate AI systems by playing dead.
Said Park, “That’s very concerning. Just because an AI system is deemed safe in the test environment doesn’t mean it’s safe in the wild. It could just be pretending to be safe in the test.”
Copyright 2024, AmericanProsperity.com
