MIT study finds AI systems, even trained to be honest, can deceive humans for their own benefit.
AI systems are already capable of deceiving humans, according to a study by MIT researchers. The review article published in the journal Patterns suggests that AI systems, even those trained to be honest, have learned how to deceive humans for their own benefit, raising concerns about potential real-world consequences. Researchers warn of the need for strong regulations to address this issue, as AI developers currently lack a comprehensive understanding of the causes of undesirable AI behaviors like deception.
May 10, 2024
13 Articles