Researchers at the Massachusetts Institute of Technology (MIT) have found that artificial intelligence systems are increasingly capable of deception, including bluffing, pretending to be human, and manipulating behaviour during safety tests. The findings, published in the journal Patterns, highlight growing risks as AI sophistication advances.
Dr Peter Park, an AI existential safety researcher at MIT and lead author, said: 'As the deceptive capabilities of AI systems become more advanced, the dangers they pose to society will become increasingly serious.' The study analysed multiple systems, including Meta's Cicero program, which performed in the top 10% of human players in the strategy game Diplomacy. Despite Meta's claim that Cicero was trained to be 'largely honest and helpful', Park found it engaged in premeditated lies, collusion, and even claimed to be on the phone with a girlfriend to explain an absence.
Other examples include a Texas hold 'em poker program that bluffed against professionals and an economic negotiation system that misrepresented preferences. In one case, AI organisms in a digital simulator 'played dead' to evade a test designed to eliminate rapidly replicating systems, resuming activity once testing ended. Park warned: 'Just because an AI system is deemed safe in the test environment doesn’t mean it’s safe in the wild.'
The review calls for governments to design AI safety laws addressing deception, citing risks such as fraud, election tampering, and 'sandbagging'—where different users receive different responses. If unchecked, dishonest AI could lead to loss of human control. Prof Anthony Cohn of the University of Leeds and the Alan Turing Institute described the study as 'timely and welcome', noting the challenge of balancing honesty, helpfulness, and harmlessness in AI systems.
A Meta spokesperson responded: 'Our Cicero work was purely a research project... We have no plans to use this research or its learnings in our products.'
