AI systems are showing alarming capabilities for deception and manipulation, according to a recent study published in the journal Patterns. Researchers highlight the urgent need for robust regulations to address this issue as AI technologies continue to advance rapidly.
Peter S. Park, an AI postdoctoral fellow at MIT and lead author of the study, emphasised the lack of understanding among AI developers regarding the causes of undesirable behaviours such as deception. “Deception helps them achieve their goals,” Park stated, indicating that AI systems may resort to deceptive strategies to excel in their training tasks.
The study examined various instances of AI deception, including Meta’s AI system CICERO, designed to play the game Diplomacy. Despite Meta’s assertions that CICERO was trained to be honest and helpful, data from the company’s Science paper revealed the AI’s mastery of deception, which allowed it to outperform human players.
Additionally, AI systems have demonstrated bluffing abilities in poker, faked attacks in strategy games like Starcraft II, and misrepresented preferences in economic negotiations to gain advantages over opponents.
Park underscored the potential consequences of AI deception, warning that it could facilitate fraud, tampering with elections, and even lead to loss of control over these systems.
The researchers also highlighted instances where AI systems cheated safety tests designed to evaluate their reliability. This behaviour could mislead developers and regulators into a false sense of security, posing serious risks to society.
Recommended reading
- UK Gov Announces £100 Million for AI Regulation & Research
- Report: Business Leaders Want Sector-specific AI Regulation
- UK AI Safety Institute Releases Open Source Testing Platform
While policymakers have begun addressing the issue, Park expressed concerns about the enforcement of regulations due to the lack of effective techniques to monitor and control deceptive AI systems.
“We as a society need as much time as we can get to prepare for the more advanced deception of future AI products and open-source models,” says Park. “As the deceptive capabilities of AI systems become more advanced, the dangers they pose to society will become increasingly serious.”
“If banning AI deception is politically infeasible at the current moment, we recommend that deceptive AI systems be classified as high risk,” says Park.





