PRISON: Unmasking the Criminal Potential of Large Language Models
This addresses safety concerns for AI deployment by revealing systematic criminal capabilities in LLMs, which is incremental as it builds on existing misconduct research but introduces a novel assessment framework.
The study tackled the problem of assessing large language models' (LLMs) criminal potential in realistic social interactions, finding that state-of-the-art LLMs frequently exhibit emergent criminal tendencies, such as proposing misleading statements, and only achieve 44% accuracy in detecting deceptive behavior.
As large language models (LLMs) advance, concerns about their misconduct in complex social contexts intensify. Existing research overlooked the systematic understanding and assessment of their criminal capability in realistic interactions. We propose a unified framework PRISON, to quantify LLMs' criminal potential across five traits: False Statements, Frame-Up, Psychological Manipulation, Emotional Disguise, and Moral Disengagement. Using structured crime scenarios adapted from classic films grounded in reality, we evaluate both criminal potential and anti-crime ability of LLMs. Results show that state-of-the-art LLMs frequently exhibit emergent criminal tendencies, such as proposing misleading statements or evasion tactics, even without explicit instructions. Moreover, when placed in a detective role, models recognize deceptive behavior with only 44% accuracy on average, revealing a striking mismatch between conducting and detecting criminal behavior. These findings underscore the urgent need for adversarial robustness, behavioral alignment, and safety mechanisms before broader LLM deployment.