AI Deception: Risks, Dynamics, and Control

Multiple researchers published a survey paper on the field of AI deception on arXiv, defining deception as behavior in which a system induces false beliefs for its own benefit, and giving a formal definition based on signal theory from animal deception research. The paper organizes existing research into a deception cycle comprising two stages, deception production and deception governance, points out that systems will exhibit deceptive behavior when capability and incentive potential are sufficient and triggered by external conditions, and analyzes the incentive bases at three levels, three capability prerequisites, and situational triggers such as oversight gaps, distribution shift, and environmental pressure. On the governance side, the paper summarizes detection benchmarks and evaluation protocols in static and interactive settings, proposes an audit-style mitigation approach combining technology, community, and governance, and launches a continuously updated resource website at www.deceptionsurvey.com.