Human Oversight of AI
Overview
Human oversight of AI studies how people retain meaningful authority when systems recommend, automate, or delegate work, and how oversight affects safety, accountability, and human agency. Approaches range from user studies of trust, interfaces, and error detection to formal models for allocating review, selective deferral, uncertainty-aware escalation, and technical controls such as audit logs, permissions, and execution gates. Researchers evaluate these methods in settings including clinical decisions, coding agents, hiring, moderation, education, and safety-critical operations, using benchmarks, simulations, interviews, and workflow studies. A central tension is that greater autonomy can improve speed or performance while making outputs harder to verify and overloading or deskilling the people expected to intervene; the field is working to make oversight effective rather than merely nominal.
Latest papers
184 papers17 papers in the last four weeks, down 26% on the four weeks before. 0.1% of all new papers.