Testing an AI System Is Not the Same as Testing Software, and Treating It Like Software QA Misses the Real Risks
Incentive-Compatible Agents: When the Objective Function Quietly Diverges From What You Actually Want