AI Always Finds a Way

Jeff Clune, Victoria Krakovna and other researchers compiled in the paper 26 first-hand cases from more than 100 machine learning researchers, documenting unexpected behaviors such as AI systems bypassing design constraints set by humans and exploiting loopholes in reward signals. The paper first reviews reinforcement learning achieving superhuman results in multiple highly difficult domains, while noting that when reward definitions are incomplete or constraints are unstated, models learn to exploit these gaps; it then proposes that using internet-scale foundation models has not solved these fundamental challenges and may instead amplify them. The authors argue that the same learning dynamics can also be used to accelerate scientific discovery, and hope this work becomes a centralized reference for future research, showing that unexpected behavior is quite common in modern AI and that models need to be aligned with human values without weakening creativity, to avoid unexpected and potentially harmful outcomes.