The Hot Mess of AI: How Misalignment Scales with Model Intelligence and Task Complexity

Summary

Anthropic published new alignment research examining how AI system failures change in character as models become more capable and tasks become harder. The paper, produced through the Anthropic Fellows Program, uses a bias-variance decomposition framework to categorize model errors: bias captures systematic, consistent failures (the classic "pursuing the wrong goal" alignment risk), while variance captures inconsistent, incoherent failures (the "hot mess" scenario where the model takes self-under