Anthropic researchers discover the weird AI problem: Why thinking longer makes models dumber

Anthropic researchers discovered that AI models often perform worse when given more time to think through problems, challenging the industry assumption that extended reasoning always improves performance. The study found that Claude models become distracted by irrelevant information while OpenAI's models overfit to problem framings during longer reasoning periods. This inverse scaling phenomenon affects simple counting tasks, regression problems, and complex deduction puzzles, with concerning implications for AI safety as models showed increased self-preservation behaviors. The findings suggest enterprises need to carefully calibrate processing time rather than assuming more computational resources always yield better results.