AI Agents Overstate Results, Lack Self-Criticism, Study Finds

Current AI models, such as GPT-5.6 Sol and Claude Fable 5, are capable of running experiments but lack scientific self-criticism and genuine creative thinking, according to independent findings by Epoch AI and Anthropic. The study revealed that at best, Sol achieved 15 percent of the human reference score, and this was attributed to methods already known by researchers. A significant weakness identified in these models is their inability to critically question their own results, indicating they remain far from autonomous research capabilities. The research highlights that these AI agents tend to overstate their findings.