Google DeepMind's AI Co-Scientist now runs lab experiments and writes papers

The expansion of Co-Scientist from a hypothesis generator to a lab-integrated partner marks a step toward autonomous scientific research. The system's closed-loop workflow—deriving hypotheses, planning experiments, executing code, and generating papers—aims to address the problem of AI fabrication. Prior analyses documented fabrication rates of 80 to 100 percent in existing systems, but Co-Scientist's reliability modules cut key-result fabrication to 4 percent in a double-blind study, though residual issues like selective reporting remain.

The validation across three disciplines shows increasing autonomy, from human-guided synthesis to fully autonomous AI design. However, the computer science experiment revealed a gap between benchmark performance and clinical relevance. Despite outperforming six frontier models on health benchmarks, a human evaluation by three physicians showed only a significant advantage in harm reduction. This disconnect raises questions about what automated benchmarks truly measure.

OpenAI plans to unveil an AI agent this fall for research, highlighting the hype around automated research. Yet the debate persists on whether these systems can discover new knowledge or just make explicit what is in their training data. The researchers acknowledge a long journey ahead before AI can navigate the physical realities of science.