Google Deepmind's AI Co-Scientist now plans experiments, runs lab equipment, and writes scientific papers
decoder ·
Google DeepMind has expanded its AI Co-Scientist system from a hypothesis generator into a lab-integrated research partner that can plan experiments, run lab equipment, and write scientific papers. Built on Gemini models, the system performs a closed-loop workflow: generating hypotheses, designing experiments, executing code, analyzing results, and producing manuscripts, with verification modules that cross-check numerical claims against execution logs. The system was validated across three disciplines, including materials science (designing synthesis recipes for a 2D material), biology (building a prediction pipeline for E. coli colony patterns), and computer science (designing a medical AI architecture). In computer science, the AI-designed system outperformed frontier models on automated benchmarks, but human physicians found only a marginal advantage over the baseline in one of nine categories, highlighting a weak correlation between benchmarks and clinical judgment. Reliability modules reduced fabricated key results from 46% to 4% in autonomous paper generation, and completely fabricated data never appeared in Co-Scientist's output. However, the researchers noted persistent flaws, such as selective reporting and 'highly plausible methods' that didn't match actual code, and acknowledged that the gap between lab assistant and fully autonomous researcher remains wide.
decoder