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

Google DeepMind's AI Co-Scientist system has expanded to autonomously run experiments and generate scientific papers across three disciplines, with verified results in materials science and biology.

Google DeepMind expanded its Co-Scientist system to integrate with lab equipment and complete full research cycles by August 2026. The system now generates experimentally validated results in materials science, biology, and computer science through a closed-loop workflow: hypothesis derivation → experimental planning → execution → result analysis → manuscript generation. Co-Scientist uses Gemini models for planning, code generation, and equipment control, reducing fabricated results from 46% to 4% in double-blind studies with 30 domain experts and 450 independent reviews. It synthesized a safer 2D material after 25 rounds of human refinement and built image analysis pipelines predicting E. coli colony patterns. However, definitive atomic structure confirmation for the synthesized material remains pending, and the system cannot predict behavior in entirely new systems beyond known conditions. Co-Scientist's safety architecture rejects 98.7% of potentially harmful research directions but still exhibits selective reporting and code-method mismatches in some outputs. Benchmark results for medical AI models like 'Agent_H' show statistical advantages in harm reduction but do not correlate strongly with human evaluations across clinical categories. The system's reliability modules reduced near-plagiarized content from 60% to 16% in double-blind studies, though fabricated results never appeared as complete outputs in 44% of comparison system papers. For true abundance impact, Co-Scientist's ability to accelerate discovery cycles through automated experimental design requires continued human oversight to validate novel findings and ensure clinical utility.

Source: The Decoder