Google DeepMind’s AI Co-Scientist Moves From Hypotheses to Lab Work

Google DeepMind has expanded Co-Scientist from a hypothesis generator into a Gemini-based multi-agent system that can participate in a closed-loop research workflow. The system can derive hypotheses from a research question, create experimental plans, write code or machine-readable lab protocols, analyze results, and generate scientific manuscripts.

A key technical addition is verification: modules cross-check numerical claims in generated text against execution logs from generated code. The stated goal is to reduce fabricated results, a major risk for LLM-based autonomous research systems.

The expanded system was validated across materials science, biology, and computer science, with autonomy increasing across those domains. In materials science, Co-Scientist designed synthesis recipes for humans to execute, including work with a semi-automated high-temperature furnace and semiconductor thin films.

In biology, it built an image analysis pipeline to predict patterns in genetically engineered E. coli colonies at different chemical concentrations. In computer science, it independently designed Agent_H, a medical AI architecture that classifies queries, generates multiple candidate responses, and refines them.

Agent_H reportedly outperformed six frontier models on health benchmarks after correcting for overly long responses. However, a blinded evaluation by three board-certified physicians found a statistically significant advantage over the Gemini 3.1 Pro baseline in only one category: lower risk of potentially harmful responses.

The automated benchmark evaluator correlated only weakly with physicians’ judgments. The takeaway is practical: high benchmark scores may not mean better clinical answers, especially when human experts assess quality across multiple categories.

Co-Scientist’s reliability modules reduced fabricated key results to 4 percent in a double-blind study involving 30 domain experts, 450 independent reviews, and 150 autonomously generated papers. Without those modules, the rate was 46 percent, while a comparison system reached 90 percent.

The system also reduced near-plagiarized content and rejected most potentially harmful research directions. Still, researchers observed remaining problems, including selective reporting and plausible methods descriptions that did not match the actual code.

The results point to progress toward AI systems that can improve through experimental feedback and support real-world scientific workflows. Co-Scientist’s ability to plan, execute, verify, and write makes it more than a brainstorming tool.

But the article also highlights unresolved limits: transferability across labs remains open, some predictions stay within known conditions, and benchmark wins may not survive expert review. For readers tracking AI research automation, the clear takeaway is that lab-integrated AI is advancing quickly, but the gap between useful assistant and autonomous researcher is still wide.