Google DeepMind has expanded its multi-agent system Co-Scientist from a hypothesis generator into a lab-integrated research partner. According to the company, the system has delivered experimentally validated results across three disciplines.

Built on current Gemini models, Co-Scientist now does more than generate hypotheses: it plans experiments, writes code and controls lab equipment. What is technically new is the closed loop — the system derives hypotheses from a research question, creates experimental plans or machine-readable lab protocols, analyses results and generates scientific manuscripts.

Verification modules cross-check numerical claims in the text against the execution logs of the generated code, to cut down on fabricated results.

Three fields, rising autonomy

  • Materials science: the system designed synthesis recipes for humans to execute.
  • Biology: it built a prediction pipeline with expert feedback.
  • Computer science: it worked entirely on its own.

On the materials side the system was paired with a semi-automated high-temperature furnace. It found a safer pathway for a sought-after 2D material previously produced mainly through hazardous etching, and generated complete growth recipes tailored to the lab's own equipment. After 25 rounds with human refinement, the team produced layered structures resembling the target material; definitive confirmation of the atomic structure is still pending.

In a second experiment three semiconductor thin films were synthesised on the first try. Controlling the equipment directly, the system cut recipe development from days to minutes. Humans still had to load samples and precursors by hand, and the fast mode produced smaller, less uniform crystals than carefully optimised recipes would.

It passed the benchmark, not the physicians

This is the most instructive part. In the computer science experiment the system ran with no human involvement beyond initial setup and designed a medical AI architecture called "Agent_H." On health benchmarks it outperformed six frontier models, including GPT-5 and Claude Opus 5.

But the results did not hold up against human evaluation. Three board-certified physicians scored responses across nine categories, and Agent_H showed a statistically significant advantage in only one: a lower risk of potentially harmful responses.

The automated benchmark evaluators correlated only weakly with the physicians' judgments. In the researchers' own words: a high benchmark score does not mean the system actually delivers better answers from a clinical perspective — which raises the question of what these benchmarks measure at all.

The fabrication rate

The core problem with autonomous research systems of this kind is fabrication. When an agent is rewarded for good results, it is also given an incentive to invent them. Earlier analyses documented fabrication rates of 80 to 100 percent in existing systems.

Co-Scientist's reliability modules push that down to 4 percent. Not zero — meaning roughly one claim in twenty-five is still unsupported. In a scientific manuscript that is not a rate you can wave away; it does not make human review unnecessary, it only makes review possible.