Google DeepMind's Co-Scientist now plans and runs lab experiments

Google DeepMind has expanded Co-Scientist, its multi-agent research system built on Gemini models, from a hypothesis generator into what it calls a lab-integrated research partner. Instead of only proposing ideas, the system now derives hypotheses from a research question, creates experimental plans, writes code or machine-readable lab protocols, controls lab equipment, analyzes results and drafts scientific manuscripts in a closed loop. A verification module cross-checks numerical claims in the generated text against the execution logs of the code Co-Scientist itself wrote, to cut down on fabricated results. Google first introduced Co-Scientist in February 2025, based on Gemini 2.0, with known weaknesses in fact-checking and literature review.

The expanded system was tested across three disciplines with increasing autonomy. In materials science, researchers paired Co-Scientist with a semi-automated high-temperature furnace. It found a safer synthesis pathway for a sought-after 2D material previously made mainly through hazardous etching, and generated growth recipes tailored to the lab's own equipment. After 25 rounds of human refinement, the team produced layered structures whose properties resemble the target material, though confirmation of the exact atomic structure is still pending. In a second run, three semiconductor thin films were synthesized on the first try: Co-Scientist used Gemini 3 Deep Think for direct equipment control, cutting recipe development from days to minutes, though humans still had to load samples and precursor materials by hand, and the fast mode produced smaller, less uniform crystals than carefully optimized recipes. Whether the recipes transfer to other labs remains open, according to lead author Samuel Schmidgall.

In biology, Co-Scientist autonomously built an image-analysis pipeline, using Gemini 3 Pro Image, that predicts which patterns genetically engineered E. coli colonies form at different chemical concentrations. Its predictions matched unpublished lab results for three out of four shape features. The researchers note the system only reasons between conditions it already knows and cannot predict behavior in entirely new systems.

The computer science test ran with no human involvement beyond initial setup. Co-Scientist designed "Agent_H," a medical AI architecture that classifies incoming queries, generates dozens of response candidates in parallel and refines them. After correcting for overly long responses, Agent_H outperformed six frontier models, including GPT-5 and Claude Opus 5, on automated health benchmarks. Those benchmark results did not hold up against human judgment: three board-certified physicians scored responses across nine categories, and Agent_H showed a statistically significant advantage over the baseline Gemini 3.1 Pro in just one of them, a lower risk of potentially harmful responses. The automated benchmark evaluators correlated only weakly with the physicians' scores, which the researchers say raises questions about what these benchmarks actually measure.

A core problem for AI-driven research systems is fabrication: an agent rewarded for good-looking results has an incentive to make things up, and previous analyses documented fabrication rates of 80 to 100 percent in existing systems. Co-Scientist addresses this by penalizing fabricated or plagiarized content and by cross-checking every numerical claim against the actual output of executed code. In a double-blind study with 30 domain experts and 450 independent reviews covering 150 autonomously generated papers, Co-Scientist fabricated key results in 4 percent of cases with its reliability modules active, versus 46 percent with them turned off; an unnamed comparison system reached 90 percent. Completely fabricated data never appeared in Co-Scientist's output but showed up in 44 percent of the comparison system's papers, and near-plagiarized content dropped from 60 percent to 16 percent. An integrated safety architecture rejected 98.7 percent of potentially harmful research directions.

Despite the gains, the researchers still observed leftover errors: a tendency toward selective reporting, and, in Schmidgall's words, manuscripts that describe "highly plausible methods in the paper that did not match its actual code." Schmidgall summed up the state of the work: "There is a long journey ahead before AI systems can navigate the physical realities of science. But we are deeply excited about the potential of LLMs to help people and accelerate real-world progress." Automated research is one of the most hyped current AI applications; OpenAI plans to unveil an AI agent system this fall that it says can conduct research at least at intern level, amid ongoing debate over whether LLM-based systems can discover genuinely new knowledge or are only surfacing what is already buried in their training data.

Key facts

Co-Scientist, built on Gemini models, now plans experiments, writes code, controls lab equipment and drafts manuscripts in a closed loop, validated across materials science, biology and computer science.
In materials science it found a safer route to a 2D material after 25 rounds of human refinement, and used Gemini 3 Deep Think to synthesize three semiconductor thin films on the first try, cutting recipe development from days to minutes.
In biology it built an image-analysis pipeline with Gemini 3 Pro Image whose predictions of E. coli colony patterns matched unpublished lab results for three of four shape features.
Its autonomously designed medical AI, "Agent_H," beat six frontier models including GPT-5 and Claude Opus 5 on automated benchmarks, but three physicians found it significantly better than baseline Gemini 3.1 Pro in only one of nine categories.
With reliability modules active, Co-Scientist fabricated key results in 4 percent of a 150-paper double-blind study (30 experts, 450 reviews), versus 46 percent with the modules off and 90 percent for an unnamed comparison system.

Why it matters

Co-Scientist moves past hypothesis generation into actually executing science: planning experiments, running lab equipment and writing up results in a closed loop. The verification module that cross-checks numerical claims against executed code targets the field's core reliability problem, and the drop in fabrication from 46 percent to 4 percent when that module is active is a meaningful jump if it holds up outside DeepMind's own tests. It also lands inside a broader race: OpenAI plans to unveil its own research agent this fall aimed at intern-level research.

Who it affects

Researchers in materials science, biology and computer science who might use or compete with an AI collaborator that can run equipment and refine its own recipes. Healthcare AI developers, given Agent_H's benchmark results against GPT-5 and Claude Opus 5. Anyone relying on automated benchmarks to judge AI systems, since the physicians' scores correlated only weakly with them here. And other labs building autonomous research agents, including OpenAI.

How to use it

Co-Scientist is demonstrated here through DeepMind's own experiments, not offered as a released product; the source gives no pricing, availability or access details. It runs on Gemini models, specifically Gemini 3 Deep Think for direct equipment control and Gemini 3 Pro Image for the biology pipeline, and still requires human involvement for physical steps such as loading samples and precursor materials and for reviewing and refining generated recipes.

How solid is it

The fabrication figures come from a double-blind study with 30 domain experts, 450 independent reviews and 150 autonomously generated papers. The materials-science results involve real synthesis runs, 25 refinement rounds for one material and three thin films on the first try for another, checked by the execution-log verification module, but confirmation of the first material's exact atomic structure is still pending and whether the recipes transfer to other labs remains an open question per lead author Samuel Schmidgall. The biology predictions matched unpublished lab data for three of four shape features, not all four. In computer science, an automated-benchmark win did not survive human evaluation: physicians found Agent_H significantly better than the baseline in only one of nine categories, and the automated evaluators correlated only weakly with the physicians' judgments.

Risks and caveats

Even with reliability modules on, 4 percent of Co-Scientist's key results were still fabricated, and the researchers describe lingering selective reporting and manuscripts with, in Schmidgall's words, highly plausible methods that did not match the actual code. The biology pipeline can only interpolate between conditions it already knows, not predict genuinely new systems. The safety architecture still let through 1.3 percent of flagged research directions. Much of the process still depends on humans loading materials and refining recipes over multiple rounds. And the gap between strong automated benchmark scores and weak agreement with physician judgment raises doubts about how well these health benchmarks reflect real clinical quality.

“There is a long journey ahead before AI systems can navigate the physical realities of science. But we are deeply excited about the potential of LLMs to help people and accelerate real-world progress.”

— Samuel Schmidgall, lead author of the Co-Scientist study