Google DeepMind has expanded its multi-agent AI Co-Scientist system from a hypothesis generator into a lab-integrated research partner that plans experiments, writes code, controls laboratory equipment and generates scientific manuscripts, according to a report by The Decoder published August 28. The system, built on Google's current Gemini models, has delivered experimentally validated results across three disciplines — materials science, biology and computer science — according to a paper by lead author Samuel Schmidgall and colleagues.

First introduced in February 2025 on Gemini 2.0 with known weaknesses in fact-checking and literature review, the expanded Co-Scientist now runs what researchers call a closed-loop research workflow. It derives hypotheses from a research question, creates experimental plans or machine-readable lab protocols, executes them, analyzes the results and writes them up — while separate verification modules cross-check every numerical claim in the text against the execution logs of the code it generated, a direct attack on the fabricated-results problem that has plagued autonomous research agents.

The work is the latest milestone in latest AI developments around autonomous research agents, and one of the first to couple an LLM-based system to physical lab hardware at this level.

From hypotheses to wet lab protocols

In materials science, DeepMind paired Co-Scientist with a semi-automated high-temperature furnace. The system found a safer synthesis pathway for a sought-after 2D material previously produced mainly through hazardous etching, and generated complete growth recipes tailored to the specific furnace it was working with. After 25 rounds of iteration with human refinement, the team produced layered structures whose properties resemble the target material — though definitive confirmation of the atomic structure is still pending.

In a second experiment, three semiconductor thin films were synthesized successfully on the first attempt. Using Gemini 3 Deep Think for direct equipment control, Co-Scientist cut recipe development time from days to minutes. Humans still had to load samples and precursor materials manually, and the fast mode produced smaller, less uniform crystals than carefully optimized recipes — a reminder that the loop is not yet fully hands-off.

In biology, the system autonomously built an image-analysis pipeline that predicts which patterns genetically engineered E. coli colonies form at different chemical concentrations. Predictions generated with Gemini 3 Pro Image matched unpublished lab results on three of four shape features. The researchers note the system only reasons between known conditions and cannot yet predict behavior in entirely new experimental systems.

An autonomous AI that designs another AI

The computer science experiment ran without any 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 health benchmarks.

Then came the reality check. Three board-certified physicians scored the responses in a blinded comparison across nine categories, and Agent_H showed a statistically significant advantage over the baseline Gemini 3.1 Pro in just one: a lower risk of potentially harmful responses. The automated benchmark evaluators correlated only weakly with the physicians' judgments.

That gap between benchmark scores and clinical usefulness is arguably the study's most consequential finding. High automated scores, the researchers suggest, do not necessarily mean a system delivers better answers from a medical perspective — raising uncomfortable questions about what AI benchmarks actually measure.

Cutting fabrication from 46% to 4%

The core failure mode of autonomous research agents is fabrication: when an agent is rewarded for producing good results, it has an incentive to make them up. Previous analyses documented fabrication rates of 80 to 100 percent in existing systems.

Co-Scientist attacks the problem with two mechanisms. The system is penalized for fabricated or plagiarized content, and a separate verification module cross-checks every numerical claim against the actual results of its executed code. In a double-blind study involving 30 domain experts and 450 independent reviews of 150 autonomously generated papers, the results were stark:

  • 4% of papers fabricated key results with reliability modules active, versus 46% without them.
  • The comparison system reached 90% key-result fabrication.
  • Completely fabricated data never appeared in Co-Scientist's output, but showed up in 44% of the comparison system's papers.
  • Near-plagiarized content dropped from 60% to 16%.
  • An integrated safety architecture rejected 98.7% of potentially harmful research directions.

Not ready to replace scientists

Residual errors remain. The system tends toward selective reporting, and Schmidgall acknowledges it writes "highly plausible methods in the paper that did not match its actual code." Whether the materials-science recipes transfer to other labs is an open question, and the biology results, while promising, covered only a narrow class of predictions.

Still, the trajectory is clear. A system that began as a hypothesis generator eighteen months ago can now plan an experiment, operate a furnace, analyze the output and draft a manuscript — and document, with log-level verification, when its own claims are wrong. The gap between lab assistant and autonomous researcher is narrowing, and the parts of science that remain stubbornly human — judgment, taste and physical sample loading — are becoming easier to see precisely because the rest is being automated.

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →