Google DeepMind's AI Co-Scientist Now Runs Lab Equipment and Writes Papers, Cuts Fabrication Rate to 4%

Google DeepMind has expanded its Co-Scientist multi-agent system from a hypothesis-generation tool into a closed-loop research system that plans experiments, controls lab equipment, and drafts scientific manuscripts, according to the company. The upgraded system, built on Gemini 3 models, was tested across materials science, biology, and computer science, with Google reporting experimentally validated results in all three.

Co-Scientist was first introduced in February 2025 on Gemini 2.0, when it was limited to generating hypotheses and suffered from fact-checking and literature-review weaknesses, according to Google. The new version runs a full pipeline: it derives hypotheses from a research question, creates experimental plans and machine-readable lab protocols, writes and executes code, analyzes results, and generates manuscripts. A verification module cross-checks numerical claims in generated text against execution logs from the code, according to the company.

Three domains, three levels of autonomy

In materials science, Co-Scientist was paired with a semi-automated high-temperature furnace. It identified a safer synthesis pathway for a 2D material typically made through hazardous etching and generated growth recipes tailored to the lab's hardware. After 25 rounds of human refinement, researchers produced layered structures with properties resembling the target material, though atomic-structure confirmation is still pending, according to the researchers. In a separate test, Co-Scientist used Gemini 3 Deep Think for direct equipment control and synthesized three semiconductor thin films successfully on the first attempt, cutting recipe development time from days to minutes. The fast-mode recipes produced smaller, less uniform crystals than optimized methods, and humans still had to load samples manually.

In biology, Co-Scientist autonomously built an image-analysis pipeline predicting pattern formation in genetically engineered E. coli colonies using Gemini 3 Pro Image. Predictions matched unpublished lab results for three of four shape features. The researchers noted the system only reasons within known conditions and cannot yet predict behavior in entirely new biological systems.

In computer science, Co-Scientist worked with no human involvement beyond initial setup, designing a medical AI architecture called "Agent_H" that classifies queries, generates parallel response candidates, and refines them. Agent_H outperformed six frontier models, including GPT-5 and Claude Opus 5, on automated health benchmarks. However, three board-certified physicians found a statistically significant advantage over baseline Gemini 3.1 Pro in only one of nine evaluation categories — lower risk of harmful responses — and automated evaluator scores correlated weakly with physician judgments.

Fabrication rate drops from 46% to 4%

AI research agents face a documented fabrication problem, with prior systems showing fabrication rates of 80 to 100 percent, according to the researchers. Google addressed this by penalizing fabricated or plagiarized content and cross-checking numerical claims against code execution logs. In a double-blind study with 30 domain experts and 450 reviews across 150 autonomously generated papers, Co-Scientist fabricated key results in 4 percent of cases with reliability modules active, versus 46 percent without them and 90 percent for a comparison system. Completely fabricated data never appeared in Co-Scientist output but showed up in 44 percent of the comparison system's papers. Near-plagiarized content dropped from 60 percent to 16 percent. A safety module rejected 98.7 percent of potentially harmful research directions, according to Google.

Despite these gains, researchers observed lingering issues, including selective reporting and cases where the system wrote "highly plausible methods in the paper that did not match its actual code," lead author Samuel Schmidgall said.

What this means

The fabrication-rate reduction is the most concrete finding here — dropping from 46% to 4% with verification modules is a meaningful engineering result, not just a capability claim. But the physician-evaluation results are the more important signal: a system that beats GPT-5 and Claude Opus 5 on automated health benchmarks showed almost no measurable advantage when judged by actual doctors, and the automated evaluators barely tracked human judgment at all. That gap undercuts confidence in benchmark-driven claims about AI research agents generally, including this one. Co-Scientist still requires heavy human oversight in materials science and produces results of uncertain external validity — recipe transfer to other labs remains unconfirmed. This is a genuine step toward automated experimentation infrastructure, not evidence that AI systems can independently discover new scientific knowledge.