Google DeepMind has quietly turned its Co-Scientist system from a hypothesis-spitting research assistant into something closer to a full lab partner — the kind that writes its own code, runs its own equipment, and drafts its own papers. The expansion, built on current Gemini models, introduces a closed-loop workflow: Co-Scientist takes a research question, generates hypotheses, creates experimental protocols or code, executes them, analyzes the output, and produces a manuscript. That’s a meaningful jump from the February 2025 version, which struggled with fact-checking and literature review.
The system was tested across three disciplines with escalating autonomy. In materials science, Co-Scientist paired with a semi-automated furnace to find a safer synthesis route for a 2D material previously made through hazardous etching. After 25 rounds of human refinement, the team got layered structures with promising properties — though atomic-level confirmation is still pending. Three semiconductor thin films were synthesized on the first attempt, with Gemini 3 Deep Think handling direct equipment control and slashing recipe development from days to minutes. Lead author Samuel Schmidgall notes the fast mode produced smaller, less uniform crystals, and whether recipes transfer to other labs remains open.
In biology, the system built an image analysis pipeline predicting E. coli colony patterns, matching unpublished results for three of four shape features. A computer science experiment ran with zero human involvement beyond setup: Co-Scientist designed “Agent_H,” a medical AI architecture that outperformed six frontier models including GPT-5 and Claude Opus 5 on health benchmarks. But board-certified physicians found a statistically significant advantage over baseline Gemini 3.1 Pro in just one of nine categories — lower risk of harmful responses. The benchmark-to-human-evaluation gap is stark, and the researchers themselves admit automated evaluators correlated only weakly with clinical judgment.
The most striking number is the fabrication rate. Previous analyses put hallucination rates in autonomous research systems at 80 to 100 percent. Co-Scientist’s reliability modules — which penalize fabrication and cross-check every numerical claim against execution logs — brought that down to 4 percent in a double-blind study with 30 domain experts and 450 independent reviews. Without the modules, the rate hit 46 percent. Completely fabricated data never appeared in Co-Scientist’s output but showed up in 44 percent of comparison system papers. The safety architecture rejected 98.7 percent of potentially harmful research directions. Still, selective reporting and other leftover errors persist, and the material science results are promising but not yet definitive — a fair reminder that ‘autonomous’ doesn’t mean ‘trustworthy.’
