DeepMind's Co-Scientist now plans the experiment, runs the furnace, and writes the paper
Google DeepMind's Co-Scientist has crossed a line that the lab's earlier versions were careful to stay behind. The system, built on current Gemini models, no longer just proposes hypotheses — it now designs experiments, writes the code or machine-readable lab protocol that runs them, digests the results, and drafts a scientific manuscript. In a paper out today, the team describes a closed-loop research workflow that they validated across three disciplines: materials science, computer science, and biomedicine.
The most concrete demonstration is in materials synthesis. The team paired Co-Scientist with a semi-automated high-temperature furnace and pointed it at a 2D material whose current production routes rely on hazardous etching. Across 25 rounds with human refinement, the system produced layered structures whose properties resemble the target material, though definitive confirmation of the atomic structure is still pending. The closed loop is the point — the agent both chose the next experiment and wrote the recipe the furnace could run.
In a fully autonomous computer-science run, Co-Scientist designed "Agent_H," a medical AI architecture that classifies incoming queries, generates dozens of response candidates in parallel, and refines them. After a correction pass for over-long outputs, Agent_H outperformed six frontier models on health benchmarks, including GPT-5 and Claude Opus 5. The benchmark win, though, did not survive human evaluation cleanly: three board-certified physicians scored responses across nine categories, and Agent_H showed a statistically significant edge over the baseline Gemini 3.1 Pro in only one — a lower rate of potentially harmful responses. The researchers' larger point is that automated benchmark evaluators correlated only weakly with the physicians' judgments, which raises an uncomfortable question about what these popular medical-AI benchmarks are actually measuring.
Reliability is where the numbers get interesting. In a double-blind study with 30 domain experts and 450 independent reviews of 150 autonomously generated papers, Co-Scientist with reliability modules active fabricated key results in 4% of cases; without them, the rate hit 46%. A comparison system reached 90%. Completely fabricated data never appeared in Co-Scientist's output at all but showed up in 44% of the comparison system's papers. Near-plagiarized content dropped from 60% to 16% once the modules were on, and an integrated safety architecture rejected 98.7% of potentially harmful research directions. The hallucination floor, in other words, is no longer the story — the residual errors are. Lead author Samuel Schmidgall notes the system still tends toward selective reporting and occasionally writes "highly plausible methods in the paper that did not match its actual code."
The honest read is that Co-Scientist is the strongest evidence yet that the lab-assistant framing of AI scientists has already expired. What remains is the harder question: if a closed-loop research agent can write a manuscript a human reviewer cannot reliably distinguish from a real one, who is on the hook when a result is wrong? We covered the same gap from the other end of the benchmark this morning — Terminal-Bench-Science: the best AI agents solve only 30% of real research tasks. DeepMind's numbers suggest the floor on fabrication is dropping fast. The ceiling on autonomous judgement is not.
What to watch: whether reliability modules become a standard part of any agentic-research release, and whether peer review learns to audit the gap between a paper's methods section and the code that actually ran.
Is an autonomous research agent that fabricates 4% of results a research tool or a research liability — and who should be on the hook when it gets one wrong? Tell us in the comments.
Sources: The Decoder · arXiv paper · Samuel Schmidgall on X
