Google DeepMind's Co-Scientist now runs lab equipment and drafts its own papers
DeepMind extended Co-Scientist from a hypothesis generator into a closed loop that plans experiments, writes code, drives a high-temperature furnace, and writes manuscripts — synthesizing three semiconductor thin films on the first try and cutting recipe development from days to minutes with Gemini 3 Deep Think. Reliability modules that cross-check every numerical claim against execution logs dropped fabricated results from 46% to 4% across 150 papers reviewed by 30 experts, the first credible answer to fabrication rates of 80-100% in prior autonomous research systems. The autonomy claim is thinner than the benchmarks suggest: Agent_H, the medical architecture Co-Scientist designed unaided, beat GPT-5 and Claude Opus 5 on health benchmarks but showed a significant edge over baseline Gemini in only one of nine categories scored by three physicians.
Source: the-decoder.com ↗
With reliability modules active, Co-Scientist fabricated key results in 4 percent of cases. Without them, the rate hit 46 percent.
Why this matters
- → Closed-loop lab autonomy cuts materials synthesis from days to minutes
- → Reliability modules slash fabrication rates from 46% to 4% in generated papers
- → Benchmark scores diverge sharply from physician evaluation on clinical tasks