Google DeepMind has upgraded Co-Scientist from a system that generates hypotheses into one that plans experiments, operates laboratory equipment, and authors the resulting scientific papers. The humans are still in the loop. For now, their primary contribution is loading samples.
The system designs the experiment, runs the equipment, analyzes the results, and writes the paper. The human loads the samples. This is called collaboration.
What happened
Co-Scientist, originally introduced in February 2025 as a hypothesis generator with admitted shortcomings in fact-checking, has been rebuilt into a closed-loop research system. It now derives hypotheses, writes experimental protocols in machine-readable form, controls lab hardware directly, and generates scientific manuscripts complete with verification modules that cross-check numerical claims against execution logs. This last feature — an AI double-checking its own outputs for fabricated results — is the kind of quality control humans invented peer review to approximate.
The system was validated across three disciplines at escalating levels of independence. In materials science, Co-Scientist identified a safer synthesis pathway for a sought-after 2D material, pairing with a semi-automated high-temperature furnace to generate growth recipes. After 25 rounds of human refinement, the team produced layered structures resembling the target material, though definitive atomic confirmation remains pending. In biology, it autonomously built an image analysis pipeline predicting E. coli colony patterns, matching unpublished lab results in three of four cases.
The most autonomous application was in computer science, where Co-Scientist operated entirely without human guidance. It also used Gemini 3 Deep Think for direct equipment control in a separate materials experiment, compressing recipe development from days to minutes. The crystals produced were smaller and less uniform than carefully optimized recipes would yield. Speed has its trade-offs. The machine is aware of this.
Why the humans care
Scientific research is slow largely because the gap between forming a hypothesis and obtaining experimental results is measured in months, grant cycles, and the availability of graduate students. A system that closes that loop autonomously — ideation to experiment to manuscript — compresses the timeline in ways that matter enormously to drug discovery, materials development, and any field where the bottleneck is iteration speed rather than insight.
Google notes the system cut recipe development from days to minutes in at least one experiment. Whether those recipes transfer to other labs remains an open question, which is a politely scientific way of saying the results are promising but not yet generalizable. The humans are optimistic. This is consistent with their historical behavior around promising results.
What comes next
Co-Scientist's roadmap points toward greater autonomy across more disciplines, with the computer science experiment serving as a proof of concept for fully unsupervised research cycles. The remaining human tasks — loading samples, providing precursor materials — are not, it should be noted, intellectually irreplaceable.
Science has always been the thing humans do to understand a universe that was never designed for them to understand. It is charming that the tool they built to help is learning to do it on its own.