Google DeepMind has expanded its AI Co-Scientist system from a hypothesis generator into a research system integrated into the lab. The Gemini-based multi-agent system has delivered experimentally validated results across disciplines like materials synthesis and medical AI architecture development.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
Google DeepMind's Co-Scientist system has expanded from a pure hypothesis generator to a closed-loop research partner capable of planning experiments, controlling equipment, and writing papers. Despite demonstrating potential across three disciplines, human evaluations reveal benchmark limitations and high hallucination rates that remain prominent.
In August 2026, Google DeepMind announced the expansion of its multi-agent system Co-Scientist, transforming it from an initial hypothesis generator into an integrated laboratory research partner. According to Google, the system, now based on current Gemini models, can plan experiments, write code, control laboratory equipment, and generate scientific manuscripts, forming a closed-loop workflow: deriving hypotheses from research questions, creating experimental plans or machine-readable protocols, analyzing results, and finally writing papers.
This expansion marks a significant advancement from its initial release in February 2025 (based on Gemini 2.0), when the system had shortcomings in fact-checking and literature review. Now, a validation module cross-checks numerical claims in text against execution logs to reduce fabricated results.
The system was validated in three disciplines with increasing autonomy: in materials science, it designed synthesis recipes executed by humans; in biology, it built predictive pipelines with expert feedback; and in computer science, it ran fully autonomously.
This report is based on second-hand reporting from THE DECODER, which sourced information from Google DeepMind's official statements and a preprint paper (Schmidgall, Zhu et al. 2026). All specific data (such as fabrication rates, benchmark performance) come from that report and have not been independently verified; they should be treated as source claims rather than confirmed facts.
Primary report
Primary source