Google DeepMind: Co-Scientist multi-agent AI partner for scientific research
Google DeepMind announced Co-Scientist on 19 May 2026 — a Gemini-based multi-agent AI system that generates, debates, and refines scientific hypotheses using 6 specialised agents in a Tournament of Ideas debate. The system was developed in collaboration with more than 100 research institutions and has already produced concrete results in liver fibrosis, ALS, cellular ageing, and infectious disease research, with analysis time reduced from months to days.
This article was generated using artificial intelligence from primary sources.
On 19 May 2026, Google DeepMind announced Co-Scientist — a Gemini-based multi-agent AI system that generates, debates, and refines scientific hypotheses through the interaction of 6 specialised agents. The system was developed in collaboration with more than 100 research institutions and marks the first time DeepMind has opened its reasoning stack for biomedical discovery at a global scale.
What is the architecture of Co-Scientist?
Co-Scientist consists of 6 specialised agents: generation (proposes hypotheses), proximity (measures similarity to existing literature), reflection (critical reassessment), ranking (candidate comparison), evolution (refinement of top candidates), and meta-review (final validation before a hypothesis is delivered to the researcher). The agents communicate through an internal method called the Tournament of Ideas — hypotheses are systematically pitted against each other in pairs until a consensus winner emerges, providing the researcher with a ranked list of questions worth further investigation.
What are the concrete results in biomedicine?
The most publicised result concerns liver fibrosis: Co-Scientist proposed molecular candidates that in in-vitro tests at partner laboratories blocked 91 percent of signalling pathways associated with tissue scarring. The system has also been used to research ALS (amyotrophic lateral sclerosis), cellular senescence, and infectious diseases, where time-series dataset analysis has been reduced from months to days.
How is the system secured for bio-research?
Bio-research carries specific security risks (dual-use, CBRN potential). Co-Scientist underwent an independent CBRN safety assessment before public launch. The system is not an open API — access is mediated through the Hypothesis Generation tool within Google’s Gemini for Science programme, and institutions must pass verification before gaining access.
DeepMind positions the system as a multiplier rather than a replacement for the researcher: the scientist defines the question and evaluates the final candidates, while Co-Scientist handles the difficult, hard-to-scale work of searching the hypothesis space. This is a different approach from OpenAI Deep Research or Anthropic Research mode, which operate at the level of synthetic reports, whereas Co-Scientist generates falsifiable claims ready for laboratory validation.
Frequently Asked Questions
- What is the architecture of Co-Scientist?
- Co-Scientist consists of 6 specialised agents — generation (proposes hypotheses), proximity (measures similarity to literature), reflection (critique), ranking (comparison), evolution (refinement), and meta-review (validation). Tournament of Ideas is the internal method in which hypotheses face each other in pairs until a consensus winner emerges.
- What was specifically discovered in the liver fibrosis research?
- Co-Scientist proposed candidate compounds that in in-vitro tests blocked 91 percent of liver scarring-related signalling pathways. The result was validated in partner institution laboratories before the system was introduced publicly.
- How is the system secured for bio-research?
- Co-Scientist underwent an independent CBRN (chemical, biological, radiological, nuclear) safety assessment before public launch. The system is not available as an open API — access is mediated through the Hypothesis Generation tool in the Gemini for Science programme and requires institutional verification.