Google's AI Co-Scientist: A Multi-Agent System to Accelerate Hypothesis Generation

Google unveiled AI Co-Scientist on February 19, 2025: a multi-agent system built on Gemini 2.0, positioned as a "virtual scientific collaborator" that goes beyond literature review to generate novel research hypotheses, proposals and experimental protocols (Google Research blog, 2025-02-19).

Six agents and a self-improving loop#

Per the official blog, the system uses a coalition of specialised agents — Generation, Reflection, Ranking, Evolution, Proximity and Meta-review — with a Supervisor assigning tasks and compute, then using automated feedback to iteratively generate, evaluate and refine hypotheses.

The blog groups the key mechanics into three: self-play-based "scientific debate" for hypothesis generation, ranking tournaments for comparing hypotheses, and an "evolution" process for quality improvement. Hypotheses are auto-scored with Elo from those tournaments; the blog notes the Elo is an auto-evaluation and not an independent ground truth.

End-to-end validation on three biomedical tasks#

Google ran three end-to-end validations with outside labs:

• Drug repurposing for acute myeloid leukaemia (AML): the repurposing candidates the system proposed were validated in vitro, inhibiting tumour viability at clinically relevant concentrations in multiple AML cell lines.
• Target discovery for liver fibrosis: the system proposed and ranked epigenetic targets grounded in preclinical evidence, which showed anti-fibrotic activity in human hepatic organoids; the blog states p-values for all suggested drugs were below 0.01.
• Mechanisms of antimicrobial resistance: asked to explain how cf-PICIs exist across multiple bacterial species, the system independently proposed that cf-PICIs interact with diverse phage tails to expand their host range. That finding had already been validated experimentally by the lab but not yet made public, and appeared in co-timed papers with collaborators including Imperial College London.

The paper and the limits#

The work first appeared as a preprint in February 2025 and was revised to v2 on June 29, 2026, with the arXiv listing marked Nature (2026) (arXiv:2502.18864, 10.1038/s41586-026-10644-y).

The limits are in the same blog post: better literature review, factuality checking, cross-checks with external tools, auto-evaluation, and larger-scale evaluation with more subject-matter experts across more research goals. In other words, every "AI-proposed hypothesis" still goes through expert review and case-by-case experimental validation.