What is an AI co-scientist?
Multi-agent systems that generate scientific hypotheses, design experiments and propose candidates for lab verification.
An AI co-scientist is a category of multi-agent systems designed to accelerate scientific research. Instead of answering a question, the system generates hypotheses, designs experiments and proposes candidates (molecules, materials, proofs) to be verified experimentally in the lab. Several specialized agents collaborate internally: one scans the literature, one formulates hypotheses, one challenges them, one ranks them by plausibility. The field has moved quickly from announcement to concrete results: systems of this kind have already produced drug-repurposing candidates verified in the lab, helped solve mathematical problems left open for decades, and proposed therapies later confirmed experimentally, yet they remain hypothesis-generation tools to be verified, not oracles that replace the scientific method or empirical validation, a distinction that matters most for anyone evaluating R&D investment and needing to stay vendor-neutral toward labs promoting their own results.
See also: AI Agents (Agentic AI) · AI systems that go beyond answering: they plan, use tools and take actions autonomously inside your processes.
Origin: Google DeepMind, February 2025
The term entered common use with the launch of Google DeepMind's "AI co-scientist" in February 2025: a system built on Gemini that supports researchers in generating original hypotheses and designing experimental protocols, with the stated goal of compressing years of lab work into weeks.
The 2026 evolution
A Turing Post census from May 24, 2026 counts 12 active systems with concrete results, no longer just promises. Google produced drug-repurposing candidates for fibrosis with a 91% block of scarring response in lab tests, and moved an enterprise version into preview with partners such as Daiichi Sankyo, Bayer Crop Science and the US National Labs. OpenAI used a reasoning model to solve a geometry conjecture left unsolved for 80 years, since 1946. FutureHouse, with its "Robin" system, proposed a therapeutic candidate for dry macular degeneration later confirmed in the lab. Axiom Math, with "AxiomProver", solved all 12 problems of one edition of the Putnam exam.
Why it matters for businesses
This is not confined to academia. R&D-intensive companies, in pharma, materials, chemicals, can evaluate these systems as discovery accelerators: they shrink the time between a hypothesis and the first lab test. The caveat to keep in mind is that they remain hypothesis-generation tools, not oracles: every candidate they produce still needs experimental verification before any investment or development decision. The value lies in narrowing the search space, not in replacing the scientific method.
Related terms
- AI Agents (Agentic AI) · AI systems that go beyond answering: they plan, use tools and take actions autonomously inside your processes.
- Reasoning models · Models that reason step by step before answering, spending more compute at inference time on complex problems.
- Foundation model · A model trained on broad, generic data and adaptable to many different tasks: the category LLMs and multimodal models belong to.
- Deep Research · A category of AI agents that autonomously explore dozens of sources and return a long, cited report, instead of a quick answer.
- Google Gemini · Google's family of models and AI assistant, integrated across Workspace, Cloud, Android and Search.