AI as Research Partner
AI systems generate hypotheses, plan experiments and write papers; in 2026 Nature published studies on Google's Co-Scientist and the AI Scientist.
Open in the interactive tree →Google's Co-Scientist (built on Gemini, announced February 2025) has several AI agents propose, debate and rank hypotheses; lab tests supported, among others, drug repurposing candidates for acute myeloid leukemia. The 'AI Scientist' from Sakana AI, UBC, the Vector Institute and Oxford wrote a paper that passed review at an ICLR workshop. FutureHouse (Robin) and Edison Scientific (Kosmos) build similar systems.
As of October 2026
Nature published the AI Scientist study on 27 March 2026 and the Co-Scientist study on 19 May 2026. Co-Scientist is designed to work under scientists' oversight, while the AI Scientist so far covers computer science only, and its authors list underdeveloped ideas and inaccurate citations as limits. Whether such systems produce reliably new knowledge at scale is still being tested.
Open steps
- Hypothesis extraction from literature High AI leverageDeploy LLM agents to systematically mine papers for testable hypotheses that current labs could test.
- Autonomous multi-day lab workflows Medium AI leverageIntegrate robots, incubators, readers and sequencers into autonomous-agent workflows that execute multi-day experiments without human intervention.
- Automated reproducibility Medium AI leverageBuild agents that parse methods sections and run equivalent experiments in robotic labs to verify and identify reproducible findings.
Where AI could help
Medium AI leverage. Systems already generate testable hypotheses, but experiments, review and reliable citations decide whether the output is new knowledge.
- Multi-agent hypothesis generation, critique and ranking tied to lab validation
- Automated literature synthesis with citation checking
- Closed-loop experiment planning in self-driving labs
Shown so far
- In May 2026 Google Research said its Co-Scientist multi-agent system had been published in Nature and cited use in research on antimicrobial resistance, plant immunity and liver fibrosis, according to Google. source
- In November 2025 a preprint from Edison Scientific described its Kosmos AI scientist, which reads about 1,500 papers per run, with independent scientists rating 79.4% of report statements accurate, according to the company. source
Prerequisites
- Scientific Method~1600-1620
- AI protein structure prediction2020
- Large Language Models2022