AI Runs Its Own Experiments to Generate Science Hypotheses
The Question Left Overnight
A researcher types a question into a text box, presses Enter, and goes home for the evening. [1] The tool she prompted is built to generate scientific hypotheses, the kind of educated guesses that normally take a trained researcher months of reading and thinking to assemble. It is called Co-Scientist. [1] Unlike the chatbots most people know, it does not answer in one breath. It dispatches a set of independent software workers, called agents, to comb through published papers, weigh rival explanations against each other, sharpen and attack candidate hypotheses, and then test those ideas against the evidence already in print. [1] The overnight gap matters because the system keeps working while the researcher sleeps.
A Tool That Argues With Itself
The design contradicts a common assumption about what an AI does. A standard chatbot produces one answer and stops. Co-Scientist does not stop. Each agent has a job, and the jobs are not the same: some gather, some synthesize, some evaluate competing explanations, some refine and critique, some test ideas against what has already been published. The system loops. It revisits its own output. It behaves less like a search engine and more like a small research group that never adjourns. The field has long treated hypothesis generation as the irreducibly human step, the creative leap no machine could make. Co-Scientist is aimed squarely at that leap. The agents do not merely propose; they evaluate, critique, and iterate, so the system is built to disagree with itself before it ever disagrees with a scientist.
Sources
1. Nature (DOI: 10.1038/d41586-026-02931-5) — Quote source (original article)