AI scientists are entering the experimental loop

From reading science to taking part in it

Most AI tools used in research begin with the literature. They search papers, summarise findings and help scientists navigate more material than any person could read alone.

Two systems published together in the 9 July issue of Nature attempt something more ambitious. They do not stop after proposing an idea. They connect hypothesis generation to experiments, analyse the resulting data and use what they find to shape the next question.

Google’s system is called Co-Scientist. FutureHouse’s is called Robin. Their designs differ, but both use teams of specialised AI agents rather than asking one model to do everything. One agent may search the literature, another may challenge a hypothesis and another may analyse raw biological data. A coordinating system keeps the work moving.

The important word is “loop”. Science advances when an idea meets an experiment and the result changes what comes next. Bringing AI into that cycle is more consequential than producing a plausible research proposal in isolation.

What Co-Scientist did

Co-Scientist is built on Google’s Gemini 2.0 models. A human scientist provides a research goal, evidence and practical constraints. Specialised agents then generate, criticise, rank and refine hypotheses. The system is designed to favour ideas that are plausible, novel, testable and aligned with the stated goal.

In one application, researchers used it to look for treatments that might work against acute myeloid leukaemia. The system proposed drug-repurposing candidates and combinations. Scientists then tested selected suggestions in vitro — experiments conducted on cells outside a living organism — and reported that some of the proposed combinations showed the expected activity.

That is more persuasive than a model judging its own prose. The claim was exposed to a physical experiment. It is still an early stage: success in cells does not establish that a treatment is safe or effective in people.

What Robin did

Robin focused on dry age-related macular degeneration, a major cause of irreversible sight loss. It coordinated literature-search agents called Crow and Falcon with a data-analysis agent called Finch.

The system proposed disease mechanisms, experimental assays and candidate drugs. Human researchers performed the laboratory work and returned the data. Robin analysed those results and generated a revised set of therapeutic candidates, creating an iterative cycle between literature, experiment and interpretation.

The authors report that Robin analysed 551 papers in about 30 minutes, compared with an estimated 294 hours of human reading, and calculate a roughly 200-fold reduction in time for the complete workflow. Those figures are estimates produced by the study team. They measure time spent on defined tasks, not the quality or speed of an entire drug-development programme.

Why several agents can help

Scientific work contains different kinds of reasoning. Finding a relevant paper is not the same as deciding whether a hypothesis is novel. Writing code to analyse gene-expression data is not the same as choosing an experiment that a particular laboratory can actually perform.

Dividing those responsibilities can make each part easier to inspect. It also creates new failure points. An incorrect paper, a fabricated citation or a mistake in data analysis can travel through the system and make a later conclusion appear more solid than it is.

Both papers therefore matter as demonstrations of structured collaboration, not as evidence that an AI can replace the scientific community. Co-Scientist allows researchers to steer the goal and constraints throughout the process. Robin’s experimental work was still performed by people. Expertise did not disappear; it moved into supervision, experimental design and judgment.

What is genuinely new

AI-generated hypotheses are no longer unusual. The stronger evidence here is that ideas were connected to laboratory tests and then revised in response to the results. That closes part of the loop that earlier systems left open.

It also offers a more realistic picture of near-term scientific automation. The laboratory is not suddenly empty. Instead, software can search and compare at a speed that changes which questions a team can afford to explore. Scientists can spend more time deciding what deserves belief, what is safe to test and what a result actually means.

What remains uncertain

Both systems were evaluated on selected biomedical problems with teams closely involved in their development. Wider tests are needed across independent laboratories, less familiar fields and cases where the available literature is sparse or contradictory.

The route from a cell experiment to a useful medicine remains long. Toxicity, dosage, manufacturing, clinical trials and regulatory review are not solved by generating a promising hypothesis. A fast research loop can also produce a larger volume of weak ideas if its checks are poor.

The papers nevertheless mark a useful threshold. AI is beginning to participate in the part of science where ideas meet evidence. Its value will depend on whether the resulting work remains reproducible, contestable and under the control of people who understand both the tools and the biology.