AI That Discovers

In 2026, two systems brought an important development to the attention of the research community: AI is not limited to summarising published literature; it can also help formulate testable hypotheses and organise parts of an experimental workflow. That does not amount to autonomous discovery. Experiments, their interpretation and independent verification remain central.

For years, the most visible use of language models in research was finding and reorganising information that had already been published. Multi-agent systems add another possibility: agents can search evidence, propose explanations, challenge hypotheses and suggest experiments. Their contribution should be judged by verification, not by the confidence of their answers.

Google describes Co-Scientist as a Gemini-based multi-agent system designed to work with researchers on generating and evaluating hypotheses. In 2026, the work was published in *Nature*. It is a contribution to scientist-led research, not a protocol that replaces peer review or experimentation.

FutureHouse introduced Robin, a system connecting literature search, hypothesis generation, experimental design and analysis. In a study of dry age-related macular degeneration, the team tested drug candidates in vitro and identified ripasudil as a possible lead for further investigation. The result is interesting precisely because the work does not end in the model: it proceeds through cell cultures, laboratory analysis and expert assessment.

A system may produce more hypotheses than a laboratory can test. The constraint then becomes the ability to decide which ideas deserve time, samples and resources. Automated analysis can also be wrong: agent-generated figures must therefore be checked against the original data and, where possible, replicated.

The published work describes candidates, experimental procedures and analysis; it should not be read as a controlled comparison demonstrating general superiority by either the agent or researchers. The stronger methodological lesson is that conclusions must remain proportionate to the data and validation described in the paper.

Speed in generating hypotheses is not enough to create an advantage. Laboratories, biobanks, high-quality data, statistical expertise and procedures for reproducing results are also needed. For universities, pharmaceutical research, cosmetics and materials science, AI may reduce time spent surveying literature and formulating leads; experimental capacity determines how many of them become evidence.

This distinction also has an organisational consequence. As the volume of hypotheses grows, researchers who can design controls, recognise fragile results and decide what should not proceed become more valuable. Productivity is not the same as the number of ideas generated.

When a system makes hypotheses cheap to produce, those who can verify them become more central. Does Europe want to compete only on models that propose ideas, or also on the infrastructure that turns those ideas into reproducible knowledge?

This paper distinguishes findings reported by the cited studies from the strategic implications proposed here. These are examples of AI-supported research; they do not justify generalisation across all disciplines or every stage of discovery.