Large models have so far worked mainly on representations: scientific papers, sequences, images, databases and simulations. A wet lab introduces an essential difference. It is where reagents, cells, instruments and protocols produce physical results that can confirm or reject a hypothesis. Anthropic’s decision to establish a biology laboratory in the Bay Area takes AI beyond the computer and into the experimental cycle.
The company has already developed Claude Science and acquired expertise in computational biology. The new laboratory is not presented as a pharmaceutical company intended to run clinical trials directly, but as a space for integrating models, automation and experiments. Its stated interests include rare diseases, difficult targets and problems on which traditional methods progress slowly.
The ideal loop is easy to describe: a model reads literature and data, proposes a molecule or experiment, the laboratory executes it, instruments measure the outcome and results return to the system. Each iteration may improve the next. What once required weeks of coordination among software, scientists and equipment can become a more continuous and documented process.
Science does not progress by optimising a score alone. An unexpected result may be a discovery, contamination or a procedural mistake. Distinguishing among them requires material experience: observing a sample, knowing an instrument’s limits and recognising when a formally correct procedure did not truly work. A model can organise hypotheses; the laboratory remains a place of judgement.
The ability to test alternatives quickly may expand research on conditions neglected by the market. If initial costs fall, it becomes more realistic to explore targets with few patients or uncertain commercial returns. Yet democratisation will depend on access to data, instruments and results. A proprietary laboratory can accelerate discovery while also concentrating knowledge and infrastructure.
Biosecurity is unavoidable. A system able to propose useful biological protocols or structures may also lower barriers to dangerous uses. Refusal by a chatbot is not enough. Controls on instruments, experiment review, synthesis limits, material traceability and separation between digital design and sensitive physical actions are all required.
Reproducibility changes as well. If a result emerges from many decisions made by agents, software and laboratory robots, a paper must document not only the final protocol but also model versions, parameters and the criteria used to select trials. Without this record, automation may create science that is fast but difficult to verify independently.
The researcher does not disappear. The role moves toward problem formulation, control quality, error evaluation and the choice of what deserves to become matter. A model can propose thousands of possibilities; a laboratory cannot execute all of them. The crucial expertise remains designing experiments that genuinely distinguish among competing hypotheses.
AI opens a real laboratory when it accepts the resistance of the physical world. A textual answer can appear plausible; a molecule must work, a protocol must be reproducible and a result must survive verification. That friction is what makes the transition promising. The machine does not become a scientist alone: it enters a system in which matter can finally tell it no.