Artificial intelligence (AI) is moving rapidly into drug development, bringing extraordinary expectations about how quickly models could identify new medicines and unearthing some concerns about giving computers a larger role in decisions that ultimately affect human health. The latest flash point came last week, when Anthropic confirmed that it has established a “wet lab” in the San Francisco Bay Area as a part of its expanded work in the life sciences. The announcement is a significant step in unifying AI and life sciences advancement but also requires some perspective. AI can meaningfully impact the process, but biology still has to prove new medicines work.

What does it mean to use AI in the life sciences? “Artificial intelligence” in drug development can describe a wide range of computational tools. Some perform specialized tasks: analyzing biological datasets, identifying drug targets, predicting protein structures or molecular interactions, or designing prospective compounds. General purpose models can increasingly work across these tasks by searching scientific literature, analyzing experimental data, writing code, proposing hypotheses, and helping researchers decide what experiment to run next.

But the AI assist is accompanied by an important corollary: No matter how sophisticated a computational model becomes, an AI-generated scientific hypothesis, molecule, or drug target remains a prediction. Biology must validate it. And prospective medicines must still ultimately demonstrate to regulators their safety and efficacy.

AI’s promise in the life sciences is not that a model spits out a finished medicine; it rests on potential efficiencies. AI can search for biological possibilities faster, generate better candidates, and shorten the cycle between a scientific question and useful experiment. Anthropic has already demonstrated its model designing protein binders that were subsequently tested experimentally, and it is developing systems that allow AI agents to interact with laboratory equipment.

That is where a “wet lab” comes in. Although it may conjure up potentially alarming images, a wet lab is simply a facility where physical experimentation is performed (think – a lab bench), as opposed to work performed entirely on a computer. Researchers use cells, proteins, compounds, reagents, and other biological materials to assess whether a computational prediction holds up. An AI-supported facility can combine experimentation with AI and laboratory automation, creating a feedback loop in which a hypothesis is tested, and the resulting data used to inform what happens next.

That feedback loop could meaningfully accelerate biomedical research. But for all the concern about AI-developed drugs, the wet lab is only the first checkpoint. A molecule that behaves exactly as predicted in a lab-based cellular experiment may not yet have shown it will successfully treat a human disease. Preclinical research probes toxicity, pharmacology, dosing, and other characteristics – but even favorable results cannot establish how a medicine will perform in patients.

That is why clinical trials will remain essential even if AI transforms much of what comes before them. Phase 1 trials principally investigate safety and dosage. Phase 2 studies begin examining efficacy while gathering more safety information. Large Phase 3 trials test whether a product provides a treatment benefit in its intended population and identifies adverse reactions smaller studies may miss. Each stage puts earlier conclusions to a harder test.

Early evidence from AI-discovered medicines illustrates how the clinical trial process provides its check on AI adoption. A 2024 analysis of AI-native biotechnology companies found Phase 1 success rates for AI-discovered molecules of roughly 80–90 percent, substantially above historical industry averages. Yet the observed Phase 2 success rate was approximately 40 percent, broadly comparable with historical experience, though the sample remains small. AI may prove very good at identifying molecules with desirable drug-like properties. The harder question is whether the identified biological mechanism makes a sick patient better: A computationally elegant molecule can still be a dud.

There is a big difference between accelerating the acquisition of evidence and eliminating the need to acquire it. Fortunately, the existing drug-approval framework is already built around that distinction. The Food and Drug Administration’s basic questions do not depend on whether a candidate was discovered through painstaking laboratory work, serendipity, conventional computational methods, or artificial intelligence. Does it work for its intended use? What risks does it present? Is the evidence rigorous enough to support those conclusions? Can it be manufactured consistently? AI does not make those questions obsolete.

There is also no particular regulatory turpitude at this point in using AI to reach a promising candidate more quickly. The existing system works reasonably well because it is largely technology-neutral: Innovation can change how researchers arrive at a prospective treatment without changing the obligation to demonstrate what that treatment does in patients. If AI can identify better targets, screen compounds faster, eliminate unproductive experiments, or improve clinical development, it should make portions of the process faster and more efficient.

What should remain intact is the evidentiary backstop. Drug development shouldn’t be subjected to unnecessary friction when new technologies make certain steps in the process obsolete. But clinical trials establishing safety and efficacy aren’t friction; they’re validation. They determine whether all the promising computational and laboratory work that came before has been translated into medicine that benefits patients. AI may change how quickly researchers reach that question, but it does not change the need to answer it.

Anthropic’s wet lab can potentially show how powerful AI tools can fit within a scientific process that remains fundamentally empirical. AI can generate a prediction. A wet lab tests its biology. Preclinical development probes further. But when that prediction becomes a medicine intended for people, clinical trials still provide the test that matters most.