AI Can Invent Drugs Faster. It Still Has to Prove They Work.

Recently, Anthropic unveiled an artificial intelligence model with a tantalizing promise: tools that could help researchers design drugs 10 times faster than current methods. If that pace tracked from idea to pharmacy, it could shrink a 10-15-year process to little more than a year — bringing cures dramatically sooner to patients who cannot wait.

Although Anthropic has since temporarily withdrawn that model amid a dispute with the federal government, the regulatory problem exposed by this breakthrough isn’t going away.

AI may make the first and scientifically hardest task — finding a promising molecule — much easier. But it does little to shorten the second, often more consequential task: proving that the medicine works in real patients, at the right dose, with risks understood enough to justify its use.

That leaves the FDA with two urgent challenges. It must evaluate a growing wave of drug candidates without becoming a bottleneck, while resisting pressure to let faster discovery justify weaker evidence. Sponsors will likely push for shorter or smaller trials — and, in some cases, no traditional efficacy trial at all. The question is how the FDA can modernize approval without shifting unresolved risk onto patients or allowing speed to substitute for proof.

After a promising drug candidate is identified, the current path to market can take many years. Most drugs entering human testing fail. Trials run for years because bodies are not computer models. Patients differ by age, genetics, illness, diet, prior treatment, and other factors we do not fully understand. A drug that looks elegant on a screen may miss its target, work only in a narrowly defined group, reveal a dangerous side effect, or prove impossible to manufacture reliably. FDA approval is not a prize for ingenuity. It is a public judgment of safety and efficacy.

Once AI can find promising molecules in a fraction of the time, the testing and approvals that follow become the rate-limiting steps in the process of getting treatments to patients. 

We approach this problem from different perspectives as co-authors. One of us has spent decades developing biological drugs and advising U.S. and European regulators on how approval pathways can better reflect modern science. The other studies the legal and regulatory systems that determine when medical innovation reaches patients, and on what proof. 

We agree that the current AI moment should not be framed as a choice between faster approval and patient protection. The better question is whether FDA can make approval smarter — faster when the science supports it, and just as demanding when patient safety depends on it.

Biosimilars, near copies of complex biologic medicines, show what a smarter approval process can look like. Modern analytical tools can often compare a proposed biosimilar with the original product in exquisite detail. In some cases, a patient trial adds little beyond what chemistry, structure, function, and pharmacology have already shown. In that field, one of us has argued — and FDA has increasingly recognized — that additional human efficacy studies may be unnecessary. The lesson is not that trials are unimportant. It is that FDA must, particularly in this new AI era, align evidentiary requirements with the uncertainty that remains.

But consider instead a hypothetical first-in-class drug aimed at a target never validated in humans — an Alzheimer’s mechanism that may look convincing in a model but has never been shown to alter the disease. There, no algorithm’s confidence can substitute for a real trial. Patients do not need a beautiful prediction. They need evidence that treatment changes outcomes that matter.

The challenge for FDA is to turn that distinction into practice: to reward AI tools that produce better evidence, while resisting efforts to use AI as a shortcut around proof.

Scientists are already debating how far AI-driven protein and drug-design tools can take us. AI will multiply candidates and may improve dose selection and biomarker discovery. But for the FDA, the key point is simpler: A model’s prediction, no matter how sophisticated, is not the same as evidence that a drug will help patients. Still, the FDA cannot force every AI-generated medicine through an old template.

To be sure, the FDA is not ignoring this shift. The agency has begun issuing guidance on how AI can support drug development decisions and is exploring AI-enabled approaches to clinical trials. This emerging approach is the right one: Do not trust the machine, but test whether the model is credible for the specific decision it supports. This work needs to move much faster. AI-generated candidates may soon arrive at a pace the old regulatory infrastructure was not built to handle.

The agency should adopt a clear, modernized framework. Every major evidence requirement should identify the uncertainty it is meant to resolve. If validated models, analytical data, real-world evidence, or smaller targeted studies can answer it, regulators should accept them. If only a randomized human trial can address the uncertainty, regulators should require it. The standard should not be the evidence tradition happens to demand. It should be the evidence patients need.

There must also be guardrails. Models used to support approval should be tested against real-world patient outcomes, not merely against other models. Their training data, assumptions, and failure modes should be visible to reviewers. Sponsors should not be allowed to hide weak clinical evidence behind sophisticated mathematics. And monitoring must continue after approval, because medicine often reveals its surprises only when it leaves the controlled world of trials.

Used wisely, AI would not replace the FDA’s gatekeeping role. It would sharpen it. The future should not be one in which drugs are approved because an algorithm is confident, nor one in which lifesaving treatments are delayed because regulators cannot distinguish a necessary trial from a ritual one.

The promise of AI in medicine is not simply faster drug design. It is a future in which patients receive lifesaving treatments sooner, with strong evidence that speed does not come at the expense of trust.