Not in the way the headlines suggest. As of late 2025, no drug discovered and designed end to end by AI has been approved by a major regulator. The closest cases are an AI-suggested repurposing, baricitinib for COVID-19, and a small group of AI-designed molecules in mid-stage trials. Phase 2 is where the claim will be settled.
In June 2025, Nature Medicine published a randomized trial of rentosertib, a pill for idiopathic pulmonary fibrosis made by Insilico Medicine. Both the drug’s target and the molecule itself came out of the company’s generative AI platform. The trial enrolled 71 patients across 22 sites in China. After 12 weeks, patients on the highest dose had gained an average of 98.4 mL of forced vital capacity, a standard lung function measure, while the placebo group had lost 20.3 mL.
Within hours, the result was being described as the moment AI finally delivered a drug. Read the paper and the claim gets smaller. The trial’s primary endpoint was safety and tolerability. The lung function numbers were exploratory, in a small study, over three months, in a disease where patients decline over years.
None of that makes the result unimportant. It makes it a phase 2a result, which is exactly what it is.
Here is the argument of this piece. “Has AI discovered a drug?” is the wrong question, because it treats discovery as a single event. The better question is which part of drug development AI has actually improved. So far the evidence says AI is good at the chemistry exam, making molecules that are safe and behave like drugs, and has not yet shown it is better at the biology exam, picking targets that change disease in people. Phase 2 is the biology exam. It is also where most drugs die.
What counts as an “AI-discovered drug”?
The label covers at least three very different things, and much of the confusion comes from treating them as one.
The first is repurposing: using AI to find a new use for a drug that already exists. The best-known example is baricitinib. In late January 2020, scientists at BenevolentAI searched their biomedical knowledge graph for approved drugs that might both block viral entry into cells and damp down the inflammatory storm seen in severe COVID-19. Baricitinib, an oral JAK inhibitor approved for rheumatoid arthritis, came out on top. They published the hypothesis in The Lancet on 4 February 2020. Eli Lilly then ran the trials, and the FDA granted emergency use authorization for baricitinib with remdesivir in hospitalized patients in November 2020, followed later by full approval.
That is a genuine AI contribution to an approved treatment. It is also a drug whose safety profile, manufacturing and dosing were already known. The AI generated a hypothesis. Conventional trials proved it.
The second is AI-designed molecules for known targets. Here the target biology is established, and machine learning speeds up the search for a compound that hits it cleanly, with good absorption and few off-target effects. This is where most of the “first AI drug in the clinic” announcements of the past five years belong.
The third, and the hardest, is AI-discovered targets: using models to find a new biological lever for a disease and then designing a molecule to pull it. Rentosertib belongs here. Insilico says it identified TNIK as a fibrosis target with its AI platform and then designed the molecule with generative chemistry. That is the full claim, and it is rare.
When someone says AI has discovered a drug, ask which of the three they mean. The answer changes how impressed you should be.
How do AI-designed drugs perform in clinical trials?
The most useful attempt to answer this is a 2024 analysis in Drug Discovery Today by Madura Jayatunga and colleagues at Boston Consulting Group, working with Wellcome. They tracked molecules from AI-native biotechs through the clinic.
Their headline finding: of the 24 AI-discovered molecules that had completed phase 1, 21 succeeded, a success rate of roughly 80 to 90 percent. The historical industry range for phase 1 is roughly 40 to 65 percent. That is a striking gap.
Then phase 2. Of 10 AI-discovered molecules that had completed phase 2, four succeeded, around 40 percent. That is roughly what the industry achieves without AI.
Two cautions sit on top of those numbers. The samples are tiny, especially for phase 2, and a couple of results either way would move the percentage a lot. And “AI-discovered” is a self-declared category, drawn largely from companies that market themselves as AI-first.
But the pattern makes sense. Phase 1 mostly tests whether a molecule is safe and behaves predictably in the body. That is a chemistry and pharmacology problem, and it is exactly what machine learning trained on decades of compound data is good at. Phase 2 tests whether hitting the target actually helps patients. That depends on whether the underlying biology was right, and no amount of elegant chemistry fixes a wrong target.
Why is phase 2 the real test?
Because it is where the money and the hope usually run out.
The recent record of AI-first companies shows the problem clearly. In 2020, Exscientia and Sumitomo Dainippon Pharma announced that DSP-1181, an obsessive-compulsive disorder candidate, had entered phase 1 as the first drug designed with AI to reach human trials. It did not progress. Exscientia later merged into Recursion, completing the deal in November 2024.
Recursion’s own experience is instructive. Its REC-994, for symptomatic cerebral cavernous malformation, met its phase 2 primary endpoint of safety and tolerability. The company presented encouraging trends on imaging at the higher dose in early 2025. Then longer-term extension data failed to confirm those trends, and in May 2025 Recursion announced it was halting that program along with others as part of a pipeline reset after the merger.
That sequence is not an AI failure. It is a normal drug development failure. Which is precisely the point. The molecule was well made. The question of whether it changes the disease was answered the old-fashioned way, by patients over time, and the answer was not good enough.
This is why the rentosertib result deserves both attention and patience. A 12-week signal in 71 patients is where the story starts. Insilico has said it is in discussion with regulators about testing the drug in larger groups. Those trials, measuring decline over a year or more, will say far more about whether AI picked the right target than any press release.
A worked example: better chemistry can make failure more expensive
Here is a hypothetical, with round numbers chosen to show the shape of the problem rather than to describe any real company.
Two biotechs each start ten molecules into phase 1. Company A uses conventional methods and sees 60 percent of molecules pass phase 1, inside the historical range. Company B uses AI and sees 85 percent pass, in line with the BCG and Wellcome figure. Both then face the same 40 percent success rate in phase 2.
Company A runs six phase 2 trials. About 2.4 molecules succeed and 3.6 fail.
Company B runs eight or nine phase 2 trials. About 3.4 succeed and 5.1 fail.
Company B ends up with more drugs heading to phase 3, which is real value. But it also pays for more phase 2 failures, and phase 2 trials cost far more than phase 1 trials. If AI only improves phase 1, it partly moves failure downstream to a more expensive stage. The economics only transform if AI also lifts that 40 percent, which means better target selection, better patient selection, or both.
Insilico’s own figures point to where the gains are real today. The company says its 22 preclinical candidates nominated between 2021 and 2024 took an average of 12 to 18 months from project start to nomination, with only 60 to 200 molecules synthesized per project. That is fast and lean by industry standards. But speed to the starting line is not the same as winning the race.
The strongest case that AI is already working
The counterargument deserves a fair hearing, and it is better than the skeptics usually admit.
First, it is early. Most AI-designed molecules entered the clinic after 2020. Drug development takes a decade, so judging the field on phase 2 results from a handful of programs is like judging a medical school on its first-year exam scores.
Second, the phase 1 advantage is not trivial. If AI reliably produces molecules that are safe and well behaved, it reduces one of the costliest forms of waste in pharma: years spent on compounds that were never going to be tolerated.
Third, the line between “AI drug” and “normal drug” is blurring fast. Large pharmaceutical companies now use machine learning throughout discovery, from structure prediction to toxicity screening. When that becomes routine, the category of “AI-discovered drug” will stop meaning much, the way nobody now talks about “computer-discovered drugs.”
All three points hold. But none of them answers the phase 2 question. They argue that AI makes discovery cheaper and faster. They do not yet show that it makes discovery more right. For patients, the second matters more.
What this means for investors, pharma partners and patients
For investors, the relevant metric is not molecules in the clinic, which AI-first companies are good at generating, but phase 2 readouts on efficacy endpoints, especially for novel targets. A pipeline full of phase 1 successes is a pipeline of open questions. The funding cycle described in our look at health venture capital in 2025, with more money concentrated in fewer companies, rewards platforms that can show clinical proof rather than platform promise.
For pharma partners, the question to ask any AI discovery company is where its model contributed. Target selection, molecule design and patient selection are different claims with different evidence. A partner who pays for “AI” without knowing which step it improved is buying a label.
For clinicians and patients, the practical message is simple. A drug being AI-designed says nothing about whether it works for you. It gets tested the same way as every other drug, and it should be judged by the same trial results. The pattern we described in AI diagnostics holds here too: the evidence is strongest where the question is narrowest, and “is this molecule safe?” is much narrower than “does this target matter?”
For regulators and policymakers, nothing about AI changes the evidence bar for approval, and nothing should. The useful policy question is upstream: whether public research funding and data sharing can help AI models learn from failed trials as well as successful ones. Models trained only on what worked will keep learning chemistry faster than biology.
Frequently asked questions
Has the FDA approved any AI-discovered drug?
No drug whose target and molecule were both discovered by AI had been approved by the FDA or EMA as of late 2025. Baricitinib’s use in COVID-19, first proposed by BenevolentAI using its knowledge graph, is the closest example, but baricitinib was already an approved rheumatoid arthritis drug. The AI contribution was the hypothesis, which conventional randomized trials then tested and confirmed.
Which AI-designed drug is furthest along?
Among drugs where AI was used to find both the target and the molecule, Insilico Medicine’s rentosertib for idiopathic pulmonary fibrosis is the most prominent. Its phase 2a trial, published in Nature Medicine in June 2025, met its safety endpoint and showed an exploratory lung function signal. Larger, longer trials are needed to establish whether it slows the disease.
Why do AI drugs do well in phase 1 but not phase 2?
Phase 1 mainly tests safety and how a drug behaves in the body, which depends on chemistry that machine learning models handle well. Phase 2 tests whether the drug helps patients, which depends on whether the target biology is right. Better molecule design improves the first. It does not guarantee the second, which is why phase 2 success rates for AI molecules look ordinary so far.
Does AI make drug discovery faster?
Early evidence says yes for the preclinical stage. Insilico reports that its preclinical candidates took 12 to 18 months on average from project start to nomination, with far fewer molecules synthesized than traditional programs. But clinical trials still take years, and they are the longest and most expensive part of development, so the overall time to approval has not yet been shown to shrink.
Is an AI-designed drug less safe?
There is no evidence that it is. AI-designed molecules go through the same preclinical toxicology, the same phase 1 safety testing and the same regulatory review as any other drug. If anything, the early phase 1 data suggest they are at least as likely to pass safety testing. Patients considering a trial should weigh it like any other trial, with their clinician.
The real test is still ahead
AI has already changed how molecules get made. It has not yet proven that it changes which diseases get treated. That proof will come from a small number of phase 2 and phase 3 readouts over the next few years, most of them unglamorous, several of them disappointing. The companies worth watching are the ones that publish those results in full, including the failures, because that is the only data from which a model can learn biology rather than chemistry.
Until then, “AI-discovered drug” is best read as a promise about process, not a claim about outcomes.
Which would convince you more: a faster pipeline, or a single AI-found target that works in phase 3? Tell us in the comments.
