AI that already saves lives: diagnosis is racing ahead, the drug is not

THE NUMBER THAT MATTERS. As of 30 March 2026, the U.S. Food and Drug Administration (FDA) had authorized 1,524 AI algorithms for medical use. Of those, 1,163 are radiology tools, 76% of the total, according to the count published by The Imaging Wire. The list passed one thousand devices in late 2025 and grows by roughly 18 new ones a month. This isn't marketing: these are products a regulator reviewed and let into hospitals. The reading is clear. The medical AI that truly works today is not a chatbot that diagnoses on its own. It is a narrow tool that helps a specialist see an image better. Radiology dominates because an X-ray or an MRI is orderly, abundant data, the terrain where these systems perform.

WHAT ALREADY SAVES LIVES. The best example is breast screening. The Swedish MASAI trial, published in The Lancet, is the first large-scale randomized study in this field: more than 105,000 women. A randomized study is medicine's gold standard, because it assigns participants at random to two groups and so isolates the tool's real effect. The result: AI-assisted reading detected 29% more cancers (6.4 per 1,000 women versus 5.0) without raising false positives, and cut radiologists' reading workload by 44%. More important still: it lowered by 12% the cancers that appear between one screening and the next, the so-called interval cancers, which tend to be the most aggressive. A German real-world study of 463,094 women points the same way: 17.6% more detection. Our reading: here, "saves lives" stops being a slogan. Catching a tumor earlier while stripping mechanical work off the doctor is, at once, better medicine and a health system that copes with staff shortages.

THE DRUG STILL ISN'T HERE. In drug discovery the story is different, and it's worth saying plainly. AlphaFold, the Google DeepMind AI that predicts the three-dimensional shape of proteins, was a genuine scientific leap and won the 2024 Nobel Prize in Chemistry. Its subsidiary Isomorphic Labs unveiled its own engine, IsoDDE, in February 2026, which the company says doubles AlphaFold 3's accuracy on the hardest cases. It sounds spectacular. But watch the dates: in January 2026, founder Demis Hassabis admitted at the Davos forum that the first human trials would not arrive until the end of 2026, not in 2025 as had been suggested. The figure that frames the whole debate is this: there are more than 200 AI-discovered drugs in clinical trials, and zero approved by the FDA. The most advanced candidate, Insilico Medicine's INS018_055 for pulmonary fibrosis, is in phase II. AI speeds up the cheap, fast part — designing the molecule — but the bottleneck is still the clinical trial: years of testing in people that no AI can skip. It's a warning against euphoria: demonstrated capability is not the same as a medicine on the shelf.

THE REAL RISK: IT HALLUCINATES WITH CONFIDENCE. The short-term danger is not a rogue superintelligence. It's that these models invent data and do so in a confident tone that deceives. In medicine this has a name: clinical hallucination, an answer that sounds plausible but is false. The figures are alarming. When the input text contains an error, some models propagate it in up to 83% of cases, according to analyses reported by specialist outlets. More than 45% of the bibliographic references they generate are fabricated. And in one survey, over 90% of clinicians said they had run into hallucinations and 85% considered them capable of harming a patient. Add automation bias: the doctor's tendency to trust the machine and lower their guard. A model can invent a measurement in a report, a wrong laterality — left for right — or a finding that doesn't exist, and from that flow decisions about biopsies or treatments. Our reading: image AI that classifies and prioritizes is already reliable; generative AI that writes free clinical text is not yet. Putting the second where the first belongs is the most expensive mistake a hospital can make.

WHO SETS THE RULES. Regulators know it. In January 2026, the U.S. FDA and the European Medicines Agency (EMA) jointly published ten "good AI practice principles" for drug development. The common thread is a simple idea: human in the loop. That is, the AI proposes, but a named person reviews and signs; the model's output is never used as-is. In Europe, the AI Act will classify medical uses as "high risk," with requirements starting to apply in August 2027. The figure that marks the limit: to date there is not a single fully authorized generative-AI device, though in March 2026 the FDA gave breakthrough designation to one such application. Translation: almost everything approved is "narrow," predictable AI, not the AI that converses. The regulatory risk, as we keep arguing, is legislating panic instead of real capability. Governance must rest on evidence and hard test tasks, not headlines.

OUR LONG-TERM READING. Let's line up the two speeds. In the short term, the honest thing is to admit the problems: models that hallucinate, biases that penalize populations poorly represented in the data, and the temptation to cut staff too soon. None of this is minor. But the horizon tilts the balance toward optimism, and rightly so. If image AI already brings a tumor diagnosis forward, and if molecule design — today the slow link — matures enough to cross the clinical trial, the direction is personalized medicine: treatments tuned to each patient's biology, diseases that are lethal today turned chronic and, at best, eradicated. This is not faith: it's the reasonable extrapolation of what is already measured. The doctor's job does not vanish; it changes. The professional who orchestrates these tools and brings judgment and human care wins. The one who only repeats mechanical tasks does not. That is the hard transition of the coming years, on the road to a more abundant and fairer medicine.