Beyond the Hype: What AI Has Really Done for Medicine
A retired veterinary internist looks at what artificial intelligence has already done for medicine, and what it might do next
Why I’m writing this
I may not be the fastest learner, but learning has always been one of my greatest joys. My approach hasn’t changed much since my years in practice: think critically, reason from first principles, and troubleshoot problems analytically, one step at a time. That habit is what drew me to artificial intelligence, and to the question of what it can actually do for medicine.
In my last article, I looked at old, inexpensive drugs being studied as possible cancer helpers. One theme kept coming up: thousands of existing medicines might have uses nobody has tested, and there is rarely money to look. AI is changing that. It can sift through more data in an afternoon than a research team could in a career.
This article walks through where AI in medicine came from, what it has genuinely accomplished, where some smart people think it is headed, and why we should stay both hopeful and realistic.
A note on the name
On September 22, 2026, President Trump told the United Nations General Assembly that United States documents would stop saying “artificial intelligence” and use “super intelligence,” or SI. His reason was that the word “artificial” makes the technology sound fake. [1]
One point of clarity is worth making. Among AI researchers, “superintelligence” has long meant something specific: a future machine that far exceeds human ability across nearly every field. The tools in this article are not that. They are specialized programs that are extraordinarily good at narrow tasks, such as predicting a protein’s shape or spotting disease on a retinal photo. I’ll use “AI” throughout, simply because that is what the research papers call it.
A short history, and two lessons
The idea is older than most people think. In 1950, the British mathematician Alan Turing asked whether machines could think. In 1956, a summer research project at Dartmouth College coined the phrase “artificial intelligence.”
Lesson one: being right is not enough (MYCIN, 1970s). At Stanford in the early 1970s, researchers built MYCIN, a program that identified the bacteria behind severe infections such as bacteremia and meningitis and recommended antibiotics, adjusting the dose for body weight. In a 1979 blinded comparison of ten meningitis cases, expert reviewers judged MYCIN’s treatment acceptable about 65 percent of the time. Faculty specialists on the same cases scored from about 42 to 62 percent. [2] Yet MYCIN was never used on real patients. A consultation meant typing every fact in by hand. The program was not tied to the laboratory. Doctors were not ready to trust it, and nobody had sorted out who would be liable if it was wrong. A good tool still has to fit into real clinical work.
Lesson two: hype is not evidence (IBM Watson, 2010s). IBM’s Watson won the game show Jeopardy! in 2011 and was soon marketed as a revolution in cancer care. It did not live up to the promise. MD Anderson Cancer Center shelved its Watson oncology project after about $62 million, most of it paid to IBM and to consultants. The system never made it into patient care. [3] In 2018, the health news site STAT reported internal IBM documents describing “unsafe and incorrect treatment recommendations.” Watson for Oncology had been trained on a small number of hypothetical patients and the preferences of a few specialists, rather than on real patient data and guidelines. [4]
Anyone who read my last article will recognize the pattern. A system is only as good as the evidence it learns from.
Lesson three: a tool can be everywhere and still be weak. By 2021, Epic’s sepsis prediction model was running in hundreds of American hospitals. An outside check at the University of Michigan found it was not very good at the job. Its ability to tell sepsis from no sepsis was only modest (an area under the curve of 0.63, where 0.50 is a coin flip). It missed about two-thirds of the sepsis cases and set off alerts on nearly one in five admissions. [5] Wide use is not the same as proof.
AlphaFold: the breakthrough that earned a Nobel Prize
If there is one result that proves AI can genuinely advance biology, it is AlphaFold.
Proteins, in plain terms. Proteins are the working machinery of every living thing: enzymes, hormones, antibodies, the receptors drugs attach to. Each one starts as a long chain of building blocks, like beads on a string. That chain then folds itself into a precise three-dimensional shape, and the shape determines what the protein does. A key fits a lock because of its shape, and so does a drug fit its target.
Figuring out a single protein’s shape used to take a lab years of painstaking work. Predicting it from the chain alone was one of biology’s great unsolved puzzles for about 50 years.
The breakthrough. In 2020, Demis Hassabis and John Jumper of Google DeepMind presented AlphaFold2, which predicted protein shapes with remarkable accuracy. They went on to predict the structure of virtually all of the roughly 200 million proteins scientists have catalogued. Those predictions were made freely available to researchers everywhere. [6]
The recognition. In October 2024, Hassabis and Jumper shared the Nobel Prize in Chemistry with David Baker of the University of Washington, who designs entirely new proteins on the computer. [6]
Why it matters for medicine. Designing a drug starts with knowing the shape of what it needs to fit. AlphaFold turned a years-long bottleneck into something closer to a database lookup.
A caution belongs beside the prize. A prediction is not an experiment. Some structures are high confidence, some are not, and a drug still has to be made, tested in animals, and then tested in people. AlphaFold removed one bottleneck. It did not remove the rest.
AI already at work in medicine
AlphaFold gets the headlines. The quiet change is in the reading of pictures.
Pictures, first. Through the end of 2025, the FDA had authorized about 1,430 AI or machine-learning medical devices. About three-quarters of them were in radiology. Cardiovascular and neurology tools made up most of the rest. Pathology, by contrast, had only a handful. [7] This is the part of the story that is already in clinics: software that flags a bleed on a head CT, a suspicious spot on a mammogram, or a polyp during a colonoscopy, and then hands the call to a human. It is assistance, not autopilot, and that is why it has traveled farther than Watson did.
Screening for eye disease. Diabetes can damage the retina and is a leading cause of blindness. In April 2018, the FDA authorized IDx-DR, the first autonomous AI diagnostic system, meaning it makes a screening call without a specialist reading the image. A primary care office takes a photo of the retina, and within minutes the software either refers the patient to an eye doctor or says to rescreen in a year. In its FDA study of 900 people, it correctly caught significant disease about 87 percent of the time, and correctly cleared people without that disease about 90 percent of the time. [8] For patients without an eye specialist nearby, that is a real gain. The system has since been marketed as LumineticsCore.
A new antibiotic lead. Antibiotic resistance is one of medicine’s most serious threats, and new antibiotics have been scarce for decades. In 2020, researchers at MIT trained an AI to recognize molecules likely to kill bacteria, then had it search huge chemical libraries. It flagged halicin, a compound originally studied for another purpose, which killed a wide range of bacteria in the lab, including resistant strains, and cleared resistant infections in mice. [9] Halicin is not yet a human treatment, but the method has since become a model for antibiotic discovery. It is also a cousin of the repurposing idea in my last article: the compound was already on the shelf.
A drug designed by AI from target to molecule. Idiopathic pulmonary fibrosis is a progressive scarring disease of the lungs with few treatment options. Insilico Medicine used AI both to pick a new target in the disease, TNIK, and to design a molecule against it, a drug called rentosertib. In June 2025, Nature Medicine published a placebo-controlled phase 2a trial of 71 patients in China. At 12 weeks, the highest dose, 60 mg once a day, was linked to a mean gain of about 98 milliliters in forced vital capacity, against a loss of about 20 milliliters on placebo. Safety was judged acceptable, though liver effects and diarrhea led some patients to stop. [10]
A sober note belongs here. A 2024 analysis found that AI-discovered drugs were clearing phase 1, the safety stage, more often than older drugs, but failing in phase 2 at about the same rate, roughly 40 percent, and the phase 2 count was only ten finished trials. [11] AI can speed up finding candidates. It does not yet guarantee that they work. Rentosertib is an encouraging 12-week signal in one country, not a finished drug.
AI and drug repurposing: the model done right
This is where my two articles meet. My last piece argued that old, inexpensive drugs may hold untested uses, but that someone has to find them and then prove them. The best example of AI doing exactly that is baricitinib.
The prediction. In early 2020, as COVID-19 was spreading, a British company called BenevolentAI turned its AI system loose on the problem. The system works like a giant map of medical knowledge, linking genes, diseases, drugs, and biological pathways drawn from millions of research papers. In a few days, it flagged baricitinib, a once-daily pill already approved for rheumatoid arthritis, as a likely candidate. The idea was published in The Lancet in February 2020. [12]
Why it made sense. Severe COVID often turns deadly not from the virus alone but from the body’s own overreaction, a so-called cytokine storm that damages the lungs. Baricitinib calms that inflammatory response, and the AI also predicted it might hinder the virus from getting into cells.
The proof. The prediction alone proved nothing. Eli Lilly, which makes the drug, tested it in large randomized trials with the US National Institutes of Health. The benefit was real and modest. In one key trial, given with remdesivir, deaths by day 29 were 4.7 percent with baricitinib and 7.1 percent without it, and patients recovered about a day sooner. In May 2022 the FDA granted full approval for baricitinib in hospitalized adults with COVID-19 who need oxygen, a ventilator, or a heart-lung machine. It had already been available under emergency use since November 2020. [13]
That sequence is the whole lesson: AI suggests, trials prove. It is exactly the process the drugs in my last article still need to go through. AI can also make that process cheaper by screening thousands of existing medicines before anyone spends money on a trial, which speaks directly to the funding problem for off-patent drugs.
Chronic inflammatory conditions: matching the patient to the drug
Baricitinib is itself an anti-inflammatory drug, a JAK inhibitor used for rheumatoid arthritis, which leads naturally to the next frontier: chronic inflammatory diseases such as rheumatoid arthritis and inflammatory bowel disease.
These conditions share a frustrating problem. We have powerful drugs, but no reliable way to know in advance which drug will work for which patient. About 40 percent of people with rheumatoid arthritis do not respond to any given biologic therapy. [14] Patients often cycle through one expensive drug after another for months while the disease keeps damaging their joints or gut.
AI is starting to help with that guesswork:
• Reading the joint tissue. In a trial called STRAP, researchers analyzed gene activity in small joint-tissue biopsies from 208 patients. Machine-learning models predicted, better than chance, which patients would respond to each of three biologic drugs. The scores were about 0.75 to 0.76, where 0.50 is a coin flip and 1.00 is perfect, and two of the models held up in a second trial. [14] Better than chance is not the same as ready for the clinic. A biopsy of the joint is also not a routine blood test.
• Reading the gut. In a 2021 study, AI analyzed the gut bacteria of patients newly diagnosed with rheumatoid arthritis. In a small follow-up group of 21 patients, it correctly flagged about 83 percent of those who would not respond to methotrexate, the standard first drug, and about 78 percent of those who would. [15] I find that one fascinating, given how much we are learning about the gut’s role in whole-body inflammation. Twenty-one patients is a hint, not a test you can order.
The realistic picture: a recent review concluded these tools are promising, but significant barriers remain before they can be used in routine care. [16] Most studies are small, and the models need testing in larger, more varied groups of patients.
Where this could go: optimism, with realism
One of the more thoughtful voices on AI’s future is David Friedberg, co-host of the All-In podcast and CEO of Ohalo Genetics, a company that uses AI to breed better crops. As someone who spends his days growing food, I pay attention when he talks. Here are a few of his recurring themes.
An age of abundance. On Chris Williamson’s Modern Wisdom podcast, Friedberg argued against the gloom that often surrounds AI. He sees AI, along with fusion energy and longevity research, driving the cost of essentials like food, housing, and energy sharply down. [17] He has also called AI a multiplier: it lets us accomplish more with the same or fewer resources. [18]
AI for everyone, not just big companies. He points out that capable open-source AI models can now run on an ordinary home computer, much as the early internet spread from labs into every household. [19] For medicine, that means university labs and small research groups, not just large drug companies, can use these tools.
AI learning the language of DNA. On All-In in June 2026, he described a genome language model from the Arc Institute, trained on genetic sequence from across life, that has essentially learned patterns in DNA. [20] The model is Evo 2. A version was published in Nature in March 2026. If AlphaFold learned to read proteins, models like this are learning to read genes. They can also design genetic sequences, including viruses, which is a power that needs a fence around it.
Now, the realism. I find his optimism refreshing, and I share much of it. But a few cautions keep me grounded:
• Predictions are not results. Abundance, age reversal, and cheap energy are forecasts. AlphaFold and baricitinib are proven. The difference matters.
• Biology is slow and stubborn. AI can design a drug candidate in months, but the human body still decides whether it works, and that takes years of trials. As noted earlier, AI-discovered drugs still fail in phase 2 at about the same rate as others. [11]
• Physical limits remain. AI cannot create fertile soil, clean water, or hospital beds by thinking about them.
• Consider the source. Friedberg runs and invests in AI-driven companies. That doesn’t make him wrong, and his track record is impressive, but in my last article I made a point of noting financial interests, and the same standard applies here.
How I approach a big dream. These cautions are not a reason to stop dreaming. Over the years, my approach to any big goal has followed the same pattern:
1. Start with the dream. Be bold and genuinely optimistic about what could be possible.
2. Build the plan. Lay out a step-by-step path that could realistically get you there.
3. Reassess along the way. Check your progress honestly, again and again, to make sure you’re still on the right path.
4. Be willing to change course. Sometimes the route you were sure was best turns out not to be, and the wise move is to adjust.
AI in medicine deserves exactly that treatment. Friedberg supplies the dream, and it’s a good one. AlphaFold and baricitinib show that the plan can work. Clinical trials, honest reviews, and the occasional retracted paper are the reassessment. And when an AI-designed drug fails, or a tool like Watson or the sepsis model falls short, the right response is not to give up but to change course.
The caveats
None of this means we should hand medicine over to machines. A few cautions apply to every AI tool, no matter how impressive:
• AI can be confidently wrong. The chat tools many of us now use can produce answers, and even research citations, that sound authoritative and are simply false. Always check the original source.
• Garbage in, garbage out. An AI learns from the data it is given. Watson stumbled partly because it was trained on hypothetical cases and a few doctors’ preferences. [4] A model trained mostly on one population may also perform worse for others. The rentosertib trial, for example, enrolled only patients in China.
• Garbage deployed is worse. The sepsis model shows that a weak tool can reach hundreds of hospitals before anyone publishes an outside check. [5]
• Privacy matters. Medical AI needs enormous amounts of patient data. How that data is protected, and who profits from it, are questions worth asking.
• A prediction is a hypothesis, not proof. This is the same lesson as the case reports in my last article. An AI flagging a drug is the beginning of the investigation, not the end.
• The doctor still matters. The best results so far come from AI and clinicians working together, with a person accountable for the decision. The eye-screening system is the exception that proves the rule: it was cleared to make one narrow call, in one setting, after a prospective trial.
Where this leaves us
I started this article curious and I’m finishing it hopeful. In my lifetime, AI has gone from a program that couldn’t fit into a hospital’s computers to a tool that solved a 50-year-old puzzle in biology and earned a Nobel Prize. It reads pictures in ordinary clinics, found a new antibiotic lead, helped repurpose an arthritis drug for hospitalized COVID patients, and may soon help match patients with chronic inflammatory disease to the right treatment the first time.
But the pattern from my last article holds here too. AI is extraordinarily good at asking the right questions quickly. Careful clinical trials are still how we get the answers.
The best thing any of us can do is keep learning, think from first principles, and look past the headlines to the evidence. That has always been my approach, and I think it’s the right one for whatever we end up calling this technology.
References
1. Trump orders AI rebrand as “super intelligence.” Axios, September 22, 2026. https://www.axios.com/2026/09/22/trump-ai-super-intelligence-rebrand
2. Yu VL, et al. Antimicrobial selection by a computer: a blinded evaluation by infectious diseases experts. JAMA, 1979. MYCIN acceptability about 65 percent on ten meningitis cases.
3. MD Anderson Cancer Center Oncology Expert Advisor. University of Texas audit: about $62 million spent; system not used in patient care. IBM support ended 2016.
4. Ross C, Swetlitz I. IBM’s Watson supercomputer recommended “unsafe and incorrect” cancer treatments, internal documents show. STAT, July 25, 2018. https://www.statnews.com/2018/07/25/ibm-watson-recommended-unsafe-incorrect-treatments/
5. Wong A, et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine, 2021.
6. Press release: The Nobel Prize in Chemistry 2024. The Royal Swedish Academy of Sciences. https://www.nobelprize.org/prizes/chemistry/2024/press-release/
7. Golshani P, et al. Radiology accounted for about 76 percent of 1,430 FDA-authorized AI/ML devices from 1995 through 2025. Reported by AuntMinnie, July 16, 2026. https://www.auntminnie.com/imaging-informatics/artificial-intelligence/article/15830219/radiology-dominates-thirty-years-of-fda-ai-device-approvals
8. Abràmoff MD, et al. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. npj Digital Medicine, 2018. FDA marketing authorization, April 2018. Sensitivity 87.4 percent, specificity 89.5 percent.
9. Stokes JM, et al. A deep learning approach to antibiotic discovery. Cell, 2020. https://doi.org/10.1016/j.cell.2020.01.021
10. Xu Z, et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nature Medicine, June 2025. https://pmc.ncbi.nlm.nih.gov/articles/pmid/40461817/
11. Jayatunga MKP, et al. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today, 2024. Phase II success about 40 percent in ten completed trials.
12. Richardson P, et al. Baricitinib as potential treatment for 2019-nCoV acute respiratory disease. The Lancet, February 2020.
13. FDA approval of Olumiant (baricitinib) for COVID-19 in hospitalized adults requiring supplemental oxygen, noninvasive or invasive ventilation, or ECMO. May 10, 2022. https://www.fda.gov/media/143825/download
14. Deep molecular profiling of synovial biopsies in the STRAP trial identifies signatures predictive of treatment response to biologic therapies in rheumatoid arthritis. Nature Communications, 2025. https://www.nature.com/articles/s41467-025-60987-9
15. Artacho A, et al. The pretreatment gut microbiome is associated with lack of response to methotrexate in new-onset rheumatoid arthritis. Arthritis & Rheumatology, 2021. Validation cohort of 21 patients: 83.3 percent of nonresponders correctly classified.
16. Machine learning shows promise in predicting RA treatment response. EMJ Rheumatology. https://www.emjreviews.com/rheumatology/news/machine-learning-shows-promise-in-predicting-ra-treatment-response/
17. David Friedberg on Modern Wisdom: abundance, energy, and AI (summary). https://summify.io/discover/everything-you-know-is-about-to-collapse-david-friedberg-8s2nO_/
18. David Friedberg: AI will produce more than humans can consume. All-In Podcast, March 2026 (summary).
19. Transcript: Everything You Know is About to Collapse, David Friedberg on Modern Wisdom. https://singjupost.com/transcript-everything-you-know-is-about-to-collapse-w-david-friedberg-modern-wisdom/
20. Arc Institute. Evo 2, a genome language model trained across domains of life. Nature, March 2026. https://arcinstitute.org/news/evo-2-one-year-later
