top of page

Welcome to the New Era of AI Medicine

David Galland
15 hours ago
31 min read

Hello rational optimist –


Stephen and I are en route to Austin, then on to SF, LA, and NYC to meet a new crop of founders and see their technology firsthand. We’ll document it on Twitter with photos and videos – follow us here and here to follow along.


Today’s Rational Optimist Diary is a Deep Dive into the stunning progress now being made in medicine thanks to AI… enjoy! – Dan Steinhart


For three years, Courtney Hofmann's young son Alex suffered pains no one could explain.


He chewed on things to cope with headaches. His gait changed, and he stopped growing. By the time he turned seven, he had been seen by 17 doctors and specialists, including pediatricians, neurologists, orthopedists, ENTs, and even dentists.


Every doctor followed the established protocols, ordering exams to test their latest theories, but ultimately none came up with a helpful diagnosis.


Finally, in desperation, Courtney pasted her son’s medical records into ChatGPT and asked for an opinion.


After engaging in a collaborative exchange about her son’s symptoms, ChatGPT came back with tethered cord syndrome, a rare condition in which the spinal cord is abnormally anchored to the surrounding tissue instead of floating freely. As Alex grew and his spine lengthened normally, the tethered cord was increasingly stretched.


ChatGPT results in hand, Courtney found a neurosurgeon who specialized in the condition. He confirmed the diagnosis, and Alex had the necessary surgery. Today, he's fine.


And so we find ourselves on the cusp of a new era of medicine and healthcare where, for the price of a modest monthly subscription, a chatbot working from the same records as 17 medical professionals correctly diagnosed a complicated medical problem the experts had all missed.


In this Deep Dive, we'll look at the astonishing progress AI is enabling across the very foundations of medicine, including disease diagnosis, drug development, and treatment.

As you’ll read, AI is almost certainly the most important development in human health since antibiotics.


And it’s happening right now, as you read this, which means we’ll be the beneficiaries of this healthcare revolution.


Let's get to it.


A Quick Historical Perspective


In the fall of 2011, I attended a Singularity Summit in New York City where I watched two fascinating presentations. The first was by Jeopardy! champion Ken Jennings, who described losing to IBM’s new Watson computer. He was followed by members of the Watson team, describing how they did it.


Beating a television game show may seem trivial until you consider every question is a riddle of nuances. The example Jennings gave involved a question which included the word "queen."


Did it refer to the Queen of England, or the sovereign of some other nation? The rock band? A playing card, or the matriarch of an ant colony? Only after identifying the correct context could the computer hunt down the answer, and it had to do so faster than Jennings could slap the buzzer in front of him.


In other words, this was serious computational skill on display. (You can watch their presentations on YouTube.)


During the same conference, the IBM team announced it was going to use the Watson AI to revolutionize cancer detection and treatment. Impressed by what I had heard, I was convinced the war against cancer would soon be won.


Brimming with what turned out to be irrational optimism, IBM partnered with Memorial Sloan Kettering, MD Anderson, and the Cleveland Clinic, ponying up an estimated $4 billion in its quest to provide every oncologist in the world with a Sloan Kettering-grade diagnostic tool in their pocket.


Didn’t happen.


What went wrong?


The problem was never computing power, of which Watson had plenty. It turned out that discerning the answer to a trivia clue on Jeopardy! and diagnosing cancer were two entirely different things, with the latter requiring making judgment calls based on incomplete information.


Watson was further handicapped by not training on real patient data, but rather on a handful of made-up cases written by Sloan Kettering doctors. At the end of the day, all it did was regurgitate their opinions.


Looked promising, but the devil was in the details, and IBM got those wrong.


Why Now Actually Is Different


The current AI-driven progress in radically improved diagnosis and healthcare is the result of three fairly obvious developments unfolding at the same time:


Data digitization – Though privacy rules and fragmented systems still create roadblocks, today nearly all medical records and associated images have been digitized, creating a pool of trillions of data points available for analysis and, importantly, training AI.


Foundation Models – A single large model can now process text, images, genomes, and waveforms—such as those produced by an ECG—all at once. The latest AI models can read your medical history, review your charts, study your scans, and weigh your genetics, processing every input together in an analysis that happens in the blink of a proverbial eye.


Compute – The AIs powering today's medical breakthroughs run on a level of computation that would have been prohibitively expensive—and perhaps impossible—just 15 years ago.


Put those three developments together, and you get something Watson could only dream of: a near-superhuman diagnostician that can weigh your personal health data against a vast library of peer-reviewed research, massive clinical databases, and archives of medical images tied to known outcomes—and apply all of it in seconds.


The consequences for you, me, and all of humanity are beyond huge.


To help you understand my enthusiasm, let’s take a brief tour of the most promising ways AI is supercharging medicine.


Breakthrough #1: Diagnosis


Whether it’s a radiologist at a busy hospital reading between 50 and 100 scans a day, a dermatologist seeing 30 patients, or a pathologist examining a cartful of slides during their shift, all three spend a large percentage of their workdays engaging in what is essentially pattern recognition.


And due to the increasing shortage of medical professionals, as often as not they’re examining patterns while fatigued.


That is quickly changing.


The New Radiology


In 2016, AI pioneer Geoffrey Hinton told an audience of researchers that medical schools should stop training radiologists because, within five years, AI would be more accurate at reading images than humans.


You hear this sort of thing—how AI is going to eliminate entire professions—a lot these days.


In most cases, including radiology, the reality is that rather than replacing professionals, AI makes them far more effective. Machines don’t take coffee breaks or nod off in front of a screen while trying to make sense of an image at the end of a long shift.


The first randomized controlled trial of AI in cancer screening involved nearly 106,000 women in Sweden. During the trial, radiologists, supported by an AI system called Transpara, detected 29% more breast cancers than using standard practices, with no increase in false positives.


Importantly, those results were produced while cutting screen-reading workload by 44%!


That, of course, allows radiologists to take on more patients, reducing pressure on healthcare delivery.


While the Swedish study was a trial, real-world results have been just as impressive. An observational study in Germany involving the routine screening of nearly half a million women found that radiologists using AI caught 17.6% more breast cancers than those not using it.


Pause for a moment to consider how many lives that will save.


Similar gains are being documented in chest and musculoskeletal imaging and in stroke care, where purpose-built tools like Viz.ai are now analyzing CT scans for large-vessel occlusions and automatically alerting the on-call stroke team the moment one is detected.


So, on the one hand, Hinton was right about AI-driven machines: pattern recognition is exactly what they do best. What he got wrong was that, rather than eliminating the profession, AI is allowing radiologists to do more, and do it more accurately, than ever before.


The Pathologist's Partner


I used to be on a sports team with a pathologist who once told me he "played God" every working day. That's because, depending on his analysis, which back then was little more than an informed opinion, patients would either receive a clean bill of health and get on with their lives, or have their world turned upside down by a cancer diagnosis.


The size of the stakes involved is exactly why pathology is such fertile ground for AI. The images pathologists examine are extraordinarily dense, containing upwards of a billion pixels each. A human pathologist, under time pressure—a constant these days—can survey only a fraction of that data before making a call based on training, experience, and intuition. And those calls are very much subject to error. 


By contrast, an AI trained on massive databases of past results can examine the entire image in minute detail, flag suspicious regions, and quantify things humans can't. For example, the exact ratio of particular cell types, or subtle staining patterns that correlate with known outcomes.


In 2021 Paige Prostate became the first FDA-cleared AI pathology tool of any kind. It was joined by Ibex's Prostate Detect, cleared by the FDA in 2025, which in trials flagged 13% of cancers that pathologists had missed on first read.


Beyond simply spotting a cancer, AI is now learning to predict how aggressive it will be. That matters because the same type of tumor growing slowly in one patient can be growing rapidly in another, and traditional pathology often can't tell the two apart.


Dermatology


As a result of spending my childhood running around shirtless under the Hawaiian sun, I now practice constant vigilance lest I suffer the delayed consequences. Caught early, skin cancer is usually survivable. Left too long, it can be one of the deadliest.


Examining moles and lesions is exactly the kind of pattern recognition machines excel at. Which is why a 2026 meta-analysis involving 2,500+ patients and 50+ doctors reported in JAMA Dermatology found AI systems now exceed the accuracy of dermatologists at diagnosing melanoma from dermatoscope images.


Specifically, in the meta-analysis, the AI identified melanoma with 80.9% sensitivity against 78.6% for the specialists, with similar outcomes in terms of false alarms.


That last part is important because AI exceeded human diagnostic accuracy without causing an uptick in benign moles being excised and sent for unnecessary biopsies.


The technology is now finding broader applications. In a pivotal study, DermaSensor, an FDA-cleared handheld device designed for primary-care physicians, roughly halved the rate of missed skin cancers, from 18% to under 9%. That matters, because most people are far more likely to visit their GP rather than see a dermatologist.


In time, and probably not all that much time, the odds are good you’ll be able to use your AI-linked phone camera as a diagnostic tool. In some underserved countries, they are already being used due to the lack of better alternatives. But for now, if you’re concerned about a mole or lesion, please see a doctor.


The Eye Knows


The retina is one of the few places in the body where you can directly see blood vessels. It turns out those vessels carry a lot more information than you might expect.


A team led by Pearse Keane at Moorfields Eye Hospital, working with Google DeepMind, built an AI that reads OCT scans (basically an ultrasound for your eyes) and can diagnose more than 50 eye diseases at specialist-level accuracy. That's a breakthrough.


And it turns out the eye reveals far more than eye disease. Separately, Google researchers showed that a deep-learning model analyzing ordinary retinal photographs can estimate a person's cardiovascular risk, guess their age to within about three years and tell whether they smoke. Other studies have since extended the approach to identifying signals for stroke and kidney disease. None of it requires blood tests, DNA sequencing, or anything beyond an image of your eye.


An AI reading the retina is already saving sight at scale. That’s because Google's diabetic-retinopathy screening system is now being deployed in rural India and Thailand to detect vision-threatening disease about as accurately as a retina specialist.


In remote areas with perhaps one ophthalmologist per million people, a portable retinal camera feeding an AI is truly revolutionary—certainly for matters of the eye, and in time, much more.


The ECG That Sees the Future


Cardiologist Paul Friedman and his team at the Mayo Clinic have spent the last decade studying what information the squiggly lines on an ECG might contain that humans have been missing for a century.


It turns out, quite a lot.


Their AI, trained on millions of tracings paired with long-term outcomes, can now look at a completely normal-looking ECG and detect with striking accuracy whether a patient has hidden atrial fibrillation the eye can't see, flagging an elevated risk of heart failure years before it appears.


And this isn't just a lab experiment. Mayo put their technology to the test in a randomized controlled trial of some 22,000 patients across 45 clinics. Doctors who used Mayo’s AI in conjunction with their ECGs were 32% more likely to diagnose a heart issue.


Mayo has since spun the technology into a company, Anumana, which in 2023 earned FDA clearance for its heart-failure test, followed by clearances for pulmonary hypertension in March 2026 and cardiac amyloidosis in April. In other words, it's coming to your doctor’s office soon, and may already be there.


The practical implications are enormous.


The First FDA-Approved Blood Test for Colon Cancer


In July 2024, the FDA approved Shield, made by Guardant Health, as the first blood test cleared as a primary screening option for colorectal cancer. Rather than having to go to a doctor for an "uncomfortable" exam, Shield involves a single routine blood draw.


That's important because roughly one in three eligible adults (including yours truly) skips regular screening.


Shield sequences the blood for faint molecular traces of cancer—including tumor DNA, altered methylation patterns, and telltale fragmentation—using a machine-learning model trained to analyze thousands of signals no human could read. In its pivotal 20,000-patient trial, it caught 83% of colorectal cancers.


In August 2024 Shield was the first such test to qualify for Medicare coverage, which should allow it to scale quickly into the general population. As colorectal cancer is now the #1 cause of cancer death in men under 50 and the second in women that age, the timing couldn’t be better.


Shield won't replace the colonoscopy, because a positive result still requires one so that any precancerous polyps can be found and removed. But that's not the point. Millions of people put off colonoscopies indefinitely. For them, the real choice isn't a blood test versus a colonoscopy—it's a blood test versus nothing.


In May 2026, the American Cancer Society added Shield to its official colorectal screening guidelines as an option for people who decline the preferred tests—and Guardant is now developing a multi-cancer version of the test, which received FDA Breakthrough Device designation in 2025.


Cancer Radar


Imagine going to your doctor for an annual check-up and, as part of the standard blood work, they conduct a test able to catch cancer while still early enough to treat easily.


That’s what companies such as Grail—spun out of gene-sequencing giant Illumina and backed by the likes of Jeff Bezos and Bill Gates—are working toward.


The company's flagship Galleri test uses a proprietary AI to screen a single blood draw for more than 50 cancers. Its "secret sauce" is the ability to detect DNA fragments shed by tumors into the bloodstream long before the cancer would normally produce any symptoms or cause serious harm.


The UK's National Health Service put it to the test in a 140,000-person randomized trial—the first of its kind for any multi-cancer blood test. The goal was to see whether using the test could produce a statistically meaningful drop in combined late-stage (Stage III and IV) diagnoses. It missed that mark.


The test did find roughly four times as much cancer as standard screening alone, and it cut the most advanced Stage IV diagnoses by about 14%. But that benefit didn't carry over to Stage III—and because the goal combined the two, the overall late-stage number didn't fall enough to count as a win.


Thus, it offers a promising but still unproven first step toward a single blood draw that can catch any number of cancers far sooner than would otherwise be the case.


Grail isn't alone. Exact Sciences has now launched Cancerguard, its own multi-cancer blood test. And Delfi Diagnostics is using AI to read the way DNA fragments in blood are chopped up—an approach called "fragmentomics"—to flag people who should get a lung CT scan based on nothing more than a routine blood draw.


The goal here is to move the diagnosis of cancer from "we found a lump" to "we found an anomaly at the molecular level," leading to much earlier and less invasive treatments.

Given the level of interest and funding going into this research, multi-cancer blood screening could well become as routine as a cholesterol check.


That shift alone could be the most significant thing to happen to oncology since chemotherapy, with the savings in time, treatment costs, and sheer stress for patients and their families.


As you’ll note from the dates throughout this article, the breakthroughs in cardiovascular and cancer screening are happening now—in 2026. Which is to say that the ability of AI to detect diseases early with superhuman accuracy is already here and will get dramatically better going forward.


And with that dose of optimism, our story continues with the 50-year-old puzzle we discussed in our Deep Dive on gene editing, Playing God.


Breakthrough #2: Drug Discovery


If diagnosis is where AI is most obviously reshaping medicine now, drug discovery is where it has the potential to most dramatically impact the future.


The Protein Folding Problem


Proteins build tissue, carry oxygen, fight infection, and perform a hundred thousand other essential functions in the body. Each of those functions depends on the protein's shape—specifically, how it folds from a simple string of amino acids into an intricate 3D structure.


That shape is everything because proteins do their jobs by physically locking onto other molecules at precise pockets on their surface.


This is also how most drugs work. When a disease is driven by a particular protein—say a cancer protein telling cells to divide or a viral protein helping a virus replicate—a drug can shut it down by slotting into one of those pockets and jamming it, like breaking a key off in a lock. But the key has to match the lock exactly.


Fit it well, and the drug binds only to its target and does its job; fit it poorly, and it either does nothing or jams the wrong proteins and causes harmful side effects. So, if you can work out a protein's precise 3D shape, you can design a molecule contoured to fit it.


Therein lay the problem. Until recently, working out the structure of a single protein required years of laboratory work. Roughly 200 million known proteins exist in nature, yet by 2020 we had "solved" only about 170,000 of them. Absent a breakthrough, finishing the job would have taken thousands of years.


Everything changed with AlphaFold.


Built by Google DeepMind under Demis Hassabis, AlphaFold 2 debuted at the 2020 CASP14 competition—the Olympics of protein-structure prediction—and utterly dominated. Experts described it as a moment on par with sequencing the human genome.


Within two years, AlphaFold had predicted the structures of essentially every known protein on Earth and made the entire database freely available to researchers.


AlphaFold 3


In May 2024, DeepMind released AlphaFold 3, which goes a step further, predicting not just the structure of individual proteins but how they interact with DNA, RNA, and the very small molecules that make up drugs.


And it did so with roughly 50% better accuracy than the best conventional tools.


This is the drug-discovery holy grail. If you can simulate how a candidate drug will dock with a target protein, you will eventually be able to screen millions of possibilities in hours instead of years. And you can save a massive amount of time and money by ruling out duds before you synthesize them.


Hassabis, who along with his colleague John Jumper won the 2024 Nobel Prize in Chemistry for AlphaFold, has gone on record predicting AI will compress the discovery phase of drug development from years to "months or maybe even weeks."


That's not to say we’ll be able to produce a finished drug in weeks. The drugs still need to plow through the clinical trials that consume most of a medicine's 10-to-15-year, roughly $2 billion journey to market. But if the discovery stage drops from years to months, the whole pipeline accelerates.


We are not there yet. But we are clearly on our way.


Note: More on other companies working with AI to revolutionize medicine can be found in the Addendum.


And with that, our journey into the bright new world of AI in medicine and healthcare continues.


Breakthrough #3: Personalized Treatment


Having looked at how AI is helping improve diagnosis and the discovery of new drugs, we turn to AI being used to match the right treatment to the right patient.


Pharmacogenomics


In researching this topic, I was surprised to learn that adverse drug reactions kill upwards of 100,000 Americans every year. Genetics is an under-appreciated part of the problem. That’s because about 93% of us carry a genetic variant that meaningfully changes how we respond to common medications. A drug that helps one patient can do nothing for another—or even kill them.


Testing for these drug-gene mismatches isn't new; a single cheek-swab panel screening dozens of genes against hundreds of medications has been available for over a decade. What's new is that those genetic results are now being connected directly into prescribing systems, where software flags a dangerous drug-gene combination in real time at the pivotal moment a clinician begins writing a prescription. 


Companies like OneOme, co-founded by the Mayo Clinic, and Genomind are already selling these tools to clinicians. In fact, the evidence convinced several leading academic centers—Vanderbilt, Mayo, and Mount Sinai—to build preemptive pharmacogenomic testing directly into their own electronic medical records years ago. Vanderbilt's program has been running since 2010.


Deployed widely, this approach could prevent thousands of deaths a year.


A Cancer Vaccine Made Just for You


Long before mRNA became a household name due to its use in the one-size-fits-all COVID shot, it was the foundation for a personalized cancer vaccine.


When a patient is diagnosed with cancer, a sample of their tumor is sequenced. AI analyzes the tumor's unique genetic mutations, identifies the ones most visible to the immune system, and designs a custom mRNA vaccine that teaches the patient's body to recognize and attack just those specific mutations.


Each vaccine works only for that one person and nobody else.


In December 2022, Moderna and Merck announced Phase 2b results from a trial combining Keytruda—considered the current best-in-class treatment—with a personalized melanoma vaccine then known as mRNA-4157. Compared to Keytruda alone, the combination reduced the risk of cancer recurrence or death by an impressive 44%.


As we were working on this Deep Dive, the same combination—with the vaccine since renamed intismeran autogene—successfully made it through a 1,137-patient Phase 3 trial, again outperforming Keytruda alone at preventing both recurrence and the spread of cancer.


It was the first successful Phase 3 result for a personalized cancer vaccine, and for an mRNA-based cancer therapy of any kind. The program has since grown to nine trials across other tumor types, including lung, bladder, and kidney cancer, with early-stage studies now underway in pancreatic and gastric cancer.


BioNTech, a German company founded in 2008, has a similar program, with early results in pancreatic cancer showing some patients remaining disease-free more than three years out. If you know anything about pancreatic cancer, that’s an amazing result.


Once again, this sort of breakthrough is only made possible by AI identifying which mutations in a given tumor will best train the immune system. It’s a problem no human researcher working without AI could solve in a lifetime. With AI, the solution can be found in hours.


Happily, the trials for this entirely new round of miracle treatments are well-advanced, meaning big changes are now visible on the horizon.


The Virtual Expert Panel


As something of an aside, but an important one, in the last two years, we have seen the introduction of multi-agent systems that are essentially teams of AIs that specialize, debate, and critique one another.


Microsoft's recent MAI-DxO study is illustrative. The system consists of a team of AI agents. One proposes diagnoses, a second challenges them, a third weighs which tests to order and at what cost, and a fourth checks the reasoning for errors. When presented with the toughest cases from the New England Journal of Medicine (NEJM), a group of experienced physicians got the diagnosis right about 20% of the time. The AI virtual expert panel got it right closer to 85%.


Efficiency is the panel’s biggest advantage, as it achieved its impressive results—outperforming experienced physicians by more than four to one—in a fraction of the time and cost of conventional methods.


That result is especially impressive when you consider these were some of the most difficult diagnostic puzzles in the profession—cases chosen by the NEJM precisely because they're extraordinarily hard.


In the foreseeable future, every tough case will be routed through some version of that multi-agent system, and the physician’s job will shift from producing the diagnosis to verifying it.


It is not a very large leap to expect that, in time, even doctors in remote areas will be able to access some sort of virtual expert panel on their mobile devices.


Bonus Breakthrough: Improved Medical Access


The World Health Organization projects a global shortfall of roughly 10 million health workers by 2030.


This is largely caused by two factors, the first being a surge in demand from aging populations, which invariably require far more care. In the U.S. alone, the 65-and-over population will grow by a third over the next decade.


Complicating things, the supply of doctors is shrinking. Roughly one in five American doctors is already past 65, with another 20% close behind, and burnout is pushing many out the door early.


A much-cited 2016 time-motion study in the Annals of Internal Medicine found that for every hour physicians spent face-to-face with patients, they spent nearly two more on the electronic health record (EHR) and other desk work—in no small part because the EHR was designed largely around billing and compliance, which effectively turned doctors into data-entry clerks.


That all changes with ambient AI scribes, increasingly installed as standard exam-room fixtures. As the term suggests, their job is to listen to the conversation between doctor and patient and automatically generate a clinical note for the record that captures the diagnosis and treatment plan and ticks all the boxes for compliance and billing.


Increasingly, they go further still, drafting the visit's follow-up orders, prescriptions, referrals, and the patient's after-visit summary. The doctor reviews and signs off. Done. Both patients and doctors win, with doctors able to see more patients per shift and spend more time focusing on the patient as opposed to the paperwork.


A Specialist in Every Phone


Earlier I mentioned the deployment of handheld devices to spot diabetic retinopathy in India and Thailand.


The importance of that sort of development becomes clear when you consider that roughly half the world's population lacks meaningful access to essential health services. In sub-Saharan Africa, there is roughly one pathologist per million people.  


Entire countries lack a single pediatric cardiologist.


In the new world of medicine, a rural doctor armed with a smartphone or purpose-built pocket-sized device can now carry a dermatologist (DermaSensor, the FDA-authorized skin-cancer scanner built for primary care), a retinal screener (Google's ARDA platform), and a cardiologist (Eko Health's AI-powered stethoscope) in their pocket—tools that a decade ago existed only in major academic centers.


Of course, tools and systems found in modern hospitals and clinics will be more sophisticated, but this sort of field-level analysis is a game changer for remote populations.


The Final Mile: Why This Is Harder Than It Looks


Everything I've described so far is real. But as with gene therapy and its delivery problem, AI in medicine has a few final-mile problems of its own.


Potentially Deadly Hallucinations 


We all know AI models can confidently make things up. In a clinical setting, such hallucinations could be deadly.


Fortunately, the hallucination problem is being reduced with each iteration of the best AIs. Case in point, when researchers tested ChatGPT on generating citations for research papers, GPT-3.5 fabricated 55% of them outright. Its successor, GPT-4, fabricated just 18%, a threefold drop in a single model generation.


The problem is far from solved, but it’s moving in the right direction, in part due to two relatively new developments. The first is "retrieval-augmented generation," which forces the model to pull from a verified knowledge base rather than its own memory.


Also helping are purpose-built medical models like Google's Med-Gemini, trained specifically for medical accuracy. The virtual expert panels discussed earlier, where multiple AIs cross-check one another's work, will also certainly be a part of the solution.


Even so, for the foreseeable future, any AI used in a clinical setting must be built on the assumption it will sometimes be wrong. That is why a human must remain involved at every crucial decision point.


The Regulatory Maze


The FDA has spent years building a framework for AI in medicine, clearing more than 1,500 AI-enabled devices (most of them in just the last five years). But nearly all of them run "locked" models that don't change on their own. The agency's 2024 Predetermined Change Control Plan lets manufacturers ship pre-approved, validated updates without a new submission each time.


Note that each of those updates is a controlled, deliberate release, not an update produced by a constantly iterating model, which is a serious blind spot in the regulatory framework.


In fact, up until early 2026, the FDA had not authorized a single device powered by generative AI or a large language model. The first to reach the starting line in the form of a preliminary "breakthrough" designation did so only in March 2026. Which means the LLM-based tools discussed elsewhere in this article, including the ambient scribes, the diagnostic panels, and the chatbots, largely operate outside the regulatory approval system.


(As one might expect, the EU's new AI Act takes an even more conservative, precautionary approach, while China is moving fast and often in a less transparent direction.)


The point is that the regulatory framework for AI in medicine and healthcare is far from settled and will remain unsettled for the next several years as regulators figure out how to handle AIs that can be updated weekly based on new information.

In a similar vein, there is still the matter of liability. Who is responsible when an AI makes a harmful mistake? Is it the doctor who followed its advice? The hospital that deployed it? The company that built it? At the moment, the answer is a legal muddle.

Until the level of trust and clarity grows, the entirely rational decision by doctors and hospitals to embrace AI cautiously is going to provide something akin to speed bumps on adoption.  


Risks and Controversies


Deskilling

If the AI catches the tumor first in every scan for 20 years, what happens to the next generation of radiologists? Do they still develop the skills and instincts of current practitioners? Or do they become rubber-stampers for a system they couldn't replicate if it failed?


The aviation industry provides a cautionary corollary. Decades of research, including the FAA's own studies, have documented how heavy reliance on autopilot erodes pilots' manual flying skills. At this point, regulators treat automation dependency as a top safety threat.


The concern is the same in medicine. How we train doctors in an AI-assisted world is a serious question.


Automation Bias

Related to deskilling, but distinct, automation bias is the documented tendency of humans to trust machine outputs more than their own judgment. If your radiologist is tired and the AI says "no cancer," it takes real effort to override that.

Systems need to be built to keep humans in the loop in genuinely meaningful ways, rather than just having them rubber-stamp the AI's conclusions. No small task.

That said, I do think the AI virtual expert panel approach will help ensure accurate conclusions, but if every decision is made by AI and the human becomes little more than an operator, how will that impact the skill levels of the human medical community going forward?


Food for thought.


Concluding Thoughts


As we begin to wrap up, let’s pause to consider pancreatic cancer, a disease that kills most patients within a year, largely because it's so rarely caught early. Around 80% of cases are found only after the cancer has spread, which is why five-year survival sits at just 13%. Catch it while the tumor is still confined to the pancreas, and survival leaps to 44%. Early detection is, quite literally, the difference between life and death.


This is exactly the sort of problem AI is so good at solving. MIT's PRISM model flags high-risk patients up to 18 months before diagnosis, catching three times as many cases as traditional screening.


In addition, Mayo's imaging AI spots the disease on CT scans radiologists had read as perfectly normal, up to three years earlier than would have been the case. And Northwell's iNav—one of TIME's Best Inventions of 2024—reads 10,000 scan reports a week across 28 hospitals, halving the time from scan to biopsy.


None of these AI tools are perfect, as each still raises false alarms. But all do one thing no single human is able to: examine a pattern of data and identify the warning signs of pancreatic cancer. For a disease where a few months' warning can move a patient from the 13% to the 44%, that is no small thing. It’s a clear and compelling case for embracing AI in medicine.


What does the future hold?


While no one, certainly not your correspondent, can reliably tell the future, based on everything I have read, the following outcomes seem reasonable.


By 2029, expect most primary care visits in the developed world to be AI-augmented. The doctor will still be human, but an AI scribe will handle documentation, an AI diagnostic assistant or expert panel will cross-check the doctor’s prognosis, and an AI decision-support tool will flag drug interactions and risks the human might miss.


By 2030, and perhaps a bit sooner, expect the first fully AI-designed drugs to be approved by the FDA. Probably in oncology or rare disease first, where the current system is weakest and the outcomes from failing to act quickly are often fatal. Once AI-designed drugs begin receiving approval, expect a cascade of new and more effective medicines to follow.


By 2035, expect a usable AI diagnostician to be standard on every smartphone. It will be free or very cheap and available globally. It may not be perfect, but it will be enormously better than what 4 billion people currently have access to.


And within the next 15 years, expect the cost of high-quality medical diagnosis to collapse—probably by 90% or more. Because the marginal cost of adding one more patient to an AI's workload is approximately zero, and that changes everything. Of course, early diagnosis—for heart disease and cancer—will mean a steep reduction in mortality from those causes.


The most advanced treatments will stay expensive for years because, at the moment, personalized cancer vaccines, gene therapies, and custom-designed drugs don't come cheap. But the first and most important step in medicine, namely figuring out what's wrong with you, is about to become almost free. Expert-level diagnosis will no longer require a specialist or a big hospital. It will be available to almost anyone, almost anywhere, often from a device already in their pocket.


I could be wrong about any of the specifics, and may be too pessimistic on the timelines mentioned, but I’m highly confident about the direction.


What about the hurdles AI faces in medicine?


In 1999, when worries about the approaching Y2K were at their height, I asked Milton Friedman over lunch if he was concerned about a global computer crash. He offered to bet me a one-ounce gold coin Y2K would amount to nothing (which I wisely declined). To paraphrase his reason for being unconcerned:


“We have known about the problem for several years, plenty of time for companies, governments, and computer programmers to adapt and sort things out. We humans are very good at problem-solving; it’s what we do best. Everything will be fine.”


In the end, he was right.


Will there be failures in medical AI? Yes. Will those failures be trumpeted in stark headlines around the world? Without question.


But the right comparison is not AI in medicine against a perfect system, but against the system we have now. A system in which millions of people die each year from diseases we already know how to diagnose and treat, simply because they don't have access to the necessary specialists.


In the current world, the mother of a desperately sick child has to consult 17 doctors in the hope of finding a cure. And patients die of pancreatic cancer that could have been caught two years earlier if only someone had noticed the pattern.


After immersing myself in the topic, I’m convinced AI is something akin to mankind’s discovery of the practical uses of fire. If used poorly, disastrous consequences are possible. However, aware of those dangers, we will proceed to harness the power of AI in ways we have only just begun to imagine.


I really hope I have succeeded in sharing my optimism about this topic, because this is going to be world-changing—the sort of development, like the discovery of antibiotics, that divides history into before and after. We are literally on the cusp of the “after” part, and it’s going to be incredible.


If you found this Deep Dive useful, please forward it to others and encourage them to join the Rational Optimist Society. It's free, so there's nothing to lose, with only lots of upbeat news and valuable information to gain.


Until next time,


—David Galland


For the Rational Optimist Society


What do you think? Have you used AI to get a second opinion on a medical issue? Would you be hesitant to embrace AI in your medical interactions?


Drop me a line at galland@rationaloptimistsociety.com. I read all my correspondence and usually respond.


In the Addendum just below, you'll find details on some of the leading companies working in the space, a reader Q&A, and sources.



ADDENDUM


Companies Building the Future


Following are more details on eight companies that, in my view, best represent where AI-in-medicine is headed. Unlike the companies featured in our Playing God Deep Dive, where most of the action was in small, CRISPR-pure biotechs, the AI-in-medicine landscape includes everything from data platforms to infrastructure giants.


Recursion Pharmaceuticals (RXRX)

Founded in 2013 by Chris Gibson (a young doctor-scientist who developed the idea during his research at the University of Utah) along with co-founders Dean Li and Blake Borgeson, Recursion took an unusual path. Instead of picking a disease first and then searching for a medicine to treat it, the company treated living cells with thousands of different chemicals. It then took detailed pictures of those cells and used artificial intelligence to spot small changes in how the cells looked. Those visual changes pointed to real effects inside the cells caused by exposure to the treatments, which have more recently included gene therapies.


Recursion called this image-based method “phenomics.” Over time it has built a database of tens of petabytes of cell images (a petabyte is a million gigabytes). By turning the shape and appearance of cells into a kind of map, the company aims to better understand how biology works and to find new medicines faster and more systematically.


In 2024, Recursion merged with UK-based Exscientia in a roughly $688 million deal—the largest AI drug-discovery combination ever—pairing Recursion's biology engine with Exscientia's AI chemistry design. Nvidia has taken a $50 million equity stake.


Primary Focus: More than 10 clinical programs spanning oncology, rare disease, and inflammation, with its platform increasingly licensed to partners including Roche, Bayer, and Merck. Its furthest-along drug, REC-4881, has posted positive Phase 2 results in a hereditary pre-cancer condition, cutting patients' polyp burden by a median of 43%—early but real evidence that a computer-guided pipeline can produce medicines that work in people.


Isomorphic Labs (subsidiary of Alphabet - GOOGL)

Spun out of DeepMind in 2021 and led by Demis Hassabis—winner of the 2024 Nobel Prize in Chemistry for the AlphaFold breakthrough that underpins the company—Isomorphic is Alphabet's bet that AlphaFold can be turned into a drug-discovery engine at industrial scale.


In 2024, Isomorphic signed deals with Eli Lilly (worth up to $1.7 billion) and Novartis (up to $1.2 billion) to design drugs using its technology. In 2025 it raised a further $600 million from outside investors to move its own drugs toward the clinic, with another $2.1 billion in Series B funding added in May 2026.


Primary Focus: Next-generation small molecule drug design, with initial programs in oncology and immunology now approaching their first human trials. Given its parentage—and a Nobel laureate at the helm—this is arguably the most well-resourced AI drug-discovery effort in the world.


Schrödinger (SDGR)

The old guard of computational drug discovery. Founded in 1990, long before "AI drug discovery" was a buzzword, Schrödinger built a platform for simulating how drug molecules interact with proteins using fundamental physics.


With the arrival of modern AI, Schrödinger has combined its physics-based simulations with machine learning, giving it a uniquely rigorous platform.


Primary Focus: Its own drug pipeline (with programs in oncology and metabolic disease) plus software and partnership licensing to roughly a third of the world's top 20 pharma companies.


Abridge (Private)

Founded in 2018 by Shiv Rao, Abridge is the leading ambient AI scribe, deployed across Kaiser Permanente, Johns Hopkins, Emory, and dozens of other major health systems.


In early 2025, Abridge raised $250 million at a $2.75 billion valuation—and then, later that year, raised again at $5.3 billion. It's one of the fastest-growing healthcare companies ever built.


Primary Focus: Ambient documentation across primary care, emergency medicine, and specialty practice, with expansion into AI-powered coding and revenue cycle management.


Nvidia (NVDA)

You can't write about AI in medicine without mentioning Nvidia. Every one of the AI models above—AlphaFold, the Mayo ECG system, Recursion's imaging pipelines—runs on Nvidia GPUs.


Beyond infrastructure, Nvidia's Clara and BioNeMo platforms have become the default toolkits for biotech and hospital AI development. Its Parabricks software has compressed genome analysis times from days to minutes.


Primary Focus: Picks and shovels for the entire industry. If you believe AI-in-medicine is a durable megatrend, Nvidia is the one certain way to own all of it at once.


Tempus AI (TEM)

Founded in 2015 by Groupon entrepreneur Eric Lefkofsky, Tempus took a different path after Lefkofsky’s wife was diagnosed with cancer. He was stunned to find that, despite all the technology in his life, none of it was being applied to her treatment.


Tempus is essentially a data platform. It sequences tumors, collects real-world outcomes from millions of patients, and uses AI to help oncologists find the treatment that best matches each specific case. It went public in June 2024.


Primary Focus: AI-guided precision oncology, with expanding programs in cardiology and neuropsychiatry. Tempus is also becoming a key partner to pharma companies running clinical trials.


As a caveat, I personally own shares in Tempus, but that played no role in my mentioning the company here.


Viz.ai (Private)

Founded in 2016 by Dr. Chris Mansi, a British neurosurgeon who watched too many patients arrive too late for stroke treatment. Viz built an AI that reads CT scans in the ambulance or ER, identifies large-vessel occlusions—the kind of stroke where minutes mean brain cells—and alerts the on-call neurologist directly on their phone.


A 2021 study showed Viz.ai reduced time-to-treatment by about 40 minutes on average. In stroke care, that can make the difference between a full recovery and a lifetime in a wheelchair.


Primary Focus: The company has expanded well beyond stroke, with FDA-cleared AI tools now covering pulmonary embolism, aortic dissection, and a widening list of time-sensitive conditions. It is quietly becoming the standard of care in acute imaging.


Xaira Therapeutics (Private)

This company started in 2024 with a huge $1 billion in early funding from major investors including ARCH Venture Partners and Foresite Capital. It is run by Marc Tessier-Lavigne, who used to lead Stanford University and was the top scientist at the big biotech company Genentech.


Xaira’s big idea is simple but bold: artificial intelligence has now gotten good enough to invent completely new proteins from scratch. These custom-made proteins (especially a type called antibodies) can become brand-new medicines. Traditional drug makers usually look for chemicals that already exist and hope they fit a disease target. Xaira wants to design the exact protein it needs, even for disease targets that no one has been able to drug before.


Primary Focus: De novo protein design for drugs that don't yet exist. The company is still very early stage. Many people in the industry see it as one of the biggest and most ambitious bets on using AI to create medicines that do not yet exist.


Some Questions and Answers


Will my doctor be replaced?

No. But their job will change.


Every serious analysis—including ones from physicians—has converged on the same answer. AI will take over the tasks of medicine that humans do poorly (monotonous pattern recognition, keeping up with a firehose of literature, integrating large volumes of data) and leave the tasks humans do well (empathy, judgment under uncertainty, advocacy for the patient) to doctors.


What I expect is the shape of the medical workforce will change. There will be more access to doctors, not less, because AI makes medical care cheaper and therefore more abundant. There will be far fewer billing specialists and administrators, because AI is far better suited to that work. And we'll probably see new roles emerge—"AI-validating physician," for example—that don't really exist yet.


Should I paste my medical records into ChatGPT?

With care, yes. But not blindly.


Four things to consider:


First, privacy. The free version of ChatGPT may use your inputs to train future models. If you're putting in sensitive health data, use the “private” mode most AI models now offer. That disables training and ensures your chats are not saved to a database.


Second, verification. Treat the AI's answer the way you'd treat a second opinion from a smart friend who went to medical school.


Third, collaborate. In the prompt, use a phrase like: “I have XYZ symptom, what possible causes should I discuss with my doctor?” You will then receive a full range of possible diagnoses, usually weighted from the most likely to the least. Feel free to push back and to ask follow-up questions; that’s how AI works best.


Fourth, awareness of limitations. AI is better at certain kinds of reasoning than others. It's extremely good at taking a long medical history and suggesting rare conditions a busy doctor might miss. It's worse at making judgment calls under high uncertainty.


As a relevant aside, in 2024 I developed a serious back issue. When I received the MRI image of my spinal column, I ran the image through ChatGPT, and it came back with a very specific and accurate diagnosis, which I verified with my doctor. So, for quick second opinions, it’s a handy tool.


Should I wait until AI medicine is more mature before using it?

Generally, no. The tools that exist today—ambient scribes, diagnostic-assistance AIs, pharmacogenomic flagging—are already net-positive for most patients. Waiting on the sidelines is waiting for a version of the technology that might be marginally better, while you forgo real benefits today.


The exception is anything billed as a "cure" or "diagnosis" from a consumer app without regulatory clearance. That's where caution is warranted. If a product is making strong medical claims and isn't FDA-cleared, that's a red flag.


In researching this article, for instance, I came across a number of phone apps purporting to be able to diagnose skin conditions. It’s coming, but for now, be skeptical of such claims.


Sources


 
 
bottom of page