On October 10, the oncologist Vinay Prasad published a post telling one of the most decorated mathematicians alive that he did not understand cancer.
The mathematician was Terence Tao, a Fields medalist and about as close to a household name as pure mathematics produces. By Prasad’s account, Tao had floated a reasonable-sounding worry about artificial intelligence in medicine: before any of us swallowed a cancer cure that a machine had dreamed up, we would want a human expert who actually understood why the thing worked. Prasad’s response, published under the blunt headline that Tao “does not understand cancer medicine”, was that this instinct, however intuitive, rests on a fantasy about how the drugs in your oncologist’s cabinet got there.
It is worth stating Tao’s side at full strength, because it is the side most people are on. Do not trust a black box with your life. Demand a human in the loop who can explain the mechanism. In mathematics you can hand a machine’s proof to a person and have them check it line by line, and extend that habit to oncology and you get a commitment that sounds like simple prudence: no cure nobody understands.
Prasad’s answer is that medicine blew past that line decades ago, and nobody flinched.
Consider the drugs we hand out without apology. Adriamycin, the red-infusion backbone of chemotherapy regimens for half a century: Prasad’s needle is that if you think you know how it works, you have not read enough about it. Docetaxel, a standard in prostate cancer, where specialists still argue over whether it acts by jamming the microtubules that shuttle the androgen receptor around the cell or simply by halting cell division. Ibrutinib and venetoclax, two of the most important blood-cancer drugs of the last decade, with no unifying theory that explains where they work and where they fail. Then he reaches past oncology for the cleanest example of all: sevoflurane, the gas that puts you under for surgery. How general anesthesia erases consciousness is genuinely unsettled, and Prasad, like the rest of us, would still take it over being awake on the table.
“A drowning man will reach for the blade of a sword,” he writes, and the oncology ward is full of drowning men reaching, taking drugs supported by a response rate in a few dozen patients, or a signal in mice, long before anyone can draw the mechanism on a whiteboard. If an AI proposed a combination nobody understood and it then cleared a real randomized trial, Prasad argues, scarcely a cancer patient alive would turn it down, because understanding was never the thing standing between the patient and the harm. The randomized trial was.
Which is where the mathematician drops out of the argument entirely.
Because the safeguard Prasad is invoking, the hard trial that shows people actually live longer, is precisely the one the FDA’s cancer-drug program spent the last twenty years declining to require. We know this because Prasad is the person who documented it. In a 2015 analysis he ran with Chul Kim, 36 cancer drugs were approved on the strength of a surrogate marker, a shrinking tumor or a delayed scan rather than a longer life. Of those 36, 18 were later shown not to extend survival at all, and only 5 were ever proven to. A later cohort, led by Bishal Gyawali, Spencer Hey, and Aaron Kesselheim, followed cancer drugs granted accelerated approval across twenty-five years. Of 93 such approvals, only 19, about one in five, had a confirmatory trial that demonstrated a benefit in overall survival.
Tao is afraid of a cure no one can explain. The cancer shelf’s trouble runs the other way: a row of drugs everyone can explain and no one has shown will buy a patient a single extra month.
This is not a gotcha at Prasad’s expense, and it would be lazy to read it that way. He is not contradicting himself; he is exposing the worship of mechanism as hollow. It is the same case he has made for a decade, that mechanism is a story we tell ourselves, that surrogate endpoints are a convenience for industry and regulators, and that the only receipt that counts is how long the patient lives. He spent part of this year running the FDA center that clears vaccines and biologics before leaving the agency this spring to return to academia, and a reader could be forgiven for noticing that the standard he demanded in print is sterner than the one the building around him ever enforced.
The useful part of the argument is the monster it correctly identifies. The dangerous cancer drug of the next decade is not some inscrutable machine-designed cocktail that cleared a gold-standard survival trial and left the experts unable to narrate why. That drug, by the only test that matters, works. The dangerous drug is the fully legible one: a molecule with a tidy mechanism diagram in the package insert, approved because it shrank a tumor on a scan, sold for years, and never once proven to help anyone outlive the disease. We understand that drug completely, and the understanding has saved no one.
A JAMA Internal Medicine commentary once called the state of cancer therapy the best of times and the worst of times, and the line has aged well. Tao, worrying about the cures we cannot explain, is guarding a door no one is coming through. Prasad, pointing at the trial as the one thing that counts, has the data on his side. And the agency he just left, the one charged with demanding that trial, has by these counts waved through drug after drug without it, and will almost certainly do so again long after this particular argument is forgotten.
Sources
- Vinay Prasad, “Terrence Tao does not understand cancer medicine” (Oct 10, 2026)
- National Center for Health Research: Kim and Prasad on FDA cancer approvals by surrogate endpoint (2015)
- Lown Institute on Gyawali, Hey and Kesselheim, accelerated approval of cancer drugs, JAMA Internal Medicine (2019)
- Lehman and Gross, “An International Perspective on Drugs for Cancer: The Best of Times, the Worst of Times,” JAMA Internal Medicine (2019)
- Clinical Trials Arena: Vinay Prasad to leave the FDA’s CBER (2026)