Gaming a Cancer Trial Is 'Cheap and Easy,' Says a Former Top FDA Official

On October 10, the oncologist Vinay Prasad told Terence Tao, a Fields medalist and about as famous as a living mathematician gets, that he did not understand cancer medicine. Tao had suggested that patients would want a human expert to understand an AI-derived cure before they took it. Prasad’s reply was that oncology runs on drugs nobody can fully explain, that the field still argues over how docetaxel works in prostate cancer, and that if AI “conjures up a cocktail to cure cancer and we don’t know how it works, but it passes a phase 3 trial, you can bet we will be on our knees thanking it.” The randomized trial was the safeguard. Understanding was optional.

The next day he published a sequel that takes the safeguard apart. Some readers had argued that Tao meant something darker than an unexplained cure: an AI that cheats its way to a phase 3 win for a drug that does not work. Could that happen? Prasad’s answer was “Yes,” followed by the deadpan that carries the rest of the post. “You don’t need AI at all,” he wrote, “in fact what a waste of tokens.” Gaming a trial, by his account, is “cheap and easy to do.” And then, flatly: “We put very little effort to catching this.”

The first trick on his list runs on attrition. Pair a toxic drug with an endpoint vulnerable to what statisticians call informative censoring, he suggests, and let the side effects do the work. Patients who quit the experimental arm stop being counted at the moment they leave, and the standard analysis treats them as though they would have fared like the patients who stayed; if the ones walking out are the sickest, as they tend to be on a punishing drug, the arm left behind looks healthier than the drug made it. Progression-free survival has a cousin of the same flaw, because a progression nobody scans for never enters the data. “What if your drug made sicker people in the experimental arm skip future scans?” Prasad asks. He offers no proof that any sponsor has done it, only that some drugs “smell like it.”

The second trick works on the comparison itself, and it reaches even overall survival, the endpoint oncologists treat as the one that cannot be fooled. A fake survival result, Prasad writes, can come from “a subpar control arm or inappropriate post protocol care,” which is to say that if the patients in the comparison group get weaker treatment than a good oncologist would give them, during the study or after it ends, the new drug’s advantage grows without the drug getting any better. The rest of his list is shorter: rewriting the statistical analysis plan after looking at the data, missing-data methods that decline to assume the worst about patients who dropped out, and, rarely, “cooking the books,” which he notes has been caught a few times.

The weak comparator is the one trick with a published audit behind it. Talal Hilal and colleagues at the Mayo Clinic went back through the trials behind FDA cancer drug approvals from January 2013 through July 2018 and reported in JAMA Oncology that of 96 approvals resting on randomized trials, 16, or 17 percent, used a control arm the investigators judged suboptimal. In 10 of those, the comparator was a drug already shown to be inferior to an available alternative, or the protocol barred combinations; in 4, the investigator’s choice of treatment was restricted; in 2, patients were re-exposed to a therapy already known not to help the second time around. The audit makes no finding of intent, and for the patients randomized to the control arm, intent is beside the point, since a new drug measured against a comparator already known to lose will tend to look like a winner. The FDA approved all 16. It was the Mayo Clinic, not the agency, that went back and counted.

The comparator problem sits on top of an older one that Prasad has been documenting for years. In a 2015 analysis with a colleague, he counted 36 cancer drugs the FDA had approved on a surrogate marker, a shrinking tumor or a delayed scan standing in for a longer life. By the end of the study period, 18 had been found not to improve survival and only 5 had been shown to; the other 13 were untested or unreported. A separate Harvard analysis by Bishal Gyawali, Spencer Phillips Hey and Aaron Kesselheim looked at 93 cancer drug indications granted accelerated approval between 1992 and 2017 and found that only 19 had confirmatory trials showing an overall survival benefit. So the randomized trial Prasad held up against Tao protects patients only when three conditions hold: it measures whether people live longer, it measures that against a fair comparator, and someone checks that the reported numbers match what happened. The approval record shows how often the first condition was waived. Hilal’s audit shows how often the second slipped through.

The machine, in the end, barely figures in the follow-up. One commenter, Kath, posted an exchange in which a chatbot told her “maximizing text probability is my architecture,” and concluded that “AI is not trained to care or know about whether the drug works in reality.” She meant it as a warning about language models. It describes a sponsor holding a toxic drug and a censoring-friendly endpoint about as well.

Prasad was not writing from the outside. He took over the FDA’s Center for Biologics Evaluation and Research in May 2025, and Commissioner Marty Makary later named him the agency’s chief medical and scientific officer as well; he left at the end of April. Most cancer drugs are reviewed by a different FDA center, so the control arms in Hilal’s audit were never his to approve. The verification gap belongs to the whole agency, though, and in his new post he names it precisely: “there is no system that verifies if 100% of primary study reports actually reflect what happened to a person.” It is an accurate sentence, and he wrote it about five months after leaving the one job at the agency that carried his name on its science.

Sources

1. Vinay Prasad, “Can AI ‘hack’ a cancer clinical trial” (Oct 11, 2026)
2. Vinay Prasad, “Terrence Tao does not understand cancer medicine” (Oct 10, 2026)
3. The ASCO Post on Hilal et al., control-arm quality in randomized oncology trials leading to FDA approval, JAMA Oncology
4. National Center for Health Research: Kim and Prasad on FDA cancer approvals by surrogate endpoint (2015)
5. Lown Institute on Gyawali, Hey and Kesselheim, accelerated approval of cancer drugs, JAMA Internal Medicine
6. BioPharma Dive, “Vinay Prasad, controversial FDA leader, to again depart agency” (March 2026)