Terence Tao posed a thought experiment about whether people would accept an AI-designed cancer drug that had passed clinical trials even though no one understood how it worked. The debate divided social media, and an oncologist reminded him that medicine has used drugs with poorly understood mechanisms for years

Terence Tao posed a thought experiment about whether people would accept an AI-designed cancer drug that had passed clinical trials even though no one understood how it worked. The debate divided social media, and an oncologist reminded him that medicine has used drugs with poorly understood mechanisms for years

On October 9, 2026, Terence Tao gave a lecture at Caltech titled “Math 2.0”. One slide imagined AI finding a cancer drug whose mechanism no one understood. The next day, a screenshot drew more than 9 thousand likes and 1,2 million views on X. Oncologist Vinay Prasad responded that medicine already uses drugs whose mechanisms of action are incompletely understood.

Tao, a recipient of the Fields Medal, the highest honor in mathematics, was developing an argument from his July lecture: mathematics is entering an era in which AI generates proofs faster than people can absorb them. Finding a solution used to deepen understanding of the field along the way; pursuing the answer alone now breaks that connection. In pure mathematics, this shift is safe: the program Lean checks proofs line by line and allows no errors through. Tao’s thought experiment illustrates why even this kind of verification is insufficient in the real world. AI selects a cocktail of compounds, Lean verifies its mathematical model, and the cocktail passes the final, third phase of human trials. Yet it remains uncertain whether the cocktail cured the cancer or exploited a weakness in the testing procedure, and no one understands its mechanism of action.

“Could the AI’s solution to this request be misaligned because it exploits a weakness in the testing procedure itself? Before having this cocktail injected into your bloodstream, would you want to know that at least one human oncologist understood how the drug worked?”

Tao had written about the same risk in early September, using the Navier-Stokes equations, which describe fluid motion, as an example. An AI-generated solution that bypasses the human process of discovery deprives the field of ideas that would have emerged during the search. In the cancer slide, the same error puts a patient’s health at stake rather than a mathematical proof.

Commentator Andrew Curran posted a photograph of the slide, arguing that if a drug works, a patient with stage four cancer does not care whether anyone understands its mechanism. User kenbwork agreed: AI should not be administering drugs whose effects are poorly understood. Another critic responded more sharply, accusing the mathematician of “laying the rhetorical groundwork for banning AI cancer drugs” and saying he would “get people killed.” All were responding to Curran’s paraphrase. None discussed the slide’s specific concern that AI might exploit a flaw in the testing procedure.

Prasad, a hematologist-oncologist and professor at the University of California, San Francisco, who twice headed the FDA’s biologics division in 2025–2026 and twice resigned following criticism, responded to Curran’s version. Six months earlier, he had established one pivotal trial instead of the previous two as the standard. The mechanisms of action of many cancer drugs remain incompletely understood: there is no unified theory of how doxorubicin works, and no one knows why sevoflurane gas causes unconsciousness.

“If AI creates a combination cancer drug whose mechanism no one understands, but it passes a full randomized phase three trial, there will not be a single cancer patient on Earth who says, "No, thank you." We currently accept far less evidence.”

Prasad is correct about existing medical practice: drugs are used even when their mechanisms are incompletely understood. Tao’s question remains unanswered, however: did the trial establish that the treatment cured the cancer, or that it could exploit the testing procedure? Tao proposes open mathematical models for interpretability and safety, which he says will be available soon. How quickly AI-designed drugs reach patients who cannot afford to wait depends on what medicine accepts as sufficient verification: trial statistics or an understood mechanism.