A duty to consult

"Today it would be malpractice for a doctor to make a diagnosis or prescription without checking AI." Those words, by Robert F. Kennedy Jr.'s account, came from Sam Altman, the chief executive of OpenAI. Mr. Kennedy, the health secretary, repeated them to the Make America Healthy Again Summit in Washington on September 29. Artificial intelligence, he added, can give patients "a second opinion that is much better informed than any doctor in the country," and could "free us from medical tyranny."

Malpractice is judged against what a reasonable doctor would have done, so the remark amounts to a proposal to move the standard of care. Yet the week's evidence showed tools that do well at some narrow tasks, poorly as all-purpose oracles, and that leave the patients they are meant to impress unimpressed. The standard should move one tool at a time, as evidence arrives: until a particular tool has been shown to help patients in ordinary practice, declining to consult it should not be negligence, and a doctor who overrules one should be protected rather than punished.

At the same summit Mehmet Oz, who runs Medicare and Medicaid, said that the health system was "hurting people" by not using AI more, The Hill reported. A day earlier McKinsey, a consultancy, estimated that AI could already perform the work behind roughly one in five American outpatient claims, mostly simple office visits and the reading of scans, though its figures describe what AI could do rather than care it now delivers. On September 30 six large physician groups, among them the American Medical Association and the American College of Surgeons, replied that statements casting AI as better informed than doctors "diminish physician expertise and risk undermining patients' trust."

The government's own machine made their case for them. On September 29 the White House launched America.gov, a chatbot built on the Gemini and Grok models that answers questions from federal websites. Tested by Dr. Céline Gounder of CBS News, it advised against vaccinating newborns for hepatitis B, following a CDC page whose change a federal judge had paused in March, and never mentioned the court order. A University of Florida study published the same day tested five AI tools for searching the medical literature. Asked a question once, even the best of them typically found well under half of the relevant studies, and between about half and four-fifths of single queries turned up none of the core evidence at all. A doctor who checks such a tool and feels reassured may simply have been told nothing.

What a tool worth consulting looks like came from dermatology. At two British hospitals an AI system reviewed photographs of skin lesions from thousands of patients urgently referred for suspected cancer, and discharged between a quarter and almost a third of them with no clinician looking. It caught more than 98% of invasive cancers and made room for more than 8,500 extra appointments; surveillance turned up six cancers it had missed. The task was defined, the error rate measured, patients agreed to it and someone counted the misses. Even so, the results so far are a congress presentation rather than a journal paper; most tools doctors will be urged to consult have far less behind them.

Patients, meanwhile, are not clamoring for it. In a study published on October 1, 1,030 Americans were shown advertisements for fictitious family doctors. Those who said they used AI were rated less warm and less competent than those who said they never did, and people were less willing to book with them or follow their advice; the penalty was larger among adults under 45. The effects were modest, and measured intentions rather than visits. Still, doctors are now told by the health secretary that skipping AI is malpractice and by patients that using it counts against them.

California offered the clearest test of who gets the last word. Governor Gavin Newsom signed legislation that his office billed as "Empowering doctors to make the final call," ensuring that clinicians can use their own professional judgment when AI tools are used in patient care. On the same day he vetoed a companion bill that would have protected clinicians from retaliation for rejecting an AI system's recommendation, because, he said, the state Labor Commissioner "lacks the medical expertise to enforce the bill's anti-retaliation rules," CalMatters reported. The governor, said Sandy Reding, president of the California Nurses Association, "vetoed the bill that would have protected us for using our judgment." Clinicians in California now hold the final call in law, with no protection for making it.

The strongest objection is that doctors make errors too, some of them fatal, and that medicine has a long record of resisting methods that later saved lives. If a cheap tool catches mistakes, refusing to look puts pride ahead of patients. That is right about the direction and wrong about the route. The standard of care has absorbed new tools after evidence showed that they helped, for a given task, in ordinary hands. A general duty to check AI would cover the government's chatbot and a literature search that returns nothing as surely as a dermatology system that counts its misses.

Medical staffs should list which AI tools are approved for which tasks, and state in writing that departing from a tool's advice, with the reason documented, is good practice. Hospitals and physician employers should put into bylaws and contracts the protection California declined to put into law: no clinician disciplined for overruling a machine in good faith. The six societies that objected this week should go further and publish, task by task, which tools have earned a place in the standard of care, with evidence like the dermatology data as the bar. California's legislators should send the governor an anti-retaliation bill enforced by the medical and nursing boards, which have the expertise he says the Labor Commissioner lacks. And the White House should test its own chatbot before the administration tells Americans that machines know better than their doctors.

Mr. Altman may one day be right that diagnosing without checking AI is malpractice. When that day comes it should be because a particular tool was tested and shown to help patients, and the checking should start with the government's own.