Americans accept AI as a healthcare helper, not as the final decision-maker

Americans are not giving healthcare AI a simple yes or no. A new Pew Research Center survey finds a layered position: people are often willing to let AI translate, schedule, take notes or analyse information, but they prefer a human to review the work and become much less receptive when the system makes a diagnosis or decides an insurance claim.[1][2][3]

The study surveyed 3,488 adults from 22 to 28 June 2026 through Pew’s American Trends Panel. That design represents the U.S. adult population, but it measures stated attitudes rather than behaviour inside a hospital or clinic. The results are still useful because they reveal what kind of deployment the public regards as legitimate.[1][2][3]

Oversight matters more as stakes rise

The pattern is more revealing than any single percentage. Translation and scheduling attract broad acceptance, including some willingness to let AI act alone. Scan analysis is also widely accepted, but almost entirely with a human final review. Diagnosis and insurance decisions produce the strongest resistance.[1][2][3]

This suggests that “human in the loop” is not one control. For a scan, people may expect a clinician to inspect and interpret an AI result. For an insurance claim, a nominal reviewer may not satisfy concerns about incentives, appeal rights and accountability. The meaningful design question is who reviews the output, what evidence they see and whether they can reverse the decision.[1][2][3]

Comfort and predicted outcomes diverge

When asked about their own doctor using AI for diagnoses and treatments, 60% said they would be uncomfortable and 30% comfortable. Yet the public is almost evenly divided on the eventual effect on health: 37% expect better outcomes, 36% worse and 26% little difference. People can therefore believe the technology may produce aggregate benefits while still disliking its use in their own care.[1][2][3]

Age differences are especially striking. Among adults aged 18 to 29, 45% now say healthcare AI will worsen patient outcomes and 31% say it will improve them. In 2022, the balance was reversed: 44% expected improvement and 29% worsening. Adults aged 65 and older moved in the opposite direction, with 42% expecting better outcomes in 2026 and 26% worse.[1][2][3]

Privacy and relationships are the public’s weak points

Fifty-one per cent think AI will make the security of personal health information worse, compared with 8% who think it will improve it. Forty-seven per cent expect the patient-doctor relationship to worsen and only 10% expect improvement. Views are more positive on access: 31% think AI will improve availability of services, while 20% expect it to worsen.[1][2][3]

Those results imply that adoption strategies built only around accuracy or efficiency will miss central concerns. A system can improve a diagnostic metric and still erode trust if patients do not know when it is used, where their data travels or whether a clinician is listening to them rather than to a score.[1][2][3]

Trust is concentrated closest to the patient

Among the institutions Pew asked about, doctors received the most balanced confidence ratings. Twenty-six per cent expressed a great deal or quite a bit of confidence that doctors would ensure AI serves patients’ interests, 34% had some confidence and 25% had little or none. Low confidence was far higher for health insurers, the federal government and AI companies.[1][2][3]

That does not give clinicians automatic legitimacy; it gives them a heavier responsibility. Health systems need to preserve a visible line of accountability from the model’s output to a person who can explain, question and correct it.[1][2][3]

The survey’s practical message

The public is not rejecting healthcare AI. It is drawing boundaries around autonomy. Acceptance is highest where the task is administrative or supportive, and where human review is explicit. Resistance rises when the system controls diagnosis, treatment, payment or sensitive data.[1][2][3]

For developers and providers, the lesson is to specify the decision architecture rather than market “AI in healthcare” as a single category. Which task is automated, what the reviewer does, how patients are informed and who handles appeals may matter as much to adoption as the model’s benchmark score.[1][2][3]

Report an error or suggest a correction