Stanford patients reveal gaps in health AI rollout

Stanford University has a long history of technical innovation — the computer mouse was invented there, and so was Google. Now the institution is producing a growing number of artificial intelligence tools for health care. But for the past 18 months, the hospital has taken a different step: asking patients about new AI tools before they go live.

A Patient Panel Before the Rollout

The medical center handpicked a group of patients and caregivers to review AI tools in development. Eric Gries is one of them. He was the primary caregiver for his wife while she was on a left ventricular assist device (LVAD) and later received a heart transplant. After that, he became the temporary caregiver for his brother-in-law, who also underwent a heart transplant.

Gries’s experience gives him a direct stake in how AI is introduced into clinical settings. The panel’s job is to flag issues that engineers or doctors might overlook — things like how a tool affects a patient’s sense of trust, or whether its recommendations make sense to someone who’s actually living with the condition.

Related: Durbin says Trump lets Big Tobacco target children

The university isn’t alone in this approach. Other academic medical centers have started similar advisory groups. But its program stands out because of the sheer volume of AI tools it’s developing and the speed at which they’re moving from research to bedside.

The group’s feedback has already changed how some tools are presented to patients. For example, one AI system that flags potential complications was redesigned after committee members pointed out that the language it used could sound alarming without context.

Why Patient Input Matters for AI Safety

Such tools in health care are often built on data from electronic health records, imaging, or wearable devices. But those datasets can miss the human side — how a patient interprets a risk score, or whether a recommendation fits their daily life. The patient panel is meant to catch those gaps.

Related: Eli Lilly acquires three small vaccine firms

“It’s easy to get carried away with the technology,” one person familiar with the program said. The advisory board forces developers to slow down and consider real-world consequences.

Some experts in medical AI ethics have raised concerns that patient panels can be tokenistic if their feedback isn’t actually used. The initiative appears to avoid that problem by integrating suggestions before deployment. But it’s too early to tell whether this model will scale beyond a handful of institutions.

Gries said his own experience as a caregiver made him particularly attuned to the emotional weight of medical decisions. When a machine recommends a course of action, he noted, the patient still has to live with the outcome.

Related: Exploring the Multifaceted Benefits of Healthy Food

The group meets regularly and reviews tools at different stages of development. Not every suggestion is adopted, but the institution says the process has already flagged several potential problems that would have been missed in a purely technical review.

The university’s approach represents a shift from the traditional model, where these systems are tested in controlled environments and then released to clinicians. By involving patients early, the hospital hopes to avoid the kind of backlash that has followed some high-profile AI failures in other parts of medicine.

The computer mouse was a breakthrough because it made technology more accessible. Patient involvement may do the same for health AI, making it more accessible.