As medical AI use grows, industry payments to providers remain opaque

STAT News – Monday, September 22, 2025
By Katie Palmer

Artificial intelligence has danced around the fringes of prostate cancer for years, but in 2024, it got a profile-raising boost: An AI tool that could use biopsy images and clinical data to predict therapy benefits and a patient’s prognosis was recommended in a set of widely used guidelines for oncology care.

“I remember being so impressed,” said David-Dan Nguyen, a urology resident at the University of Toronto who focuses on prostate cancer. But he was also curious: How exactly does a new device make it into clinical standards? He started to look into whether its manufacturer had made any payments to hospitals or doctors in the leadup to its guideline inclusion — to support research, perhaps, or to cover fees for consulting or marketing.

When he looked at Open Payments, the federal database of industry payments to clinicians and teaching hospitals, he didn’t see any. “That really started sparking a thought in the back of my mind,” said Nguyen, who decided to keep looking for payments made by manufacturers of AI devices authorized by the Food and Drug Administration. In a research letter published Monday in JAMA, he and his co-authors found fewer than 10% of nearly 850 AI devices were linked to payments between 2017 and 2023.

“We are aware of blind spots that exist for AI devices and their deployment in medicine,” said Nguyen. “But there are more than the ones that we think about already — one being conflicts of interest.” As AI becomes rapidly adopted in medicine, it’s important to understand whether financial incentives have the potential to color decisions about technology in clinical care. But the researchers’ findings suggest that gaps in disclosure requirements may result in missed payments made by AI manufacturers.

“AI devices often come to market with limited testing, so research funding could, in principle, help fill critical evidence gaps. But without transparency, those payments risk blurring into marketing rather than science,” said Ken Mandl, a Harvard Medical School professor who directs the Boston Children’s Hospital Computational Health Informatics Program.

Some AI device manufacturers likely didn’t appear in the federal database because they simply didn’t make any payments to providers and hospitals. “But there’s also a large majority who likely had these relationships, but did not have to report these payments,” said Nguyen.

Manufacturers are only required to share payments to health providers with the Centers for Medicare and Medicaid Services when they fall into specific categories — ones that apply most naturally to traditional drugs and medical devices. Reportable payments apply to devices that are directly reimbursed by Medicare, Medicaid, or Children’s Health Insurance Program and that require premarket approval or notification by the FDA.

But most medical products that use artificial intelligence do not go through the FDA; the prostate cancer device, for example, wasn’t authorized by the FDA at the time Nguyen got curious about it. Even among FDA-authorized AI devices, a very small subset are reimbursed in the traditional sense. There’s also a fair amount of gray area in the detailed reporting criteria that could make it unclear to AI device manufacturers whether they’re required to disclose payments.

Among payments that were disclosed, “it was striking how quickly it grew in the six years that they studied it,” said Rita Redberg, a University of California, San Francisco, cardiologist and device safety researcher. In 2017, industry payments totaled $17.3 million; by 2023, they had reached $24.6 million. Increases in general payments — those that do not go toward research, but things like consulting fees, gifts, food, and entertainment for health providers — drove most of that increase, growing from $6.6 million to $13.3 million.

Payments from the AI industry totaled $120 million over the six-year period. Nearly half, $59 million, were associated with cardiovascular devices, followed by $40 million for radiology devices. “The payments usually track how lucrative the market is,” said Redberg, “and that’s what this did, too.” Typically payments for devices are highest for cardiovascular and orthopedics, but radiology devices have long accounted for more than three-quarters of FDA-authorized AI devices, followed by cardiovascular devices.

While those categories absorbed the biggest payments, “influence still extends across a wide swath of clinicians and hospitals,” said Mandl. The concentration of payments, and their reach, “makes transparency and compliance with reporting all the more urgent.”

Payments from radiology device manufacturers might escape reporting because of the way they’re paid for, said letter co-author Genevieve Kanter, an associate professor of public policy at the University of Southern California who researches conflict of interest.

“When you’re buying big equipment or software that’s going to be used at a hospital, there’s no splitting out that reimbursement,” said Kanter. A health system might purchase an AI device that’s part of a bundled payment for interpreting a radiological image, for example. Cardiology devices are more frequently provided directly by a provider to a patient and billed to insurance; an implanted heart failure monitor with an AI-based algorithm to interpret its output would qualify more obviously for payment disclosure.

Nguyen and his co-authors call to adapt payment transparency policies to the reality of this growing class of AI devices. “No one here is trying — or few people are trying — to hide payments,” he said. “I think if you have the foresight to present these ties transparently, you avoid the risk of future harm and loss of trust in the care that you’re providing.”

Kanter suggested that payments reporting criteria could expand to capture products that aren’t subject to FDA authorization. Also worth consideration: payments to recipients other than health care providers. “When we think about how these technologies get adopted, it’s through a technology assessment committee at the hospitals,” said Kanter. Payments to non-physician health care administrators in those groups are not included in current reporting requirements.

“The real risk is unseen commercial forces shaping both how AI tools are built and how they’re adopted, tilting decisions toward profit over patients,” said Mandl. Until underreporting of payments from AI manufacturers is addressed, “that gap leaves patients, clinicians, and policymakers in the dark about potential conflicts of interest.”

Access this article at its original source.