Hospitals Race Ahead of AI Oversight as FDA Devices Top 1,000

What happened

A Forbes Business Council essay is circulating a leadership concept called the "Trusted Cognitive Hospital," arguing that hospitals should measure responsible AI not by how many algorithms they buy but by whether they preserve clinical judgment, protect patient data, monitor performance and maintain clear accountability 1. The essay is framed as opinion rather than investigative reporting, but it lands at a moment when the underlying numbers make its argument hard to dismiss. Across regulatory filings, hospital surveys and peer-reviewed research, a consistent picture emerges: AI has moved from pilot project to infrastructure inside American hospitals faster than the systems meant to govern it have matured.

The clearest evidence of scale comes from the FDA's own device database. MedTech Dive's analysis found 950 AI- or machine-learning-enabled devices authorized as of August 7, 2024, up from just six in 2015 and 221 in 2023 alone 6. A follow-up legal analysis put the count at 1,016 by March 2025 7, and the American Hospital Association, in a letter to the FDA, cited more than 1,240 authorized devices with most clearances coming in just the last three years 8. A separate industry summary corroborates the trajectory, noting the FDA cleared its first AI device in 1995 but didn't see broad uptake until 2015, after which authorizations passed the 1,000 mark by March 2025 18.

The evidence gap behind the numbers

What the authorization count does not capture is confidence in real-world performance. A JAMA Health Forum study of 691 AI/ML devices cleared through 2023 found that FDA decision summaries frequently omitted basic evidence: 46.7% did not describe study design, 53.3% did not report training sample size, and 95.5% provided no demographic information 1011. Only six devices — 1.6% — cited data from randomized controlled trials, and just three reported patient outcomes 10. The same analysis identified 489 adverse events tied to 36 devices, including one death, and found 40 devices had been recalled a combined 113 times 1011. A separate 2024-focused study found modest improvement — cybersecurity considerations appeared in 54.2% of summaries and predetermined change plans in 16.7% — but concluded that transparent performance and demographic reporting remained limited even in the most recent cohort 9.

That gap is compounded by how devices get to market. Roughly 97% of authorized AI devices have gone through the FDA's 510(k) pathway, which relies on similarity to an already-cleared predicate device rather than new clinical trial data 6891018. More than three-quarters of all authorized devices are radiology products 6918, reflecting the fact that imaging generates the large, structured datasets that machine learning models are best suited to exploit.

Hospitals are already deploying, unevenly

Separate from FDA device counts, hospital-level survey data shows adoption is already widespread. A federal data brief from the Assistant Secretary for Technology Policy found 71% of hospitals reported using predictive AI integrated with their electronic health records in 2024, up from 66% the prior year 1213. A companion report from UPMC's Center for Connected Medicine and KLAS Research similarly found most health systems now using AI in some form 5. Adoption is sharply uneven: system-affiliated hospitals reported predictive AI use at 86%, versus 37% for independent hospitals, and urban facilities outpaced rural ones 1213. The fastest-growing uses were administrative — billing automation jumped from 36% to 61%, and scheduling assistance from 51% to 67% — while higher-risk clinical uses like treatment recommendations grew far more slowly 1213.

Generative AI is following a similar path. A JAMA Network Open survey of 2,174 nonfederal acute-care hospitals found 31.5% were already using generative AI tied to their EHR in 2024, with another 24.7% planning to within a year — suggesting more than half of U.S. hospitals could be using it by the end of 2025 1415. The same study flagged a digital divide: hospitals with weaker predictive-AI evaluation practices were less likely to adopt generative AI responsibly, meaning some organizations may be moving fastest with the least oversight in place 1415. Separately, the Washington Post's health newsletter noted that more hospitals are integrating AI into records systems even as personnel questions swirl at HHS 2.

Regulators are trying to catch up

The FDA's response has been to regulate AI's capacity for change rather than treat every update as a new product. In August 2025, the agency finalized guidance on Predetermined Change Control Plans, allowing manufacturers to pre-specify future modifications, the validation method for those changes, and an assessment of their risk — all reviewed once, up front, rather than through repeated new submissions 1617. The guidance applies across the 510(k), De Novo and PMA pathways and includes worked examples, such as patient-monitoring and feeding-tube-placement software, illustrating when a change falls inside an authorized plan versus when it requires new FDA review 17.

Beyond the FDA, oversight is fragmented across agencies. ONC's HTI-1 rule imposes transparency requirements on predictive algorithms embedded in certified health IT, which underlies systems used by more than 96% of hospitals 19. ONC's decision-support resource guide further directs developers to document validity, reliability, fairness, safety and governance for predictive tools, and to disclose how models are locally validated and monitored 20. Congress and agencies outside health care are moving in parallel: ARPA-H recently committed $62.7 million to fund AI agents for heart failure care under its ADVOCATE program, explicitly noting that Medicare and the FDA are still developing regulatory pathways for AI that can prescribe drugs directly to patients 4.

Where the reporting agrees

Across sources — Forbes commentary, MedTech Dive's device analysis, FDA guidance documents, JAMA research and federal hospital surveys — there is no real dispute about direction: AI authorizations and hospital adoption are both rising sharply, radiology dominates the device landscape, the 510(k) pathway dominates clearance routes, and evaluation and governance practices lag behind deployment 689101218. Every source that addresses adoption also flags unevenness — between large and small hospitals, system-affiliated and independent facilities, urban and rural, well-resourced and under-resourced — as a persistent feature rather than a temporary lag 12131415. And every technical source, regulatory or academic, agrees that FDA clearance is a market-access threshold, not proof of clinical benefit; the JAMA studies make this explicit through their evidence gaps, while the FDA's own PCCP guidance implicitly concedes that initial clearance cannot anticipate everything a deployed model will do 101617.

Where it doesn't

The most visible discrepancy is the device count itself, and it is a function of timing rather than contradiction: MedTech Dive's 950 devices as of August 2024 6, the 1,016 figure from a legal analysis dated March 2025 7, and the AHA's letter citing over 1,240 8 are sequential snapshots of a constantly updated FDA database, not competing claims about the same moment. Readers should treat the FDA list as a moving count, not a fixed fact.

A sharper divergence is in framing rather than figures. Forbes' essay treats responsible AI primarily as a leadership and culture problem — governance, trust, human-centered design 1 — while the JAMA Health Forum and PMC studies treat it as an evidence and transparency deficit measurable in missing trial data, missing demographics and unreported adverse events 1011. The federal hospital-survey data occupies a third position, describing governance as already partially institutionalized — 82% of hospitals say they evaluate models for accuracy, 74% for bias — while simultaneously showing that a meaningful share of hospitals, 15 to 21%, simply don't know whether such evaluation happens at all 1213. That gap between reported governance and actual certainty is not something the Forbes essay or the device-count reporting engages with directly; it emerges specifically from the ASTP/ONC and JAMA Network Open survey data 12131415.

There is also a difference in emphasis worth noting: ARPA-H's $62.7 million heart-failure program is reported as a forward-looking bet on patient-facing AI that can prescribe medication, explicitly ahead of settled FDA and Medicare rules for that use case 4 — a more aggressive posture than anything described in the device-authorization or hospital-survey research, which characterizes current AI largely as decision support rather than autonomous prescribing.

The reading the evidence supports

Taken together, the material does not support a narrative of either runaway hospital adoption of unsafe tools or adequate regulatory readiness — it supports a narrative of velocity mismatch. Authorizations, hospital deployments and generative-AI rollouts are compounding annually, while the evidentiary basis for most cleared devices remains thin and hospital governance, though widely reported, is inconsistently verified even by the hospitals themselves. The FDA's PCCP guidance is a genuine, substantive attempt to close part of that gap by regulating change over a device's lifecycle rather than only at launch, but it does not resolve the deeper problem the JAMA studies document: that most currently deployed AI was cleared with limited prospective evidence in the first place. The practical conclusion is that institutional leadership frameworks like the one Forbes describes are not marketing exercises so much as a necessary response to a regulatory system that authorizes products faster than it can compel proof they work as intended in the hospitals actually using them.