The warning: a claim becomes "accepted wisdom" before it's checked
In April 2026, Harvard economist David Cutler and co-author Lev Klarnet argued in JAMA Health Forum that falling per-capita imaging volume had "allowed the U.S. to meet the need for imaging without an increase in radiologists" — and floated AI as a way to stretch the existing radiologist workforce further, especially in underserved areas.
Researchers at the Harvey L. Neiman Health Policy Institute, led by executive director Elizabeth Rula, PhD, pushed back hard. Their JACR editorial — covered by Radiology Business and AuntMinnie — called the assertion "counter to recent collective experience of the radiology community" and said it "ignores a decade of evidence." The core problem: the per-capita imaging data Cutler and Klarnet cited runs only through 2016, while Neiman HPI's own research projects total U.S. imaging use will grow 17% to 27% by 2055, driven by an aging population, and documents rising examination complexity, RVU workload and radiologist turnover in the years since.
The editorial's real warning isn't about one economist's paper. It's about what happens when an unverified claim gets repeated until it functions as fact — shaping residency choices, staffing plans, payer policy, and now, implicitly, how much AI capacity is assumed to fill a gap that may not exist the way it's described.
The same test applies to AI vendor marketing
If a workforce claim deserves scrutiny because it could shape policy, an accuracy claim on an AI-reporting vendor's homepage deserves at least as much — it directly shapes what a radiologist trusts to flag or miss a finding. The evidence base is not reassuring. An independent audit of 173 CE-certified radiology AI products from 90 vendors, published in European Radiology in 2025, found that 66% had any peer-reviewed supporting evidence at all — up from 36% in a 2020 baseline, real progress, but still leaving a third of commercial products with no published validation.
Even among products with published evidence, most studies stop at diagnostic accuracy on a single dataset: the same 2025 audit found only about 22–24% of supporting studies examined real clinical or economic impact, a share that has barely moved, while independently (non-vendor) authored studies actually declined as a share of the literature, from 49% to 45%. A statistical sanity-check of outperformance claims across medical-imaging AI papers more broadly — published as a 2025 preprint by Christodoulou et al. — found that more than 80% of papers introducing a new method claimed to outperform prior approaches, but a large share of those claims (86% for classification tasks, 53% for segmentation) had a high statistical probability of being false rather than genuine.
"FDA cleared" is not the same claim as "clinically validated"
Buyers often treat FDA clearance as a proxy for proven accuracy. It isn't designed to be one. A 2025 analysis of 2024 FDA authorizations, published in Biomedicines, found that 94.6% of ML-enabled devices cleared that year went through the 510(k) pathway — which requires showing substantial equivalence to an already-cleared predicate device, not an independent clinical trial — with radiology accounting for 74% of those clearances. Writing in Radiology: Artificial Intelligence, researchers Diab and Lotter made the distinction explicit: FDA validation rigor and public transparency about that rigor are separate questions, and marketing language frequently conflates "authorized" with "rigorously validated."
The gap matters most when a model leaves the site it was built on. A systematic review of radiology AI generalizability studies published between January 2022 and June 2025, in the Annals of Medicine and Surgery, found that external validation at a different institution produced a median AUC drop of only about 0.03 — but specificity losses of up to 24 percentage points in some cases (one example dropped from 94% to 70%). A headline accuracy number measured in a vendor's own validation set is a real number — it just may not be the number that shows up at a buyer's site.
Regulators are already drawing this line
This isn't only a radiology-journal debate. The FTC's "Keep your AI claims in check" guidance states plainly that performance claims are deceptive unless backed by competent, reliable evidence, and that any claim an AI product outperforms a prior or competing product requires actual comparative testing data — not an assumption. Under "Operation AI Comply," the agency has brought more than a dozen enforcement actions against unsubstantiated AI capability claims since 2024. The Neiman editorial and the FTC's guidance are really making the same point from different directions: a claim that isn't backed by evidence someone else can check isn't a finding, it's marketing.
What a defensible claim looks like
| Question to ask | Hype answer | Defensible answer |
|---|---|---|
| Where was accuracy measured? | Internal validation set only | External, multi-institution test set, disclosed |
| Who ran the study? | Vendor-funded, vendor-authored | Independent or peer-reviewed, vendor-disclosed as funder |
| What does "FDA cleared" mean here? | Implied equivalent to a clinical trial | Named pathway (510(k)/De Novo) and predicate stated |
| What outcome is claimed? | "Improves accuracy" (vague) | Named metric, dataset, and comparison baseline |
| Is it comparative? | "Outperforms radiologists/competitors" | Backed by head-to-head testing data, or not claimed |
Where xAID fits
None of this is an argument against AI in reporting — it's an argument for buyers reading claims the way the Neiman editorial wants policymakers to read workforce statistics: by checking what evidence actually backs the number. For imaging centers and teleradiology companies evaluating vendors, that means asking for the dataset behind a validation figure, not just the figure. It's also the reasoning behind how xAID's workflow is built: the AI produces a structured report draft, xAID's in-house radiologist reviews every preliminary, and the result reaches a client ready-to-sign — so the claim being made is about a reviewed draft, not an unsupervised verdict.
Frequently asked questions
What did the new radiology editorial say about misleading imaging claims?
In October 2026, leaders at the Harvey L. Neiman Health Policy Institute published a JACR editorial warning that an influential claim — that falling imaging volumes mean the U.S. no longer needs more radiologists — is misleading because it rests on data through 2016 and ignores a decade of evidence showing rising imaging volume, complexity and radiologist workload. They warned that an unfounded claim, once repeated, can quickly become accepted wisdom and shape workforce and policy decisions.
Does the same scrutiny apply to AI-reporting vendor marketing claims?
Yes, by the same logic. An independent 2025 review found that only 66% of 173 CE-certified radiology AI products had any peer-reviewed evidence behind them, and most of that evidence stops at diagnostic accuracy rather than clinical outcomes. The FTC has separately warned that AI performance and comparative-superiority claims are deceptive unless backed by competent, reliable evidence — the same standard the JACR editorial is asking policy claims to meet.
Does FDA clearance mean an AI tool's accuracy claims are validated?
Not necessarily. About 94.6% of FDA-authorized AI/ML medical devices, including most radiology AI, are cleared through the 510(k) pathway, which requires showing substantial equivalence to an existing device rather than independent clinical validation. Researchers writing in Radiology: Artificial Intelligence have noted that regulatory rigor and public transparency about that rigor are distinct and often conflated by marketing language.
What should imaging centers and teleradiology companies ask AI vendors before buying?
Ask for the regulatory pathway (510(k) versus De Novo), whether performance was validated on an external, multi-institution dataset rather than only the training site, whether the evidence is peer-reviewed and vendor-independent, and whether the claim is about diagnostic accuracy only or about real clinical outcomes. A published systematic review found specificity can drop by as much as 24 percentage points when AI models are tested outside their training institution — a gap marketing materials rarely disclose.
Sources: Holland & Knight, "Operation AI Comply" Two Years Later (2026); Radiology Business and AuntMinnie on the Neiman Health Policy Institute's JACR editorial; Cutler & Klarnet, JAMA Health Forum (2026); Antonissen et al., European Radiology (2025); Christodoulou et al., arXiv preprint (2025); Almarie et al., Biomedicines (2025); Diab & Lotter, Radiology: Artificial Intelligence (2025); Suleman et al., Annals of Medicine and Surgery (2025); FTC, "Keep your AI claims in check" (2023). Figures are rounded as reported.