A regulatory surge, with an unanswered question
Artificial intelligence is spreading through health research and regulation faster than its use can be independently checked. A narrative review of regulatory, health technology assessment, industry, journal-policy and survey records through September 2026 found 1,614 AI-enabled devices authorized by the US Food and Drug Administration. Radiology accounted for 1,230 of them, or 76.2 percent. The FDA had also received more than 500 drug and biologic submissions containing AI components since 2016.
The record shows a steep rise in device authorizations, from 18 in 2016 to 335 in 2025. But the review found no documented case in which a regulator or health technology assessment body accepted or rejected an AI-derived endpoint as such. An endpoint is the measured result used to support a regulatory, reimbursement or policy decision. Growth in the number of tools does not by itself show that their outputs are trusted for consequential evidence.
The FDA's January 2025 draft guidance proposed a seven-step, risk-based framework for judging the credibility of AI models. As of September 2026, it remained a draft and explicitly excluded AI used in drug discovery and operational efficiencies. The framework therefore did not settle the wider question of how AI-generated evidence should be evaluated across health economics and outcomes research.
Speed claims are easier to make than to verify
Companies reported striking time savings in work linked to regulatory submissions and payer dossiers. The figures ranged from 40 percent to 97 percent for regulatory writing and reached approximately 80 percent for payer-dossier drafting. Yet each figure came from a single-arm comparison without a concurrent control, and the source was the company deploying or selling the tool. The results therefore describe reported efficiency, not an independently established effect.
The review also found a large imbalance between the number of AI products and the number of published studies examining their economics. It identified 19 economic evaluations of AI in health care published from 2020 through 2024, compared with 1,614 authorized AI-enabled devices. That comparison does not measure quality or value directly, but it shows how little formal economic evidence sits alongside the expanding regulatory record.
The authors describe four failure modes as documented well enough to be expected rather than exceptional. One randomized vignette study of 457 clinicians found that standard AI raised diagnostic accuracy by 2.9 percentage points from a 73.0 percent baseline, while systematically biased AI lowered accuracy by 11.3 points. Explanations supplied with the AI offered no significant protection in that study. Other cited evaluations reported hallucination rates of 39.6 percent for GPT-3.5, 28.6 percent for GPT-4 and 91.4 percent for Bard, while a separate evaluation found fabricated citations in 55 percent of GPT-3.5 outputs and 18 percent of GPT-4 outputs.
The paperwork does not match the reported use
The clearest sign of that mismatch appeared in health technology assessment documents. AI use was disclosed in only 3 of 1,391 documents audited, or 0.2 percent, even though surveys reported use ranging from 25 percent to 81 percent. The audit measures declared use, not actual use, so it cannot show how often AI was used without being reported. It does show that readers of the published record have very little visibility into where AI entered the work.
The review treats disclosure as a methodological issue, not a matter of naming a software product. Reviewers would need to know whether AI selected evidence, defined cohorts, derived endpoints, generated code or changed the reimbursement decision. Without that information, it is difficult to reproduce the analysis or judge whether the human review was proportionate to the task.
A test for whether AI changes the answer
To make AI-assisted evidence more auditable, the authors propose TRACE-HEOR, a 10-item reporting standard extending CHEERS 2022 and PALISADE. The proposal is designed so that each item can be checked from the submission itself. It is not an established consensus instrument and has not undergone a Delphi process, a structured method for building expert agreement.
The proposal's most consequential test asks what happens when the AI component is removed. Researchers would re-estimate the primary result using a pre-AI alternative or a plausible worst case, then assess whether the incremental cost-effectiveness ratio crosses the relevant threshold or whether the recommendation changes. In plain terms, the test asks whether AI merely made the work faster or whether it could have changed the decision.
That focus reflects the review's central concern: speed and validity are different questions. The evidence assembled here does not show that AI improves the truth, validity or clinical utility of health-economics evidence. It also does not provide a single pooled estimate of AI performance, cost effectiveness or productivity.
A record that still needs stronger checks
The authors caution that their evidence base is a narrative rather than systematic review, so negative results may be less visible. Several important quantities came from conference abstracts or company communications, and the source records used different tasks, populations, comparators and outcome definitions. Those differences prevent one pooled estimate from summarizing AI's effects.
The study received no funding. Its author reports no conflicts of interest related to the work, while also reporting an academic affiliation and a health-economics consultancy affiliation; the text says no client had a role in the review or reviewed the manuscript.
The article says its large-language-model classifier was itself an object of study rather than an assistive writing tool. Its prompts, inputs and outputs were archived in the public replication package, and no AI tool was listed as an author. The proposed standard, however, remains a proposal until it is tested prospectively and validated by broader agreement.
Paper data and sources
Original title: Faster but Not Truer, Part 2: The Regulatory Record, the Economics, and the TRACE-HEOR Reporting Standard
Authors: Not provided
Journal/Repository: Journal of health economics and outcomes research
Status: Peer-reviewed
First online: 2026-10-19
DOI: Not available
Original paper · Full text