Warning letters from the U.S. Food and Drug Administration almost never mention artificial intelligence. A retrospective document analysis of every warning letter the agency posted between January 1, 2022, and September 28, 2026, found that just 7 of 3,011 letters, or 0.23 percent, referred to AI in any form. The finding cuts against a wave of news coverage suggesting that regulators have begun issuing AI-specific enforcement actions; instead, the study shows that when AI does surface in FDA correspondence, it appears alongside long-established violations rather than as a new category of charge.
The analysis, published in Discover Pharmaceutical Sciences, searched the full text of all letters in FDA’s public warning letter database for the terms “artificial intelligence,” “AI,” “machine learning,” “large language model,” “neural network,” “deep learning,” and “chatbot.” Every letter containing a search term was then read in full. Eleven letters contained a term, but four were excluded on review: in two, “AI” meant adequate intake, a nutrition term, and in two others the abbreviation appeared only in a document title or an action-item number. That left seven letters for the final count.
The yearly breakdown underscores how rare these references remain. In 2022, none of 689 letters mentioned AI. One of 608 letters did so in 2023, two of 569 in 2024, three of 736 in 2025, and one of 409 in 2026 through the September 28 search date. The authors note that with only seven letters, no trend can be established, and the analysis is descriptive rather than explanatory.
The seven letters split across three FDA programs. Three came from the device program, three from the Center for Tobacco Products, and one from the Center for Drug Evaluation and Research. Together they cited 36 violations, ranging from 2 to 9 per letter, and AI appeared in the supporting text of six of them. In no letter was AI the only violation cited.
In the device letters, AI described what a product claimed to do or what its unvalidated software could do. In letters to Exer Labs and SeniorLife Technologies, claims of “artificial intelligence-based algorithms” to screen, diagnose, or “pre-diagnose” health conditions gave each product an intended use that required premarket review. FDA charged the products as adulterated under section 501(f)(1)(B) of the Federal Food, Drug, and Cosmetic Act because no premarket approval application or investigational device exemption was in effect, and as misbranded under section 502(o) because no 510(k) premarket notification had been submitted. In the letter to WAVi, the AI reference was the firm’s own statement, recorded under a software validation violation, that its unvalidated software had “Artificial Intelligence (AI) codes & capabilities.”
The three tobacco letters took a different tack. FDA charged electronic nicotine delivery system products as adulterated and misbranded for lacking marketing authorization, and listed an “AI chat” feature among smartphone-like characteristics that may appeal to young people. AI was not part of any cited violation in those letters; it figured only in the agency’s description of product features.
The letter that stands apart is the April 2026 warning letter to Purolea Cosmetics Lab, the only one in the period to treat a firm’s use of AI as a violation. The letter cited insanitary conditions, three current good manufacturing practice violations, and unapproved new drugs. A separate subsection stated that failing to review AI-generated specifications, procedures, and master records “is a violation of 21 CFR 211.22(c)” and that any AI output used in current good manufacturing practice activities “must be reviewed and cleared by an authorized human representative” of the quality unit. The authors identify it as the earliest CDER warning letter they found addressing a firm’s use of AI, and the only one in the period reviewed.
The Purolea letter does not create a new enforcement category. FDA cited no AI-specific requirement, and the letter’s other findings, including insanitary conditions, missing microbiological testing, and unapproved new drugs, describe a firm with broad manufacturing failures in which unreviewed AI output was one problem among several. The practical lesson for quality units, as the authors put it, is specific: if a firm uses AI to draft manufacturing documents, a qualified member of the quality unit must review and approve the output before it is used, and the firm should be able to show inspectors how that review is done. That is the same approval requirement that applies to documents written by people. FDA’s draft guidance on AI for regulatory decision-making and joint FDA-European Medicines Agency principles take a similar risk-based view, asking whether an AI output is credible for its context of use.
Inspection records reveal a second, quieter route by which AI enters FDA’s written record. Form FDA 483s and establishment inspection reports are released selectively on request under the Freedom of Information Act and are not full-text searchable on the agency’s website, so the authors searched FOIA-released records for the same terms, restricting “AI” to phrases such as “AI software.” Three records referred to AI. A June 2022 establishment inspection report for AltaThera Pharmaceuticals, from a postmarketing adverse drug experience inspection, recorded that the firm did not use automated algorithms or AI in drug safety operations; similar entries in reports from 2020 and 2021 suggest a standard inspection question. A January 2024 Form 483 for Imagen Technologies described the firm as an “AI software” device manufacturer and cited complaint-handling procedures. A November 2024 Form 483 for Alexion Pharmaceuticals noted that free-text adverse event data were “first sent for review by artificial intelligence” in August 2024. No inspection record cited a firm’s use of AI as a deficiency, and because the records are released selectively, this set is neither complete nor representative.
The study has limits the authors spell out plainly. FDA’s database lists letters after they are posted, so letters issued before the search date but not yet posted were not captured. Keyword searches miss references that use other wording, such as “automated algorithm,” without any of the search terms. Screening and coding were performed by one author with the assistance of a large language model, without a second coder, though the verbatim text supporting each decision is provided in a supplementary file. Warning letters also reflect FDA’s choices about which cases to pursue and publish, so the results describe the agency’s written enforcement, not the extent of AI use in industry. The review did not examine FDA’s treatment of AI in product approvals or in software as a medical device.
What remains is a grounded picture of how AI has entered regulatory correspondence. In the device letters, AI describes what a product does or is claimed to do, and FDA applies long-standing premarket and quality system requirements. In the tobacco letters, “AI chat” is one of several features the agency considered attractive to youth. Only the Purolea letter treats a firm’s use of AI as the violation, and it rests on an existing requirement, the quality unit’s responsibility to approve procedures and specifications. An enforcement precedent exists, the authors conclude, but it is built on existing rules, not new ones, and AI-related citations remain a small and heterogeneous subset of what FDA is actually citing.
Subject of Research: FDA warning letters and inspection records that reference artificial intelligence between 2022 and 2026
Article Title: Artificial intelligence in FDA regulatory correspondence, 2022 to 2026: a retrospective document analysis
Article References: Kwiecinski, G., & Nandy, R. (2026). Artificial intelligence in FDA regulatory correspondence, 2022 to 2026: a retrospective document analysis. Discover Pharmaceutical Sciences, 2(1), Article 22. https://doi.org/10.1007/s44395-026-00061-5
Image Credits: AI Generated
DOI: 10.1007/s44395-026-00061-5
Keywords: FDA, warning letters, artificial intelligence, generative AI, regulatory enforcement, CGMP, medical devices, tobacco products, quality unit review, pharmacovigilance, document analysis, drug regulation
News Source: Blake Davidson. (November 4, 2026). FDA warning letters rarely mention AI, retrospective analysis finds. Scienmag.
