On September 9, 2025, the FDA did in a single day what it had not done in the previous two years combined. It issued roughly 100 letters in a coordinated crackdown on deceptive drug advertising, 41 untitled letters and 8 warning letters to traditional pharma manufacturers, plus 58 to online pharmacies, telehealth outfits, and compounders. For scale: the FDA's Office of Prescription Drug Promotion issued only five enforcement letters in all of 2023 and 2024 combined. Five letters in two years, then about a hundred before lunch.
Every compliance newsletter that week reached for the same word: AI. The crackdown landed in the same year the FDA published its first draft guidance on artificial intelligence, and the agency has openly named AI-generated health content as a concern. But reading the wave as an AI story sends your team chasing the wrong fix. Read the letters and the violation is almost always the same old one: the ad minimized or omitted risk. The FDA does not care whether a copywriter or a language model wrote the sentence. It cares whether the final communication is misleading.
The numbers, and what they are actually counting
Across 2023 and 2024, OPDP sent five enforcement letters total. In the September 2025 action, the FDA and HHS issued on the order of 100 letters in one coordinated sweep, and the agency's broader 2025 enforcement posture has run past 200 letters for the year. The 49 aimed at traditional manufacturers (41 untitled, 8 warning) are the ones that matter for brand teams, because those are the companies running MLR-governed promotional review.
The temptation is to read a 20x jump as a 20x change in what is illegal. It is not. The rules did not change. Risk-minimizing, misleading promotion was a violation in 2022 and it is a violation now. What changed is enforcement appetite and operating tempo: the FDA decided to stop sending the occasional letter and start sending them in waves. The substance of the violations is strikingly conventional. The single most common problem across the DTC letters is the omission or minimization of risk information, the same category that has topped OPDP citations for a decade. A brand that read the 2025 letters expecting a novel "AI violation" found instead a familiar list: benefit overstated, risk buried, claims unsupported by the cited evidence.
So the correct reading is not "the FDA is now policing AI." It is "the FDA is now policing the same things, far more often, in public, at scale." If you were comfortable that your promotional output was clean under the old, slow enforcement regime, the new regime is a stress test of whether that comfort was earned or just untested.
Where AI actually enters the picture
The FDA has explicitly placed AI-generated health content and chatbot interactions inside the scope of what it is watching. When a model drafts patient-facing copy, or when a branded chatbot answers a question about a therapy, the output is a communication, and a communication that is false or misleading is a violation regardless of its authorship. The FDA's published focus areas for 2025 name AI-generated content and chatbot-mediated interactions directly. The agency is not treating "a machine wrote it" as a defense, and it is not treating it as an aggravating factor either. It is treating the machine as irrelevant to the question of whether the final claim holds up.
This is the part teams keep getting backwards. The risk AI introduces is not that the FDA will single out AI-drafted content for special punishment. The risk is that AI makes it trivially easy to produce the exact violations the FDA is now hunting, faster than any review function can catch them. A model asked to make a benefit claim "more compelling" will, helpfully, overstate it. A model summarizing efficacy data will, helpfully, compress the safety section into a clause or drop it. The same fluency that makes generative tools useful is the fluency that minimizes risk language, and minimized risk language is the number-one citation in the September wave. AI did not create a new category of violation. It built a high-throughput factory for the oldest one.
The guidance map: a real white space, and the FDA said so
There is no FDA guidance specific to AI in promotion. None.
The one document that exists is about something else. In January 2025 the FDA published its first draft guidance, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. It introduces a risk-based credibility-assessment framework for AI models used in regulatory submissions, the kind that supports a dossier or a safety analysis. The comment period closed April 7, 2025. It says nothing about using AI to draft an HCP email, a patient brochure, or a banner ad. The entire domain of AI-assisted promotion sits outside its frame.
So promotional teams operate in genuine regulatory white space: a high-enforcement environment with no AI-promotion-specific safe harbor to hide behind. The FDA has not signaled that a promotion-specific rulebook is coming. Its posture is that the existing standard already governs: the communication must not be misleading, and that applies to AI output exactly as it applies to everything else. No safe harbor is on the way to bail anyone out.
Why "the machine drafted it" makes the problem harder, not safer
There is a quieter reason AI-assisted promotion is a compliance hazard, and it has nothing to do with the model's intentions. It has to do with the paper trail.
Ask the people who do this work and they will tell you they do not trust the tools yet. In a 2025 survey, 65 percent of US pharma marketing and promotional-review professionals said they do not trust AI for creating regulatory compliance submissions. Their reasons are precise and worth memorizing, because they are the exact failure modes the FDA penalizes. The top concern was hallucination, named by 40 percent: the model states something that is not true. Second, at 20 percent, was the lack of a traceable audit trail: no record of where a claim came from or who approved it. Third, at 12.5 percent, was the lack of transparency and explainability: nobody can reconstruct why the model wrote what it wrote.
Put those next to the FDA's enforcement standard and they line up one to one. A hallucinated benefit is an unsupported claim. A dropped contraindication is risk minimization. And the missing audit trail is the thing that turns a defensible mistake into an indefensible one, because when the FDA asks how a misleading claim cleared review, "the model generated it and we are not sure why" is not an answer that survives an inspection. The traceability gap is not a side issue. It is the difference between a correctable deviation and a warning letter.
A worked example: how Varigel earns a letter without anyone deciding to break a rule
Take a fictional brand, Varigel, approved for one narrow indication with a known contraindication in patients on a common comorbidity medication. The label is precise. The MLR file is immaculate. Here is how a clean brand ends up in the September pile anyway.
The team adopts a generative tool to scale content for an unbranded-to-branded patient campaign. A marketer prompts it to "write a warm, reassuring 150-word description of what Varigel does for patients." The model returns lovely copy. It describes the benefit in confident, human language, slightly broader than the indication, because "reassuring" pulled it that way. It mentions the most common side effect and, to keep the warmth, compresses the contraindication into a soft "talk to your doctor" line that a reader skims past. Nothing here was malicious. The marketer did not decide to overstate the benefit or bury the risk. The model did, helpfully, in service of the brief.
Now multiply. The same prompt pattern produces forty variants for forty audiences and channels. MLR, staffed for the old volume, reviews them the same careful way it always has, and most get caught. But "most" is the problem. A couple slip, because the review function is now the narrow end of a pipe that generation just widened. Those ship. One of them is the ad that a clinician, a competitor, or an FDA reviewer running a sweep reads as minimizing risk. The letter that arrives does not say "you used AI." It says the communication omitted material risk information. Varigel is now a data point in a hundred-letter day, and the root cause was not a bad actor. It was speed without an upstream check.
The lesson is not "do not use AI." The lesson is that the violation is manufactured upstream, at the moment of generation, and caught or missed upstream, before MLR drowns. That is the only place the leverage is.
The rational response: self-govern upstream, do not wait
If the standard is "do not be misleading," it already applies to AI output, and no promotion-specific safe harbor is coming, the strategy follows directly. Govern AI-assisted promotion at the source, with a documented, human-overseen pre-check, before anything reaches MLR and long before anything reaches a patient. Three moves.
Check the claim against the label at the moment of drafting, not at the end of review. The violations in the September wave are label-relative: benefit broader than the indication, risk lighter than the label requires, evidence that does not support the claim. A pre-check that compares each AI-assisted asset against the approved label, before it enters the MLR queue, catches the exact failure mode the FDA is citing, while it is still cheap to fix. This is what Juncture's Pre-check does: it scores every asset against the approved label and surfaces drift before a reviewer ever opens it.
Keep the audit trail the survey says is missing. The 20 percent who named traceability were naming the gap that converts a mistake into a warning letter. Every AI-assisted claim needs a record: what it asserts, what label section it maps to, who reviewed it, and a signed approval. A 21 CFR Part 11 trail with an e-signature sign-off is not bureaucratic theater. It is the artifact that lets you answer the FDA's "how did this clear" question with a sentence instead of a shrug.
Treat the pre-check as the throughput fix, not a tax. Upstream governance is also the only thing that lets you use the generation speed at all. Catching drift before MLR means MLR reviews cleaner assets, the queue moves, and the extra output the model produced stops dying in a folder. The compliance move and the velocity move are the same move. We unpack the mechanics in why the MLR bottleneck got worse, not better and what a pharma pre-check actually does.
Where this leaves you
The September 2025 crackdown is not a referendum on AI. It is a referendum on whether your promotional output can survive scrutiny at volume, delivered by an agency that has stopped sampling and started sweeping. The five-to-a-hundred jump is what enforcement looks like when the FDA decides the old tempo was the only thing protecting sloppy content, and removes it.
The brands that read this as "the FDA is coming for AI" will waste the next year waiting for a rulebook that is not being written, or banning a tool their competitors are using to move faster. The brands that read it correctly will see that the standard never changed, only the odds of getting caught, and that the one place to act is upstream: a documented, human-overseen pre-check that catches the misleading claim at generation, keeps the trail that survives an inspection, and clears the MLR queue as a side effect. Juncture is built for that seam, joining the pre-MLR Pre-check to the Answer Monitor that watches what AI engines say about your brand after launch, both measured against the same approved label.
Bring one brand and the AI-assisted assets you are about to send into review. We will show you which ones drift from the label, where the risk language thinned, and what a clean upstream pre-check would have caught, before any of it becomes a letter.
Sources
- U.S. FDA, "FDA Launches Crackdown on Deceptive Drug Advertising," September 2025. fda.gov
- Sheppard Mullin, "FDA's Wave of Untitled Letters Signals Stricter Scrutiny for DTC Pharma Ads," 2025. sheppard.com
- IntuitionLabs, "FDA AI Drug Advertising Enforcement and Compliance," 2025. intuitionlabs.ai
- U.S. FDA / Federal Register, "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products," January 2025. federalregister.gov
- Klick Health / Momentum Events survey on AI for regulatory compliance submissions, reported by FiercePharma, 2025. fiercepharma.com