Net Impression Is the Standard Your Claim Checker Cannot See

On 9 September 2025 a presidential memorandum directed the Department of Health and Human Services to act on transparency and accuracy in direct-to-consumer prescription drug advertising, and FDA moved the same day. Across the rest of 2025 the agency issued more than 200 enforcement letters challenging prescription drug advertising and promotion. Seventy-four of those went to drug and biologic manufacturers — 10 warning letters and 64 untitled letters — and only five of the 74 were issued before 9 September.

The interesting part is not the volume. It is the grounds. Arnall Golden Gregory's read of the wave puts it plainly: "OPDP emphasized the net impression of promotional pieces, evaluating each piece based on the overall takeaway conveyed to the intended audience, rather than analyzing individual claims in isolation." That sentence is a specification, and it describes something no claim-level checker computes. A retrieval system that verifies each sentence against an approved claims library returns a pass on a piece whose every claim is substantiated and whose overall effect is misleading. The failure is not a tuning problem. The unit of analysis is wrong.

Which matters more now than it did in 2023, because the volume of promotional material a large brand generates has gone up while the review capacity has not. Generative tooling produces derivative assets fast and cheaply, each one individually defensible and collectively unaudited for the thing OPDP is actually reading.

What FDA actually issued in 2025, and the two numbers that do not check out

Two figures circulate in industry decks about the September 2025 wave: that 59 warning letters concerned brand website content, and that 96% of formal warnings targeted digital and social channels. Neither survives a check against the published tallies, and repeating them will cost you credibility with the one person in the room who has read the letters.

Here is what the record supports. King & Spalding's 2025 year in review counts 74 letters to prescription drug and biologic manufacturers — 10 warning letters, 64 untitled letters, two of them from CBER and the rest from CDER's Office of Prescription Drug Promotion. Of those 74, 42 targeted direct-to-consumer television advertisements. Sixteen challenged promotion directed at healthcare professionals. Nine concerned spokespeople, of which two were DTC television spots and seven were media appearances by celebrities, executives and clinicians. The remainder covered social media, virtual backgrounds, newsletters, sales aids, exhibit booth panels, print advertisements, sponsored links, online videos and webpages. A separate 32 warning letters went to medical device firms from the Office of Product Evaluation and Quality.

Covington's October 2025 update records that in Q3 2025 CDER issued more than 50 warning letters and more than 50 untitled letters over drug advertisements, and notes a CBER untitled letter to AstraZeneca over FluMist dated 9 September 2025. Hall Render describes nearly 100 enforcement letters in a single week. The larger "approximately 100 cease-and-desist letters" figure FDA itself used includes compounding and telehealth firms, not only manufacturers — which is why the headline number and the manufacturer number never reconcile.

So the channel story is the opposite of the one usually told. Television, the most heavily lawyered surface in the entire promotional estate, drew the majority of the manufacturer letters. Digital drew a minority. That does not weaken the argument for worrying about generative content; it sharpens it. If the most-reviewed channel is where the composite-impression failures are being found, the least-reviewed channels are not safer — they are unmeasured.

Where "net impression" actually lives in the rules

This is the part to get right, because the difference between a binding clause and a draft one is the whole of the competence question.

Read that table twice. The most-cited articulation of the net impression standard — FDA's draft guidance on presenting risk information, announced in the Federal Register on 27 May 2009 with comments due by 25 August 2009 — has sat in draft for seventeen years. Guidance in draft is not binding on anyone, including FDA. The enforceable hook is the regulation: an advertisement fails 202.1(e)(5)(ii) where effectiveness information is given greater scope, depth or detail than risk information. OPDP's letters then apply "net impression" as the analytical method for reaching that conclusion.

The one genuinely new binding constraint is the major statement final rule, which codified five standards at 21 CFR 202.1(e)(1)(ii)(A)–(E): consumer-friendly language; audio matching the rest of the advertisement in volume, articulation and pacing; dual modality in television, meaning concurrent audio and readable text; legible font, contrast and placement; and no audio or visual elements likely to interfere with comprehension. That last standard, (E), is a net-impression rule written directly into the CFR. It makes a scene change, a swelling soundtrack or an on-screen graphic a compliance object.

And the widely-anticipated end of broadcast advertising is not law. HHS and FDA announced on 9 September 2025 an intention to remove the "adequate provision" option that has allowed broadcast advertisements since 1997 to point to other sources for full risk information. That is now RIN 0910-AJ14 in the Unified Agenda, with a notice of proposed rulemaking expected in December 2026 and final action targeted for 2027. As of 30 August 2026 no proposed rule has published. Anyone telling you to plan a 2026 budget around a finalised ban is selling you something.

Why claim-by-claim checking is blind by construction

An automated promotional reviewer built the standard way does three things. It segments the asset into claims. It retrieves the nearest approved claim or reference from a library. It scores similarity and flags divergence. That architecture is genuinely good at a specific class of defect: the unsupported sentence, the off-label indication, the reference that does not say what the claim says it says, the paraphrase that has drifted from approved language. Those are real violations and catching them early is worth money.

What that architecture cannot produce is a judgement about the asset as a whole, because it never forms one. Consider what the 2025 and 2026 letters actually cite:

• Relative prominence. Benefit copy above the fold, risk copy below it. Every claim passes. The piece fails.
• Sequencing and pacing. Risk narration delivered over a scene change, competing with a visual the viewer is watching instead. Now expressly covered by 202.1(e)(1)(ii)(E).
• Visual implication. Imagery depicting activity or outcomes the label does not support, alongside text that says nothing untrue.
• Indication scope drift. OPDP's twenty-first untitled letter of 2026, to Sanofi on 15 July 2026, concerned healthcare professional emails referring to treating "RSV disease" where the approval is specifically for RSV lower respiratory tract disease. Four words, and no claim-similarity score will fire on it if the approved claim is in the library with the same shape.
• Composite risk minimisation. No omitted risk, no false statement, but the sum of the choices leaves risk information subordinate.

Every item on that list is a property of arrangement, not of content. A retrieval system compares text to text. Prominence, order, contrast, dwell time and relative weight are not text. They are layout, timing and design decisions that live in the rendered asset, and most review pipelines never look at the rendered asset at all — they look at a copy deck and a reference pack.

The honest version of the mechanism claim is this. Claim checking reduces the rate of one failure mode while leaving the dominant failure mode unmeasured, and it does so while increasing throughput. More assets, each individually cleaner, none assessed on the standard being enforced. If your first-pass approval rate improves and your OPDP exposure is unchanged, the metric moved and the risk did not.

For scale on what that throughput means: Indegene's Future of MLR report, which is vendor-published and should be read as directional rather than as a benchmark, puts the manual state at roughly a 21-day review cycle, about three review rounds, $2,500–$5,000 per asset and a 2–3% compliance error rate, with AI-augmented pilots reporting three to five days and 1.2 rounds. Take the cycle-time numbers with appropriate scepticism. Take the implication seriously: the same review committee is being asked to absorb several times the asset count.

The 2026 letters moved the ground under the claims library

There is a second problem, and it is worse than the first, because it breaks the premise that an approved library is a stable ground truth.

OPDP's 2026 output has repeatedly found fault with claims that were previously understood to be permissible. Gardner Law's survey of the 2026 untitled letters records letters to ImmunityBio over Anktiva in January 2026 for survival and cystectomy-avoidance impressions drawn from a single-arm trial; to BeOne Medicines over a Brukinsa television commercial for creating an impression that patients could safely undertake extended travel and vigorous activity without adequately conveying risk; to ARS Pharmaceuticals over Neffy for emphasising the absence of needles in a way that could minimise seriousness; to argenx over Vyvgart Hytrulo; and to Novo Nordisk in February 2026 over overstated weight-loss expectations for the Wegovy tablet.

The clearest instance is a March 2026 untitled letter analysed by Sidley, concerning an injectable incretin product. FDA objected to the statement that the product has the most FDA-approved uses in its class, on the basis that a factually true comparison of approved indications creates a misleading impression of superiority. A 2005 warning letter concerning Zyrtec had taken the opposite position, that FDA does not object to truthful, non-misleading statements about approved indications. The same sentence, twenty-one years apart, on opposite sides of the line.

An approved claims library encodes yesterday's interpretation. When the interpretation moves, every asset that passed a similarity check against that library inherits the stale reading — and the more of them you generated, the larger the recall. This is the specific reason to instrument which assets used which library version, and to keep that mapping queryable. Most content supply chains cannot answer "which live assets rest on claim 4471, revision 3" without a manual audit.

Also worth noting for anyone building a business case on volume: the frequently quoted Veeva Pulse observation that roughly 80% of approved pharma content is rarely or never used in the field is a vendor figure, not an independent one. If it is even directionally right, a substantial share of the generative throughput now being celebrated is producing regulated liability that nobody will ever see, which is the worst available trade.

The visual half is worse, and there is a number for it

Text at least has a retrieval story. Images do not.

A study published in Plastic and Reconstructive Surgery — Global Open in April 2026, "A I naccurate Anatomy", screened 4,734 manuscripts and identified 37 articles containing facial anatomical depictions. AI-generated content appeared in 10.8% of the relevant 2025 articles — four of 37 — a marked rise on 2024. Gross anatomical inaccuracies were present in every surveyed image, rendering them unsuitable for medical education. The error breakdown by structure: muscular structures wrong in 29 of 32 cases, ligaments in 18 of 18, bones in 10 of 14, blood vessels in six of six, nerves in two of two, glands in one of one. The authors describe structural distortions, anomalies with no known human correlate and, in places, features resembling nonhuman morphology. A complementary web survey of online anatomy course sites found nearly half using erroneous AI-generated imagery in promotional material.

Be precise about what that establishes. It is a small sample in one specialty — facial anatomy in aesthetic medicine — screened by human reviewers with domain expertise. It is not a general error rate for all generative image models on all medical subjects. What it does establish is that the images passed peer review, in journals, with editors and reviewers looking at them. If that filter did not catch a leg with two femurs, a commercial review committee working through a queue of mechanism-of-action visuals under cycle-time pressure will not catch a misplaced fascial plane either.

And a mechanism-of-action illustration is a promotional claim. It says the drug does this, here, by this route. An MLR pipeline that runs text through retrieval and passes images through on a reviewer's glance has automated the tractable half of the problem and left the intractable half exactly where it was, while raising the number of images per campaign.

Europe adds a marking duty the United States does not have

If assets cross into the EU, there is one further live obligation that is frequently misdated.

The AI Act's Article 50 transparency obligations applied from 2 August 2026. Article 50(2) requires providers of AI systems generating synthetic image, audio, video or text content to mark outputs in a machine-readable format detectable as artificially generated. The Digital Omnibus — political agreement in May 2026, European Parliament endorsement on 16 June 2026, Council adoption on 29 June 2026, signature on 8 July 2026, entry into force on 27 July 2026 — did not defer Article 50 generally. It deferred the Annex III high-risk obligations to 2 December 2027 and the Annex I obligations to 2 August 2028, and gave systems already placed on the market before 2 August 2026 until 2 December 2026 for the Article 50(2) marking duty. The prohibited-practice rules and the Article 4 AI literacy obligation have been live since February 2025.

Promotional content generation is not high-risk under Annex III. That is the point people get wrong in both directions: it is not caught by the high-risk regime and its deferrals, and it is caught by the transparency and literacy obligations that are already in force. Marketing operations staff prompting a model to produce regulated promotional material are covered by the Article 4 literacy duty today, not in 2027.

What this means in practice

On Monday, four things.

Instrument the composite, not the claim. Add a rendered-asset review gate that produces an explicit judgement on relative prominence of benefit and risk, sequencing, and whether visual or audio elements interfere with comprehension. Write it against 21 CFR 202.1(e)(5)(ii) and 202.1(e)(1)(ii)(A)–(E) so the reviewer's finding maps to a clause. This is a human judgement with a structured output, not a model score. A model can prepare the evidence — extract the risk-copy word count against benefit-copy word count, timestamp when risk narration overlaps a scene change, capture above-the-fold screenshots at three viewport sizes — and a named reviewer signs the conclusion.

Version the library and keep the mapping. Every generated asset records the claims-library revision it drew on. When an interpretation moves, as it did between the 2005 Zyrtec position and the March 2026 letter, you need a query, not an audit. Budget this as a data engineering task in the content system, not as an AI project.

Separate the two things your pre-check is doing. Report claim-substantiation catch rate and composite-impression review coverage as different numbers on different lines. If the first improves and the second stays at zero, the dashboard is telling you the risk fell when it rose. Insist that any tooling business case state which of the two it addresses; a mechanism that scores claims against a library addresses only the first, whatever the marketing says.

Treat generated imagery as unapproved until a qualified human has verified the anatomy. Not a reviewer glance — a named medical or scientific reviewer with the relevant anatomical competence, with the verification recorded. The peer-review filter failed at this; yours will too unless someone is specifically accountable for it. If that person does not exist in your organisation, the honest conclusion is that you should not be generating anatomical illustrations yet.

Who signs: the promotional review committee chair owns the composite judgement, because it is a regulatory conclusion and not a tooling output. What it costs: a rendered-asset gate adds review time per asset, which is exactly the saving the automation was bought to deliver, and that trade should be made explicitly rather than discovered in a letter. What goes wrong: teams add the gate as an optional step, throughput pressure removes it within two quarters, and the only evidence anyone ever sees is an untitled letter naming a piece that passed every automated check it was given.

The standard itself is old. What changed is the volume of material being assessed against it, and the fact that the systems producing that material optimise for fluency, which is precisely the property that makes a misleading net impression easy to construct and hard to detect one sentence at a time.