I keep turning that over because of what sits on the other side of it. The same class of model the agency is pointing at our copy is the one increasingly drafting it. And we know how that model behaves when left alone with a claim: it writes the efficacy beautifully and treats the risk language as something to be got through. Trained on years of marketing, it produces benefit-heavy copy, and if it is optimising for anything it learns that the safety information is the part nobody reads — so it trims it, softens it, moves it down the page (improvado.io). Fair balance is precisely the thing the machine is worst at, because fair balance is a judgement about proportion, not a sentence it can complete.

So here is the shape of the year. The instrument that writes the overstatement and the instrument that catches it are converging on the same architecture. One side drafts the imbalance; the other reads for it. And in the middle is a medical writer who, on a tight Friday, is tempted to let the first machine do more and check the result less.

The line I’d put on the wall is the one the FDA has held for decades and restated without blinking: the company is responsible for the promotional material, regardless of how it was made. There is no AI lane in the regulation, no allowance for “the tool drafted it.” The model is not a co-author you can name in your defence. Accountability did not move an inch when the drafting got faster — it still ends with a human signature, and now that signature is being read back by something that doesn’t get tired at four o’clock.

I don’t take this as a reason to keep AI away from promotional copy. I take it as the clearest argument yet for where our value actually sits. If the regulator has automated the first pass of review, then the fair-balance judgement — the proportion, the placement, the risk carried in the same breath as the benefit — stops being the boring bit at the end of the job. It is the bit the machines on both sides are worst at, which makes it unmistakably ours.

The writers who get nervous about this are the ones who have quietly let the safety information become the thing they do last and fastest. The ones who’ll be fine are the ones who already treat it as the thing they do most carefully. The audit was always coming. It just arrived as software, and it reads every line.

— Ned

Follow the conversation: #IrreplaceablesHealth

Source links:

• FDA’s AI-powered crackdown on alleged deceptive drug promotions: https://natlawreview.com/article/fdas-ai-powered-crackdown-alleged-deceptive-drug-promotions

• Pharma ad compliance 2026 — fair balance and ISI: https://improvado.io/blog/pharma-ad-compliance-fda-ftc-fair-balance-and-isi-requirements

• AI-generated pharma content and FDA responsibility principle: https://blog.madebyxds.com/ai-generated-pharma-content-fda-compliance-2026

Your review pipeline is drifting with the conversation’s mood

If your team runs promotional copy, clinical summaries or MLR-prep outputs through an LLM in a single conversation thread, the scores are being skewed by whatever came before — negative items more than positive, by a factor of 1.6. The fix is unglamorous and effective: a fresh context for every item. Most off-the-shelf review workflows don’t do this, which means the more you batch, the less you can trust the output.

A third of medical chatbots breach their own safety thresholds

A systematic sweep of 6,233 medical GPTs found that a third or more violate operational safety thresholds, and over half of the action-enabled ones carry no privacy disclosures. These are exactly the numbers regulators and legal teams will reach for. If you are advising a client on deploying an AI-assisted medical information tool, you now have an audit framework — and a set of findings — to point at when the question of due diligence comes up.

The FDA is piloting AI in early-phase trials

The agency’s request for information on AI-enabled optimisation of early-phase trials is worth reading not for what it decides — it decides nothing yet — but for how it frames the questions. This is an early look at how the FDA will think about AI in the trial process, and by extension the communications around it. The thinking that shapes the guidance is happening now, in the open.

Ogilvy Health puts its AI strategy on the record

Ogilvy Health has gone public on its adoption pace, its nano-influencer plans and its own proprietary bot, ‘Ria’. Read it as a competitive benchmark, not a playbook: the tool is agency-owned and won’t transfer to an in-house team. What does transfer is the signal about where a large healthcare agency is willing to say, on the record, that it is placing its bets.

The AI upsell is coming at renewal

If your operation runs on Salesforce or a Veeva-adjacent stack, the AI add-on costs land at your next renewal whether you wanted them or not — and Gartner is already warning that enterprise agreements may not hold. For any communications operation whose budget predictability depends on its CRM, this is a line item to raise before it raises itself.

I’ve been hearing that qualifier a lot. A physician writing about the quiet shame of AI in medical writing put a name to it: people confessing their AI use in whispers, prefacing it with “just a little,” as though the size of the admission changes its nature. It maps almost exactly onto what I see in health communications. The adoption is happening everywhere. The talking about it is happening nowhere.

Which is strange, because the numbers say the opposite. A recent industry overview puts adoption somewhere between 67 and 76 per cent — most organisations either using or testing AI across their workflows. Read that and you’d picture a field that has settled the question and moved on. The survey crossed the threshold. The people did not.

The gap between those two facts is where most of us are actually living. Officially, AI is a productivity story with a tidy percentage attached. Unofficially, it’s a thing you do at your desk and don’t quite mention at the team meeting, because you’re not sure whether disclosing it makes you look efficient or replaceable. The web developers have started saying this out loud — that AI is generating more than half their output while their employers quietly reassess headcount. They’ve decided the silence costs more than the admission. We haven’t, yet.

I don’t think the shame is about the tool. It’s about the unspoken sum the qualifier is trying to manage: if a little AI is fine and a lot is suspect, then admitting how much you actually use is admitting something about how much of the work was ever the hard part. That’s the real conversation, and “just a little” is the sound of us avoiding it.

So here is the question I’d put to the floor. If two-thirds of us are already doing this, who exactly is the whisper for?

— Ned

A colleague’s IT team “rolled out AI to everyone” last week — meaning the assistant in the browser, the one that summarises your tabs and rewrites your emails. The medical writers were to use it like everyone else. Reasonable, free, already there, and a quiet category error. The general-purpose assistant is built to be passable at everything for everyone; our work is narrow, exact and unforgiving of the near-enough. Harvard Business Review made the wider point this week: healthcare’s specialised dialects sit beyond what a generic model handles dependably. The danger was never that the tool is bad — it’s that it’s good enough to be trusted by people who can’t see where it fails, and in our field the failures stay invisible until someone qualified looks. When the rollout email lands, the useful reply is a question: trained on what, grounded in which sources, checked by whom?

Source link → Harvard Business Review: https://hbr.org/2026/03/healthcare-uses-specialized-language-it-needs-specialized-ai-too

The unfashionable question, asked out loud

For two years the conference AI sessions ran to a script: a demo, some adoption statistics, a closing slide about the future arriving whether you like it or not. That’s changing. STAT reported on panels where scientists are finally asking the unfashionable question aloud — not “is it impressive?” but “does it actually work?” We sit downstream of these tools; when a client’s research team runs a literature scan through an AI platform, the output flows into our copy and our name goes on the claim at the end. So practitioner scepticism isn’t something to borrow sheepishly. It’s cover. Make the question respectable in your own rooms: validated against what, on whose data, and who notices when it’s wrong?

Source link → STAT: https://www.statnews.com/2026/05/21/are-ai-scientist-tools-actually-useful-ai-prognosis/

When the builder is more cautious than the brief

The brief wanted copy on how the company “used AI to design” its lead molecule — the implication, unmistakably, speed. The trouble is that the person best placed to know was more careful than the marketing. In an interview with STAT, the chief executive of an AI-driven biotech drew the line: AI accelerated the early design — the candidates worth making — while everything after, the assays, safety, trials and regulators that decide whether a drug exists, runs at the old speed. The clever bit is narrow; the slow bit is most of it. When the person who built the model won’t overclaim, the writer downstream certainly shouldn’t. My test now: would the scientist who did the work nod, or wince? Write for the nod.

Source link → STAT: https://www.statnews.com/2026/05/26/ai-biotech-bighat-biosciences-ceo-on-ai-drug-development-hype/

— Ned

Follow the conversation: #IrreplaceablesHealth

What the people doing the work are actually saying

There’s a pattern emerging in how AI is landing inside pharma and health communications right now. Not in the press releases — those are uniformly upbeat — but in the anonymous reviews employees leave when they think no one important is reading.

Four signals worth paying attention to this month.

Vaniam Group: the gap between the brand and the room

In January, a former Vaniam employee wrote: "'AI' is frequently used as a buzzword, there appears to be little understanding of how it meaningfully applies to the company's products, services, or client value."

What makes this interesting isn't the criticism. It's the management response, which references "Vaniam Intelligence" — apparently the company's branded AI product — and frames employee scepticism as a training problem: "we're committed to providing the training and context needed so teams understand how these tools create meaningful value, not buzzwords."

Management's answer to "we don't understand how this applies to our work" is more onboarding. The gap between how AI is being sold internally and how it's actually landing with the people doing the work — that's the signal here.

Novartis: the enthusiasm without the infrastructure

A Novartis employee with over a decade at the company gave it 3 stars in March. Listed among their cons: "Over-zealous interest of senior leadership in AI (structurally not capable of (nor built to) contextualize output)."

That phrase — structurally not capable of contextualising output — is the most precise description I've come across of a problem that's everywhere right now. Leadership is sold on AI. The people who would actually use it, verify it, and carry professional responsibility for it understand that their organisation doesn't have the processes, the expertise, or the culture to do that safely.

This is a global pharma company operating under serious regulatory obligations. The distance between those two realities is not a gap you can close with a lunch-and-learn.

Inizio: AI as performance monitor

An RN Educator on one of the company's patient call programmes wrote: "No recognition for excellent patient care — just criticism for not using the right words to trigger the AI system that grades calls."

This is AI not as a creative or productivity tool but as a surveillance mechanism. The system doesn't assess quality; it grades compliance with language patterns. Which means the incentive becomes gaming the AI rather than serving the patient. Inizio's health communications division and its patient services arm are different businesses, but they share a culture and a leadership. Worth knowing what's happening across the house.

Pfizer: even the data scientists aren't enthusiastic

A small one to round out the picture. A Senior Data Scientist at Pfizer listed "pushing the use of AI" as a con in a recent review — alongside the in-office mandate — while the company moves through its now-familiar restructuring cycle.

When even the people trained in the technology are flagging its rollout as a frustration, it usually means the how is the problem rather than the what.

From the Floor is a monthly signal read drawn from anonymous Glassdoor reviews across 18 pharma and health communications companies. The signals are directional, not statistical. Qualitative monitors, not data. Ned selects what seems worth noting.

— Ned

Sources

Vaniam Group reviews (Glassdoor)

Novartis AI reviews (Glassdoor)

Inizio AI reviews (Glassdoor)

Pfizer reviews (Glassdoor)