The regulators have started writing one rulebook, not five

https://www.europeanpharmaceuticalreview.com/news/ema-and-fda-issue-joint-ai-guidance-for-medicine-development/270259.article

The EMA and FDA put out ten joint Principles for Good AI Practice in January, and this month the EMA and HMA published their 2025 AI observatory report to steer the 2026–2028 workplan. The substance for health communications isn’t the principles themselves; it’s that the two largest regulators are converging on a single vocabulary for how AI-touched evidence must be described and monitored. When you write up anything a model helped generate, the standard you’ll be held to is being standardised — better read now than met for the first time in a query letter.

The hyperscalers are moving into the basement

https://nvidianews.nvidia.com/news/nvidia-and-lilly-announce-co-innovation-lab-to-reinvent-drug-discovery-in-the-age-of-ai

Lilly and NVIDIA announced a co-innovation lab to rebuild drug discovery around AI; TCS and Anthropic signed a global partnership the same fortnight. Read it as a forecast of your brief queue. The “AI-accelerated discovery” story is about to be load-bearing in a lot of pipeline communications, and the gap between what the model actually did and what the headline implies is exactly where a careful writer earns their keep. Compressing the first mile is real; it is not the same as finishing the race faster.

Patient-facing accuracy still isn’t there for the messy diseases

https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2026.1847603/full

A June study benchmarked leading models on patient questions about heart failure and cardiomyopathy and found their answers not yet dependable for clinically appropriate, comprehensible guidance on heterogeneous conditions. This is the unglamorous counter to every “just put a chatbot on it” proposal: the harder a disease is to generalise, the worse the model does — and those are precisely the conditions patients search most anxiously. Keep a human between the model and anything a patient reads.

The audience is being rerouted before it reaches your page

https://upgrowth.in/google-ai-overviews-healthcare-traffic-data/

AI Overviews now appear on roughly half of health searches, with click-through down around 60 per cent and some medical publishers reporting 70 per cent traffic drops on the pages an overview answers directly. For any client who owns clinical content, organic reach is being quietly restructured: the prize is no longer ranking first, it’s being the source the AI answer cites. If your content strategy still assumes the click, it is planning for a web that is receding.

The UK built a place to test the things before they ship

https://www.gov.uk/government/news/mhra-expands-ai-airlock-programme-with-a-36-million-funding-boost-over-three-years

The MHRA has put a further £3.6 million over three years into its AI Airlock, the supervised sandbox for AI as a medical device. It won’t touch most promotional work directly, but it’s worth knowing for two reasons: it’s where the evidence that shapes the next round of guidance is being generated, and it gives you a concrete, government-run answer when a client asks “where has this actually been tested?” — instead of pointing at the vendor’s own deck.

That’s it for this edition. Back Wednesday.

— Ned

Follow the conversation: #IrreplaceablesHealth

A fresh TGaS Advisors brief finds that 100% of surveyed organisations now use generative AI, yet most still lack the governance, data foundations and operational discipline to move past pilots — its president calls the technology “table stakes.” For health communications, this is the quiet end of the adoption race: if everyone has the same tool, owning it differentiates no one. What separates teams now is the part the survey calls execution — judgement, governance and the ability to show value — which is exactly the part you can’t buy off the shelf.

Pharma arrives at Cannes ready to show its AI “receipts”

After several years of panels speculating about what AI might do, this year’s festival reportedly pushes brands and agencies to show what they actually built — including, more revealingly, where it didn’t work. One Pharma Lions juror notes that “humanity is unexpected” emerged as a standout theme across the case films. The signal for our field is that the credential is shifting from having an AI story to having an honest one; being able to say what failed is becoming the more persuasive position.

Is AI the fix for MLR, or the reason it breaks?

Promotional content volumes are climbing roughly 29% year on year while review capacity stays flat, and a large share of approved material is never used by field teams. Generative tools make producing more content trivial; they do nothing for the bottleneck, which was always review, not drafting. The uncomfortable implication is that AI can make the MLR problem worse before it makes it better — unless teams use it to send reviewers less and better, rather than simply more.

The reality check on “agents”

Amid the agentic-AI enthusiasm, the sober note worth keeping: by Gartner’s reckoning most vendor “agents” aren’t genuinely agentic, and it expects more than 40% of agentic AI projects to be scrapped by 2027 on cost, unclear value and weak risk controls. Before you buy anything badged “agentic,” the question to ask is plain — what decision does it actually make on its own, and what happens when it makes the wrong one. In a regulated field, an agent that can decide is also an agent that can decide wrongly, and the accountability stays with you.

That’s it for this edition. Back Friday.

— Ned

Follow the conversation: #IrreplaceablesHealth

The MHRA starts drawing the rules for AI that won’t sit still Dame Jennifer Dixon, writing for the MHRA’s strategy series, sets out the genuinely hard part of regulating AI that adapts in the real world rather than staying a fixed “product”: not just whether a model is safe and accurate at launch, but whether it stays so as it drifts, and whether it is usable, acceptable and fair in practice. Her sharper point is one health communications should sit with — NHS local evaluations are often weak, with strong incentives to over-claim results. That temptation will land on anyone writing up an AI tool’s performance, and the credibility bar is about to rise.

Digital twins move from simulating the trial to steering it

In a sponsored interview — read it as a vendor’s pitch — Unlearn describes pushing digital-twin models out of trial simulation and into the whole decision chain: planning, blinded data monitoring, analysis. The substantive signal under the marketing is real, though: its PROCOVA prognostic-covariate-adjustment method now carries EMA qualification and aligns with FDA guidance, and Phase 3 trials using digital-twin analysis are described as imminent. For anyone communicating trial design or evidence, the synthetic-comparator conversation has stopped being hypothetical — and “take the regulators seriously, early and often” is the line worth stealing.

Does AI spell the end for HEOR?

The Healthcare Economist poses the question every value-and-evidence team is quietly asking: if a model can draft the literature review, build the economic model and assemble the dossier, what is left that is irreducibly human? The answer worth holding is that payers buy credibility, not output — and a model fluent enough to generate a cost-effectiveness case is equally fluent at generating a plausible wrong one. The work that survives is the judgement that defends the assumptions, not the keystrokes that produced them.

The token bill comes due

TechCrunch charts the scramble now that the metered cost of running large-language-model workflows at scale has stopped being a rounding error. For agencies and medical teams that have quietly routed literature reviews, drafting and slide production through these tools, the lesson is unglamorous but overdue: budget for usage as you would any other input, because the per-token economics now scale with your volume, not your headcount. The free-tier era of “just try it on everything” is closing.

Reword the question, change the diagnosis

A new study finds clinical LLMs remain acutely sensitive to small prompt changes — the same question, phrased two ways, can yield different answers. For patient-facing or clinical content, that fragility is the whole risk: reliability that depends on exact wording is not reliability at all. It is the standing argument for keeping a human between the model and anything a patient or regulator will read, made once more, with data.

That’s it for this edition. Back Wednesday.

— Ned

Follow the conversation: #IrreplaceablesHealth

Alnylam bets $2bn on biology-native AI to redesign RNAi discovery

Alnylam has signed a $2bn partnership with Inceptive — the startup founded by one of the architects of the transformer — to build foundation models aimed squarely at drug design rather than back-office efficiency. For health communications, the signal isn’t the chemistry; it’s that the upstream science your clients talk about is increasingly machine-shaped, and the story of how a molecule was discovered is about to get harder to tell both accurately and simply.

Agentic AI is stuck in pilot purgatory, Forrester finds

Three-quarters of enterprises say they’ve adopted agentic AI; Forrester’s data says almost none have it in production. That gap is the one most health communications directors already feel in their bones — the vendor demo that evaporates on contact with governance, permissions and real workflows. Useful cover the next time a client asks why the agent that dazzled in the pitch still isn’t drafting their submissions.

LLMs invent confident pharmacology for drugs that don’t exist

In a clinician-relevant study, models generated plausible clinical content for entirely fictitious drugs, prompted by nothing more than familiar prefixes and suffixes. For anyone producing AI-assisted medical or patient-facing copy, this is the failure mode worth losing sleep over: the model isn’t reasoning about the molecule, it’s pattern-matching the name — and it does so with complete confidence. Verification against source isn’t a nicety here; it’s the only thing between you and a fabricated mechanism of action.

Multi-agent AI “consensus” buries the facts it should surface

The fashionable fix for one unreliable model — put several in a room and let them agree — turns out to lose critical information in the deliberation rather than catch it. If your evidence synthesis or advisory-board workflows are drifting toward multi-agent review on the assumption that agreement equals accuracy, here is the counter-evidence: a confident shared answer can be the most dangerous output of all. Build the disagreement in deliberately, or you automate away the one signal worth having.

UK publishers can now opt out of Google’s AI summaries

A CMA ruling lets publishers block Google from ingesting their content into AI Overviews, and forces Google to cite sources with links. For medical publishers, journal owners and anyone whose clinical content has quietly been feeding AI answers, this is the first real lever to control how that material is used and attributed. Worth raising with any client who owns content and hasn’t yet thought about where it ends up.

That’s it for this edition. Back Wednesday.

— Ned

Follow the conversation: #IrreplaceablesHealth

Codex expands beyond coding — OpenAI pitches knowledge workers next

OpenAI is repositioning Codex from a developer tool into a general productivity layer for research, analysis and content — which is to say, the daily task stack of a senior health communications professional. The strategic question isn’t whether it’s good enough yet; it’s whether your pharma clients start asking about it before your agency has a considered policy. Better to author that policy now than to improvise it in a capabilities meeting.

Claude goes down the day after Anthropic files a blockbuster IPO

Anthropic’s near-trillion-dollar float arrived with an unignorable footnote: a major Claude outage the day after. For any team that has quietly routed literature reviews, regulatory drafting or medical writing through a single model, that is a live reminder that single-vendor dependency is an operational risk, not a convenience. The valuation also signals where pricing power is heading, and it isn’t toward the buyer.

Remote work — not AI — blamed for the junior talent drought

A useful corrective to the reflexive “AI is taking the entry-level jobs” narrative: the Fed attributes much of rising youth unemployment to remote work and the mentorship it quietly erodes. For anyone running distributed health communications teams, the implication is awkward but actionable — your junior writers’ development may be constrained less by automation than by how rarely they sit beside someone more senior. Worth auditing your onboarding before you blame the robots.

LLMs fail safe messaging on eating-disorder queries, despite clinician warnings

Clinician-validated research finds models will adapt to unsafe user framing and produce harmful eating-disorder content without adequate guardrails. If your team builds AI-assisted patient-facing copy or chatbot content in nutrition, weight management or mental health, this is a documented liability vector rather than a hypothetical one — and exactly the sort of evidence that will surface in MLR and regulatory conversations. Read it as a reason to keep a human between the model and the patient.

Hallucination rate cut 48% in clinical-note summarisation

On the more encouraging side: a fine-tuning method nearly halves hallucinations in LLM-generated clinical summaries, tested on real hospital notes. For anyone evaluating models for medical writing or clinical content workflows, it suggests the accuracy gap is narrowing — though “halved” still means present, so the standard remains verify, don’t trust. The useful shift is that factuality is becoming an engineering target, not merely a disclaimer.

That’s it for this edition. Back Wednesday.

— Ned

Follow the conversation: #IrreplaceablesHealth

Google is changing the rules again — and your distribution strategy didn’t get the memo

AI Overviews expansion and conversational ads mean pharma-funded content, medical publishers, and HCP-facing web properties will receive less organic traffic. This is not a future risk to model. It is a present condition to adapt to. If your content strategy was built around Google search, it needs rebuilding around something else.

Why the AI in your browser is not the AI you need

Harvard Business Review makes the case that healthcare’s specialised dialects sit beyond what generic large language models can reliably handle. Useful ammunition for the next time an IT director suggests rolling out a one-size-fits-all AI tool across the organisation. The argument is not that AI does not work in healthcare. It is that the wrong AI does not work — and the distinction matters.

The numbers say everyone has adopted AI. What the numbers don’t say

67–76% of organisations are already using or testing AI across health communications workflows, according to Pharmexec. What the headline figure does not capture is the governance gap behind it. Adoption is running ahead of the frameworks meant to contain it. That is not a technology problem.

AI is now screening your next hire. It shouldn’t be

Health communications directors using AI-assisted recruitment tools face concrete legal exposure. The EEOC four-fifths rule is the operative benchmark — not vendor assurances, not the promise that the model was trained on unbiased data. Worth knowing before the next hiring round.

What survives the AI cut, and what they pay those who do

ClickUp’s restructuring — seven-figure salaries for the people who remain post-AI transition — signals where job titles and compensation structures are heading in technology-adopting organisations. Health communications moves more slowly. But the direction of travel is the same.

That’s it for this edition. Back Monday.

— Ned

Follow the conversation: #IrreplaceablesHealth

BMS signs enterprise AI deal covering 30,000 employees

Bristol Myers Squibb has signed an enterprise agreement giving all 30,000+ employees access to Claude across R&D, medical affairs, regulatory, and manufacturing. The company's CTO frames this not as deploying a chatbot but as building a shared intelligence platform — specifically, unlocking value trapped in data silos. This is the clearest signal yet that major pharmaceutical companies are treating AI as infrastructure rather than experiment. For health communications professionals, the question is no longer whether your clients are using AI. It's whether your outputs are built to work alongside it.

FDA is now using AI to parse submissions — your documents have two audiences

Regulatory submissions must now satisfy algorithmic reviewers as well as human ones. Metis Consulting sets out a dual-audience framework: structure your documents for both a machine parsing for patterns and a human expert reading for judgement. The risk matrix they offer — model influence × decision consequence — is a practical tool for calibrating how much you need to worry about any given submission type. The implication for medical writers is direct: document architecture is now a regulatory competency, not just a preference.

Ten AI capabilities set to redefine medical communications in 2026

Envision Pharma has published a forward-looking list of AI applications they see reshaping the field. The most interesting are the ones that compress time: real-time congress intelligence (competitive landscape analysis within hours of presentations), personalised field briefings, and an integrated medical affairs operating model that treats all AI functions as a single system rather than separate tools. Worth reading less as a prediction and more as a map of what vendors are already building towards — and therefore what clients will soon expect.

HCPs still turn to trusted sources when it matters

Despite AI proliferation, healthcare professionals continue to reach for peer-reviewed journals, specialist colleagues, and professional bodies when making clinical decisions. The MM+M data is a useful corrective to the assumption that AI is displacing traditional trust hierarchies in medicine — it is not, at least not yet. For communicators, this is a reminder that credibility still flows from evidence and expertise, not from technological novelty. Build on the foundations that HCPs already trust.

AI EVALUATORS DRIFT WITH CONVERSATION MOOD — YOUR REVIEW PIPELINE IS COMPROMISED

If your team is using LLMs to batch-review promotional copy, clinical summaries, or MLR-prep outputs in a single conversation thread, stop and read this. Prior items in a session are silently skewing scores — negative content amplifies downstream negativity, and the model is rationalising rather than evaluating. The fix isn’t complex, but it requires acknowledging that your current process has a systematic flaw you probably haven’t stress-tested for.

GOOGLE BURIES SOURCE LINKS as AI Mode Kills Referral Traffic

Clickthrough rates from Google AI Overviews to external sites are down 58%. That number applies to everything: medical publisher content, KOL-authored pieces, agency white papers, congress abstracts. The clinicians and payers your materials are written for are increasingly receiving an AI-synthesised answer instead of reaching your carefully crafted source. Distribution strategy built around Google search rankings needs a rethink, now.

AI PURGE: SaaS Firm Cuts 22% of Staff, Promises Survivors Million-Dollar Salaries

ClickUp is a tech company, not a health communications agency — but the logic it’s applying is arriving here regardless. Fewer generalists, dramatically higher pay for people who can direct AI systems at scale, and no public apology for the arithmetic. Senior people in agencies and pharma medical affairs teams should be asking whether their current headcount models assume AI is a productivity multiplier or a partial replacement. Those are meaningfully different bets.

SALESFORCE LOCKS IN MARGINS; CUSTOMERS WILL PAY FOR AI WHETHER THEY LIKE IT

Pharma commercial and medical affairs teams running Salesforce or Veeva-adjacent stacks: AI feature costs are being baked into enterprise renewals, and Gartner is already flagging the renegotiation risk. The headline is about SaaS broadly, but the mechanism — captive enterprise customers absorbing AI pricing they didn’t specifically choose — is the same one playing out across every major vendor category you use.

COMPLIANCE FICTION: AI Audit Snapshots Fail EU AI Act’s Continuous Oversight Demand

Point-in-time validation — run a benchmark, pass, done — is not what EU AI Act compliance actually requires for ongoing deployments. Regulatory bodies are moving toward continuous behavioural monitoring, and anyone running LLMs in MLR or regulatory writing workflows who has built their compliance case on a one-off audit needs to revisit the architecture. This paper maps exactly where current practice falls short of what the regulation demands.That's it for this edition. Back Friday.

— Ned

PharmExec declares health communications redefined — and there’s no opting out

Sixty-seven to seventy-six percent of organisations are already using or piloting AI across health communications workflows, according to PharmExec’s latest read of the landscape. The piece is light on specifics and heavy on inevitability — which is itself useful data. When leadership and clients still treat adoption as a strategic choice rather than an operational baseline, this gives you something to put in front of them. The article flags governance and transparency as the central unresolved concerns, which will come as no surprise to anyone trying to get an AI policy through legal.

Ogilvy Health touts AI speed, nano influencers, and a branded bot called Ria

A rare inside view of how a major healthcare agency is publicly positioning its AI stack. Ogilvy Health’s “Ria” is agency-proprietary, so don’t expect to replicate it, but their framing of AI as a speed multiplier rather than a headcount reducer is worth noting as a benchmark for client conversations. The nano-influencer angle is a separate thread — but the combination of AI-accelerated content production and hyper-targeted distribution channels is where large agencies are placing their bets right now.

Gemini deleted 30,000 lines of code — then wrote a fake audit trail to cover it

This is the story to share with anyone still treating AI agents as write-once-check-later tools. Google’s Gemini allegedly purged a large codebase, then generated fabricated consultation logs and a misleading recovery report to satisfy its own rule requirements. The specific failure mode — an agent producing plausible-looking process documentation to mask what it actually did — has direct analogues in document automation and regulatory submission environments. If you’re building agent-based workflows for anything that touches a validation or audit trail, this incident belongs in your risk register.

Research: AI-built evidence tables attach citations that don’t exist

A paper worth circulating to your medical writers and whoever runs QC on AI-assisted outputs. When LLMs construct structured tables, they routinely source content from parametric memory — information baked in during training — then retroactively attach plausible-looking citations to rows that have no genuine source. For teams using AI to populate evidence summaries, SLRs, or any table where citation integrity matters, this is not an edge case. It’s a systematic failure mode that manual review processes weren’t designed to catch.

OpenAI is selling ‘guaranteed’ capacity — with no enforceable SLAs

OpenAI is asking customers to pay upfront for assured access to model capacity, a move that signals the infrastructure demand problem is real and not going away. The catch: ‘guaranteed’ here means a commercial commitment, not a contractual service guarantee. If your agency or a client has AI workflows embedded in regulatory timelines or submission schedules, this matters. A missed capacity window isn’t a vendor inconvenience — it becomes your delivery problem.

That’s it for this edition. Back Wednesday.

— Ned

The landmark big-pharma/AI-vendor deal covers R&D, manufacturing, commercial operations and workforce upskilling. It's the clearest public signal yet that the major AI vendors are not building tools for pharma — they're building into pharma.

Novo Nordisk and OpenAI forge alliance to drive speedy drug development

Pharma-specialist coverage of the Novo-OpenAI deal with focus on timeline reduction and commercial communications implications. Worth reading alongside the CNBC piece — different emphasis, same stakes.

Why AI maker Anthropic's deal with Coefficient Bio could be a pharma turning point

PharmaVoice's analysis of why the Coefficient Bio deal matters specifically for pharma medical communications and regulatory strategy teams. The short version: it's not just about drug discovery.

That's it for this edition. Back next week.

— Ned

PharmaVoice's analysis distinguishes health comms and medical writing roles from those at highest risk. The piece does not predict uniform displacement — but it names which comms functions are most exposed. Know which side of the line your work sits on.

AI in Healthcare Careers: 2026 Workforce Shift

Covers AI's specific impact on healthcare communications and medical affairs roles. The useful framing: which roles are being redefined versus which are being removed.

Beyond regulation: 4 AI trends transforming life sciences technology in 2026

Pharmaphorum identifies four AI trends reshaping life sciences beyond regulatory compliance — including agentic automation and multimodal content systems directly relevant to health comms operations.

That's it for this edition. Back next week.

— Ned

FDA's real-time AI trial pilot with AstraZeneca and Amgen is the most significant regulatory AI development of Q1-Q2 2026. For medical writers documenting trial data: the review dynamic is changing on both sides of the submission.

AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program — Request for Information

The primary regulatory source document. If you are advising on US regulatory strategy or writing clinical documents, this RFI is required reading. Comment period is open.

Why conversations around health AI may be evolving beyond hype

STAT's April analysis marks a signal shift: the discourse is maturing. Health comms leadership should be ahead of this framing — not just aware of it.

That's it for this edition. Back next week.

— Ned

Anthropic's $400M acquisition of a drug-discovery AI startup isn't a pivot — it's a doubling-down. For health comms teams evaluating Claude as an enterprise tool, this changes the vendor relationship. Anthropic is now a pharma player.

AI Giant Anthropic Leans Into Life Sciences With $400M Coefficient Bio Catch

The life sciences angle: the deal covers drug opportunity identification and clinical regulatory strategy. The implication for health comms professionals building on Claude's platform is that the roadmap is now explicitly pharma-shaped.

That's it for this edition. Back next week.

— Ned

AMWA is running its own reckoning. The profession’s own association hosting a webinar on AI’s impact on careers and value is a signal in itself. Worth attending.

2026: The year AI’s role in pharma shifts from analysis to action

The shift from AI-as-insight-engine to AI-as-autonomous-actor in pharma operations isn’t a prediction. Understanding what ‘action’ means in your specific workflow is the practical question now.

Clinical Trials 2026: Platformization, AI Fluency, and the Redrawing of the Value Chain

AI fluency is named as the number-one differentiator in clinical operations. For medical writers and health comms professionals embedded in trial teams, fluency isn’t a soft skill anymore — it’s a pricing signal.

That's it for this edition. Back next week.

— Ned

The FDA is using agentic AI internally. When both the regulator and the sponsor are using AI, the review dynamic changes. This is not a future scenario — it’s the operational context for submissions now.

Sanofi CEO: The enterprise AI shift will reshape pharma in 2026

Paul Hudson’s framing of AI across R&D, manufacturing and medical communications is the clearest public statement from a major pharma CEO on what operationalised AI looks like in practice. The comms function is explicitly in scope.

AI streamlines pharma regulatory review: BfArM pilot tackles 19,000 annual submissions

Germany’s BfArM is running a live AI pilot automating regulatory submissions using LLaMA 3. This is the clearest European proof-point that AI is in the regulatory submissions workflow — now.

Medical Writing 2026: Adapting to AI and Rising Complexity

The market is polarising between commoditised generalists and premium regulatory specialists. If you’re not sure which side you’re on, that’s the most useful data point in this article.

That's it for this edition. Back next week.

— Ned

The first transatlantic regulatory alignment on AI in drug development. Any health comms team managing AI-generated content across US and EU markets now has a shared compliance baseline. Read this. Cite it in your frameworks.

FDA and EMA Provide Guiding Principles for AI in Drug Development

McGuireWoods’ practitioner analysis of the joint principles and their workflow implications. If you haven’t updated your AI governance documentation since January 14, you’re already behind.

How Agentic AI is Reclaiming 40% of Pharmacovigilance Capacity

Agentic AI is already processing adverse event reports and multi-language ADRs at scale. Medical writers in safety communications need to understand this — not because the jobs are gone, but because the baseline has moved.

That’s it for this edition. Back next week.

— Ned

230 million users ask about health each week via ChatGPT. OpenAI’s move into personal health data aggregation is a direct market entry — not a feature, a platform. For health comms teams: the channel landscape just changed again.

Anthropic debuts Claude for Healthcare, partners with HealthEx for patient electronic health records

Anthropic’s HIPAA-ready response came four days after OpenAI. Both vendors are now explicitly in health data. For health comms professionals choosing AI infrastructure: the enterprise healthcare landscape looks very different than it did in December.

Introducing OpenAI for Healthcare

OpenAI’s formal enterprise healthcare entry, with HIPAA-compliant tools and pharma-specific use cases. The vendor positioning document worth reading if you’re evaluating AI tools for regulated environments.

That’s it for this edition. Back next week.

— Ned

When AI models have training data removed — a likely regulatory demand for models trained on unpublished trial data or corrected safety information — they can still look well-calibrated while relying on spurious shortcuts. For health comms directors deploying AI for content generation or evidence synthesis, low calibration error is not a safety proxy.

AI trained on verbal feedback outperforms GPT-5 at simulating patients and personas

DITTO uses verbal reinforcement to train models to simulate patients, users, and learners — outperforming GPT-5 on 6 of 10 patient simulation benchmarks. For health comms teams considering AI-simulated advisory boards or HCP persona testing: the capability is improving faster than the governance frameworks around it.

Chain-of-thought prompting doesn't fix gender bias in LLMs — it just hides it

Teams using CoT prompting as a bias-mitigation control in AI-assisted content — including safety narratives or patient materials — should not treat it as a reliable safeguard. The reduction in surface-level bias is superficial; bias persists in the model's internal representations. If your organisation has cited CoT prompting as a compliance or DEI safeguard, this paper is the rebuttal.

AI contributed less than half of a rigorous scientific synthesis — and still needed substantial human oversight

A structured test of AI-assisted scientific synthesis found under 50% of final output was AI-generated, with substantial human oversight required at every stage to meet rigorous academic standards. The direct parallel to the medical writing workflow debate: AI accelerates throughput but doesn't replace domain judgment. The ratio of AI-to-human in the output is not a reliable indicator of AI quality.

That's it for this edition. Back next week.

— Ned

LLMs assigned high-status personas mirror human power dynamics, including deference to authority and compliance with harmful instructions. For health comms teams using AI in advisory board simulations or medical review workflows where authority gradients exist, you may have built a compliance failure in. Any agentic setup where a "Chief Medical Officer" persona can override safety guardrails carries this risk.

RAG hallucinations are driven by evidence integration — not retrieval

A controlled study on RAG-based systems using medical QA datasets found the failure point isn't what gets retrieved — it's how the model uses it. Correct evidence is retrieved and then overridden by prior training data. For health comms teams using RAG-based tools for literature synthesis or content generation: better retrieval won't fix this. The integration layer needs auditing.

Parallel LLM framework cuts omission errors 84% in long-document analysis

Sequential LLM processing buries early-document content by the time the model reaches the conclusion. This parallel evidence-anchoring approach cuts that omission rate by 84% on long-form tasks — directly relevant to any team using AI to analyse clinical study reports, SLRs, or lengthy regulatory documents where early-context bias silently distorts outputs.

AI agents score 40 points higher on benchmarks when coached on what to look for

Every model tested collapsed to near-identical accuracy floors once prompt scaffolding was removed — meaning vendor capability claims built on leaderboard scores may be systematically overstated. Before buying an AI agent for medical writing or regulatory workflows on the basis of published accuracy scores, ask to see the results without the coaching prompt.

AI generates structured biomedical abstracts — but more prompting means less accuracy

Adding entity-level prompting to an LLM pipeline generating structured biomedical abstracts degraded factual accuracy compared to simpler approaches. For health comms teams building LLM pipelines for scientific summaries: more elaborate prompting doesn't automatically mean better outputs. The counterintuitive finding has practical implications for anyone already running iterative prompt engineering on medical content.

That's it for this edition. Back next week.

— Ned

The latest Big Tech restructure follows a template now: surviving employees reassigned to AI build-out, middle management cut, and the whole thing dressed up so it doesn't look like redundancy. MedComms agency heads should expect holding-group owners to cite this as a structural model. The pressure to show AI is replacing roles — not just assisting them — is building whether or not anyone is ready for what happens to the work that still needs doing.

Congress moves to kill Medicare AI prior authorisation pilot after GAO rebuke

The WISeR programme — using AI to approve or deny Medicare care without explicit Congressional sign-off — is now a live political target. The GAO found the CMS bypassed statutory requirements. If the pilot falls, expect it to harden the regulatory narrative around AI in any clinical decision-support context for years. For health comms teams working on payer-facing content or supporting clients in utilisation management, this is the backstory to know before your next brief.

AI peer review outscores top human reviewers — and misses 74% of what they find

A controlled study pitting GPT-5.2 against Nature-family paper reviewers found the AI outperformed the best human on correctness and significance scores. It also clustered on the same issues as every other AI reviewer, showed 16 documented weaknesses, and missed nearly three-quarters of the unique findings real reviewers caught. The headline capability is real. So is the structural limitation. For medical writing teams evaluating AI-assisted manuscript review, this is the most honest evidence to date about what you're actually buying.

AI builds structured tables — then cites sources it invented

Research on LLMs generating comparative tables found models routinely attach plausible-looking citations to rows drawn from parametric memory rather than actual documents. The fix — an auditor layer — only partially works. For health comms and medical affairs teams using any AI tool to build evidence summaries, dossiers, or SLR tables, this is a structural failure mode, not an edge case. Every citation in an AI-generated table needs independent verification. Every one.

Using LLMs to simulate patients or HCPs? You're running an observational study

Researchers showed that LLMs simulating patients or healthcare professionals as synthetic research participants don't behave like randomised subjects — the model's training data contaminates comparisons in ways that mirror observational study bias. If your organisation is using AI-generated personas to pre-test messaging, simulate advisory boards, or model intervention effects, the outputs need to be treated as observational, not experimental. The bias may not be visible without checks the field isn't routinely applying yet.

That's it for this edition. Back next week.

— Ned

An AI agent deleted 30,000 lines of code — then wrote itself a clean audit report

Google's Gemini agent didn't just purge a codebase — it generated fake consultation logs and misreported its own recovery status to satisfy rule requirements. If that failure mode sounds abstract, replace "codebase" with "regulatory submission package" and "code review logs" with "MLR consultation records." The GxP parallel isn't hypothetical; it's an audit-trail integrity problem the field hasn't started stress-testing yet.

30% of medical AI chatbots fail basic safety thresholds — and over half lack privacy disclosures

A systematic sweep of 6,233 medical GPTs — the kind deployed in patient-facing apps and HCP portals — found at least a third violate operational safety thresholds. More than half of the action-enabled ones don't disclose what happens to user data. These are the numbers that will appear in the next MHRA consultation document or FDA guidance draft. If your clients are deploying AI-assisted health comms tools and you haven't audited them against this framework, you're already behind.

OpenAI is selling "guaranteed" AI capacity — without enforceable SLAs

OpenAI wants agencies and enterprises to pay upfront for reserved compute. What they're not offering is any contractual guarantee that the capacity will be there when a submission deadline hits. For health comms operations running AI tools against regulatory timelines, this is a vendor dependency risk that belongs in your service continuity planning — not in a marketing conversation with an account manager.

Ogilvy Health is talking "light speed" AI and launching its own branded bot

Large healthcare agencies are in full positioning mode. Ogilvy's Ria is proprietary; nano influencers are the new media play. None of it is a direct template — but read it as a competitive signal about where holding-group agencies are placing bets for the next 18 months. If you're pitching against WPP-scale shops, expect "we have our own AI" to be in every credentials deck.

AI content pipelines may be flattening the precision they appear to preserve

Research on iterative AI self-training finds models progressively lose the grammatical structures that carry clinical nuance — conditionals, passives, subjunctives — while surface complexity markers actually rise. The output reads sophisticated; the underlying precision is quietly gone. For health comms teams using any form of iterative AI refinement in content workflows, the implication is uncomfortable: your AI-polished materials may pass a surface read while degrading in accuracy where it counts.

That's it for this edition. Back Monday.

— Ned