Medical affairs exists to keep the science honest. Hold that next to the first global survey of how medical science liaisons actually use AI: 27% of MSLs now lean on AI tools to prepare for meetings with key opinion leaders, and the single most common use is literature review, at 22% [Cureus, 2026]. The function built to catch a fabricated citation is adopting the one technology best known for inventing them. And only 33% of the companies doing it have written a policy for how [Cureus, 2026].
The Problem
Medical affairs isn’t customer support in a lab coat. Its reason to exist is the line between scientific exchange and promotion, the line the FDA’s Office of Prescription Drug Promotion enforces. An MSL can discuss the data a sales rep legally cannot. The accuracy is the license. A fabricated study or an off-label claim slipped into a KOL conversation is not a typo. It’s a compliance event with the company’s name on it.
Now look at where the survey says AI is landing. Across 367 medical affairs professionals in 48 countries, the top MSL use cases are literature review (22%), data analysis (20%), and building presentations (16%) [Cureus, 2026]. Those are exactly the tasks where a large language model (LLM) fails in its signature way. It does not return “I could not find that.” It returns a confident, properly formatted citation to a paper that doesn’t exist. The Cureus authors say it without hedging: AI-generated scientific content “may contain inaccuracies, incomplete interpretations, or fabricated information” [Cureus, 2026]. The technology is being adopted fastest into the corner of the job where its worst failure mode does the most damage.
Picture the actual moment. An MSL preparing for a cardiology KOL asks an AI assistant to summarize the latest outcomes data and pull supporting references. It produces a clean paragraph, three citations, a relative risk reduction, and a journal name. Two of the citations are real. One is a plausible-looking fabrication, right author, right journal, wrong study, a number that was never published. On a slide, in front of a physician who trusts the science because medical affairs is supposed to be the honest channel, that single invented line is the whole problem. It is off-label by accident, unbalanced by accident, and indefensible on purpose once anyone checks.
The governance numbers make it worse. A third of companies (33%) have a policy governing MSL use of AI, and 31% of respondents do not even know whether their company has one [Cureus, 2026]. At the same time, 87% of these professionals say learning AI matters to staying competitive in the role [Cureus, 2026]. Read those together and you get the real state of play: near-universal pressure to adopt, almost no rules, and a workforce where 39% rate themselves not knowledgeable about the tool they feel pressed to use [Cureus, 2026]. Call it plainly: not a technology gap but a control gap. And control is the entire product medical affairs sells.
The Insight
The comforting story is that AI hallucination is a temporary defect the next model version fixes, but it’s not. Fabrication is structural to how these systems work. They predict plausible text, they don’t retrieve verified facts, and a confident wrong answer is indistinguishable on the page from a correct one. The proof is sitting inside the regulator. When the FDA deployed its own generative AI tool, Elsa, built on a frontier model inside a secure government cloud, reviewers found it “confidently hallucinates,” invented studies and regulatory citations, and was trustworthy only for tasks nobody had to double-check [Medical Economics, 2025]. CNN reported the tool was making up studies and misrepresenting research [Medical Economics, 2025]. If the agency that polices drug claims cannot keep fabrication out of its own AI, a medical affairs team running an off-the-shelf AI assistant should assume it can’t either.
So the useful question stops being “is the AI accurate” and becomes “what verifies the AI.” That reframes the entire build.
“The medical affairs AI that wins will not be the one that drafts the fastest. It will be the one that can prove where every sentence came from.”
Here is the part a cautious team would rather not put on a slide. Everyone is racing to automate the medical, legal, and regulatory review bottleneck, to push approved content out faster. That is backwards. The review bottleneck is the moat. Speeding up generation without scaling verification doesn’t remove the risk, it relocates it downstream to the KOL conversation and to OPDP, where it is most expensive to fix. The edge is not a faster writer. It is a verifier: a system that grounds every claim in approved, source-linked content and refuses to make claims it can’t trace. The FDA already signaled where this goes. Its draft guidance on AI used to produce evidence for regulatory decisions is built around a risk-based credibility assessment, judging a model by the stakes of its context of use, not its fluency [FDA, 2025]. Origination, not speed, is the regulatory tell.
This is familiar ground for pharmacy, which is why the parallel earns its place. Pharmacists already run a rank-and-release discipline where the machine flags and the human holds the authority to release. The output is never trusted because it sounds confident. It’s trusted because it was checked against the source, by a named professional, before it reached the patient. Medical affairs needs that same posture before any AI output reaches a KOL. The survey’s warning is simple: lean too hard on the AI draft and your scientific rigor fades, because the more an MSL trusts it, the less they question it.
In Practice
Not every medical affairs task carries the same failure cost, so the verification gate should scale with what a fabrication would actually break.
| AI use case | What a hallucination costs | Verification gate |
|---|---|---|
| Internal meeting notes, admin drafting | Wasted time, caught in-house | Light: spot-check, low stakes |
| Literature search and evidence synthesis | A fake or misread citation enters the scientific record | Hard: every reference traced to the primary source before use |
| Content for KOL and HCP scientific exchange | Off-label or unbalanced claim, OPDP exposure | Hard gate: human medical review, source-linked, logged |
| Field insight capture and KOL targeting | Skewed strategy, biased prioritization | Medium: human review of the model’s logic and inputs |
Two rules make this work. First, no AI-written scientific claim goes to an outside audience until a specific person checks it against the original source it cites. Not the summary the AI read, the actual source. Second, treat “can this tool show me where each sentence came from?” as a deal-breaker when buying it, not a bonus. A tool that can’t point to the exact source for every sentence isn’t saving you time. It’s a liability with a nicer interface. The stakes climb higher with evidence generation and real-world-evidence work: when AI helps shape an analysis that lands in a regulatory submission or a published paper, the model’s credibility becomes the evidence’s credibility, and that’s exactly what the FDA’s risk-based framework is designed to test [FDA, 2025].
The Bottom Line
The survey describes a job moving fast with the brakes barely bolted on. AI is already in use, mostly for literature review and data analysis, but only a third of companies have any rules governing it [Cureus, 2026]. The risk isn’t spread evenly. The teams most exposed are the ones treating AI as a shortcut for scientific exchange, because the moment a made-up citation reaches a KOL or a regulator, the “we’re faster now” story turns into a compliance problem, and compliance is the only story a general counsel cares about. The teams that come out ahead are the ones already built around verification, that buy or build AI whose first job is to prove where its facts come from, not to write quickly.
For investors watching the medical affairs software wave, the same line separates winners from losers. The space is packed with tools that draft and summarize, and they all demo beautifully. The companies that last will be the ones that make source-tracing the actual product, because that traceability is about to go from nice-to-have to legally required, and a feature that becomes a requirement is a business. Speed is a demo. Knowing where every fact came from is the moat.
The MSLs who get ahead won’t be the earliest adopters. They’ll be the ones who never let the tool speak for them without checking, in a job where getting the science right was always the whole point.