Does AI drafting make wrong-word errors on drug labels moot?

No. AI changes the kind of error, not the need to check. Whatever drafts the label, someone still has to compare it against the approved text and read the high-stakes lines for meaning. A correctly spelled wrong word is fluent, plausible and wrong, and fluent, plausible and wrong is exactly what language models are known to produce.

What regulators have said

Europe's medicines regulator has addressed this almost word for word. In its reflection paper on AI in the medicinal product lifecycle, the European Medicines Agency (EMA) writes, in section 2.3.5 (Product information):

AI/ML applications used for drafting, compiling, editing, translating, tailoring, or reviewing medicinal product information documents should be used under close human supervision. Given that generative language models are prone to include plausible but erroneous or incomplete output, quality review mechanisms need to be in place to ensure that all model-generated text is both factually and syntactically correct before submission for regulatory review.

The US picture is different, and worth stating carefully. FDA's draft guidance on using AI to support regulatory decision-making expressly leaves out AI used for drafting or writing a regulatory submission. So FDA has not written an AI-specific rule for drafted labeling text. It also hasn't relaxed anything. The existing labeling controls apply to the final label no matter how the first draft was made.

The same fluent-but-wrong failure, documented in clinical text

We have not found a published study that measures correctly spelled word swaps in AI-drafted drug labels, and we won't pretend one exists. What the peer-reviewed literature does document is the same failure in nearby text:

• Clinical summaries. Studies of AI-drafted clinical notes and emergency department summaries found hallucinations and omissions, with hallucinations most common in the plan section, the part that tells people what to do next. In one study of AI notes written from consultations, 1.47% of sentences contained a hallucination, and the authors flagged negations, where a statement's meaning flips, as the most concerning type.
• Over-correction. Research on AI grammar correction in general English describes models rewriting text that was already correct to make it read more smoothly, a problem that persists even with carefully designed prompts and examples. On a label, approved wording that gets smoothed is wording that has changed.
• Translation. In one study of machine-translated emergency department discharge instructions, 2% of Spanish and 8% of Chinese sentences had errors with potential for significant harm. In another, AI translations of pediatric discharge instructions held up against professional translators in Spanish but did worse in Chinese, Vietnamese and Somali.

None of that says AI put a wrong word on a specific label. It says the mechanism is real: the output reads well, and reading well is what lets it through. A contextual word error on FDA DailyMed works the same way. A heart-risk warning said "attach" where it needed "attack", and the line still scanned. Read the case: FDA DailyMed NSAID warning said "attach" instead of "attack".

The fair counterpoint

AI is not the villain here, and the evidence doesn't say it is. Tuned pipelines with guardrails have matched people or beaten earlier systems on some tasks. At an online pharmacy, a domain-guarded system for medication directions cut direction near-misses (errors a pharmacist caught before they reached the patient) by 33% compared with the system it replaced. On the same task, the best general-purpose model, given only example prompts, produced 4.38× more near-misses than the guarded system. A study of hospital discharge summaries found AI and physicians made a similar number of hallucinations, with similar harm per error, although the AI summaries had more errors overall (2.91 vs 1.82 per summary).

Look at how those results were reached, though. Each one paired the model with an error taxonomy, pharmacist review or clinician annotation of the output. The verification step was part of the win. And matching people is the argument for keeping the check, since people's errors are why the check exists.

What this means for a labeling team

• Use AI where it helps. First drafts, reformatting, consistency passes.
• Compare to the approved text. Treat any reworded required statement as a change until someone confirms it.
• Read the key word on each high-stakes line. Warnings, directions, indications and first aid get read for what they tell the reader to do.

More on the review side: does AI-drafted labeling still need human review? To see what gets through when that step is missed, browse our case index.

FAQ

More on contextual errors in FDA DailyMed labels.

We scan freshly published FDA DailyMed labels every day and pull out these high-stakes wrong words.

Every Tuesday and Friday, a new case like this lands in your inbox if you subscribe.