Generative AI can now draft large parts of regulatory submissions, from clinical study reports to summaries, in a fraction of the time human writers need. Regulators do not prohibit this. But the sponsor remains fully accountable for every sentence and number submitted. AI moves the work from writing to verification, and verification is where the risk now sits.
A clinical study report for a single trial can run to hundreds of pages. It describes how the trial was designed and run, who took part, what happened to them, and what the data show, table by table. A marketing application bundles many such documents with summaries, safety analyses and product information, in the structure set by the international Common Technical Document. For decades, teams of medical writers have produced these documents by hand, over weeks and months, as a project’s critical path waited.
That is exactly the kind of work large language models are good at, and the industry has noticed. Novo Nordisk, in a case study published by Anthropic, described an internal system called NovoScribe that it says cut clinical study documentation from more than ten weeks to about ten minutes, with drafts going to human reviewers for approval. Vendors across the industry now promise similar compression for protocols, summaries and responses to regulators’ questions.
Treat vendor figures with the usual caution. But the direction is not in doubt. Drafting is becoming cheap.
Here is the argument of this piece. When drafting becomes cheap, the scarce and valuable work becomes checking. A submission is not a piece of writing. It is a set of claims a company makes to a regulator, under legal obligations, about data that determine whether patients get a drug. Generative AI can write those claims. It cannot be accountable for them. The sponsors who benefit most will be the ones that redesign review around that fact, not the ones that simply produce drafts faster.
Can AI write a regulatory submission?
Large parts of one, yes.
Much of a submission is highly structured and derived from other documents. Methods sections restate the protocol. Results sections describe tables and figures produced by statistical programming. Safety narratives summarize individual patient records in a standard format. Summaries condense clinical study reports. Language models can generate first drafts of all of these, cross-reference sections and harmonize terminology.
Other parts are harder. Integrated summaries of efficacy and safety require judgment about how to present conflicting results, what caveats matter and how findings relate to the proposed use. Responses to regulators’ questions need an understanding of what the reviewer is really concerned about. The benefit-risk discussion is, in effect, the company’s argument for approval. A model can draft these, but the draft is a starting point for expert argument, not a substitute for it.
What do regulators say about AI in submissions?
Regulators have mostly addressed a related but different question: when AI produces data or analyses that support decisions about a drug’s safety, effectiveness or quality.
In January 2025, the FDA published a draft guidance, “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products.” It sets out a risk-based credibility assessment framework: sponsors define the question an AI model addresses and its context of use, then assess risk based on how much the model influences the decision and how serious the consequences of an error would be. The more influence and the higher the stakes, the more evidence of credibility the FDA expects.
The European Medicines Agency’s reflection paper on AI across the medicinal product lifecycle, published in draft in 2023 and finalized in 2024, takes a similar human-centered, risk-based view and places responsibility on applicants and marketing authorization holders to ensure AI tools they use are fit for purpose.
Neither document is mainly about using a language model to write prose. The underlying principle carries over anyway. Regulators judge what is submitted, not how it was typed. Existing obligations, including the requirement that submissions be accurate and complete and the legal consequences of false statements to a regulator, apply whether a sentence came from a writer or a model.
Regulators are also adopting these tools. In June 2025 the FDA launched Elsa, an agency-wide generative AI assistant to help staff with reading, writing and summarizing, including in clinical protocol reviews, stating that its models do not train on data submitted by industry. The likely future is AI-drafted submissions being read, in part, by AI-assisted reviewers. That raises the value of submissions that are precise, consistent and easy to verify.
Where does AI-drafted regulatory writing go wrong?
Three failure modes stand out.
Plausible errors. Language models produce fluent text that reads as correct. In a submission, the dangerous error is not an obvious mistake. It is a transposed number in a safety table description, an adverse event described with the wrong severity, or a sentence stating that a result was statistically significant when it was not. These are errors a hurried reviewer, reading fluent prose, is likely to miss.
Invented references. Models can generate citations to studies that do not exist or misattribute findings to real ones. The risk is not hypothetical in public documents. In May 2025, the US Make America Healthy Again Commission’s report on childhood chronic disease was found by journalists to cite studies that did not exist, and the text was revised. A submission with a fabricated reference would seriously damage a sponsor’s credibility with the reviewer.
Automation complacency. When drafts are usually good, reviewers start skimming. We made this argument about clinicians in the real risk of AI-drafted patient messages: the risk is less the model than the human who agrees with it. Regulatory reviewers face the same pressure, amplified by deadlines.
A worked example: one wrong number in a safety section
Consider a hypothetical phase 3 program, with illustrative details.
A sponsor uses a language model to draft the safety section of a clinical study report from its statistical output. The draft is excellent: clear structure, consistent terms and every table referenced. In one paragraph, it states that serious liver injury occurred in 3 patients on the drug and 3 on placebo. The table it was drawn from shows 5 on the drug. The model conflated two rows.
The medical writer reviewing the draft checks the conclusions and the flow. Everything reads well, and the numbers look plausible. The paragraph passes. The integrated safety summary, also drafted with AI from the study reports, inherits the error and says liver injury was balanced between arms.
Months later, a regulatory reviewer, cross-checking text against tables, finds the discrepancy. At best, the sponsor faces an information request and a correction. At worst, the reviewer now distrusts every narrative in the application and asks for more analyses, delaying the review far more than AI saved in drafting.
What would have caught it? Not better prose. A verification step designed for the way AI fails: automated checks that every number in the text matches its source table, a requirement that every factual claim link back to a source, and human reviewers told to verify numbers rather than read for sense. The time saved in drafting has to be spent, in part, on this.
The strongest case for AI-drafted submissions
The counterargument deserves a hearing, and parts of it are strong.
Human-written submissions already contain errors. Writers working late on a deadline copy numbers wrongly, carry forward outdated text and describe results inconsistently across documents. An AI system built to draft directly from source data, with every number traceable, could make fewer transcription errors than tired people, not more.
There is also a patient benefit. If AI shortens the time to prepare submissions, effective drugs can reach review sooner. Writing is rarely the longest part of drug development, but on the critical path at the end of a program, weeks matter.
And AI can raise consistency. A single application contains dozens of documents written by different people. Models can harmonize terminology and detect contradictions across documents faster than humans can.
All of this is persuasive, under one condition: the system is designed so that every claim is traceable to its source and checked. Deployed that way, AI can make submissions both faster and better. Deployed as a fast drafting tool feeding a traditional review process, it produces more polished documents with errors that are harder to spot.
What this means for sponsors, medical writers and regulators
For pharmaceutical sponsors, the implementation question is not “which model?” but “what does our review process look like now?” Budget time saved in drafting for verification. Require source traceability for every number and claim. Prefer systems that generate text from validated data outputs over systems that summarize summaries. Document how AI was used, even where not yet required, so that questions from regulators can be answered quickly.
For medical writers, the job is changing rather than disappearing. The value shifts from producing text to designing templates, checking accuracy, and shaping the argument in the parts of a submission that require judgment. The lesson from ambient scribes, discussed in our piece on AI scribes and burnout, applies here: time saved gets spent somewhere, and organizations that spend it on more volume rather than better checking will regret it.
For regulators, clear expectations would help. It would be useful to say whether and how sponsors should disclose AI use in document preparation, and to reinforce that accountability does not change. Regulators using AI to assist reviewers should be equally transparent about how those tools are validated.
For patients, the stakes are indirect but real. Submissions are how regulators learn what a drug does. Speed that comes at the expense of accuracy in safety data is not a trade anyone should accept on their behalf.
Frequently asked questions
Does the FDA allow AI-written regulatory submissions?
The FDA has not prohibited using AI to help draft submissions. Its focus, set out in a January 2025 draft guidance, is on AI used to generate information supporting regulatory decisions, assessed through a risk-based credibility framework. Regardless of tools used, sponsors remain responsible for the accuracy and completeness of everything they submit, and existing legal obligations apply.
What parts of a regulatory submission can AI draft?
AI is most useful for structured, data-derived content: methods sections based on protocols, descriptions of results tables, patient safety narratives and summaries of longer documents. It is less reliable for content requiring judgment, such as benefit-risk discussions and responses to regulators’ questions. In every case, drafts need expert review and verification against source data.
Who is responsible for errors in an AI-drafted submission?
The sponsor. Regulators hold the company submitting an application accountable for its content, regardless of whether a person or a tool drafted it. Internally, companies assign responsibility to the qualified people who review and approve documents. Using AI does not transfer accountability to the software vendor or reduce the consequences of submitting inaccurate information.
How do you check AI-generated regulatory documents?
Effective checking is designed for how AI fails. That means automated comparison of every number in the text with its source table, links from each factual claim to its source, verification that every cited reference exists and supports the statement, and instructions for reviewers to check facts rather than read for flow. Consistency checks across documents in the same application also help.
The takeaway
Generative AI is about to make regulatory writing much faster. That is good news only if review changes with it. A submission is a set of promises made to a regulator about patients’ safety, and every one of them still needs a person willing to stand behind it.
The winning teams will not be the ones that draft fastest. They will be the ones that can prove, line by line, that what they drafted is true.
If your team uses AI in regulatory writing, what changed in your review process? Tell us what works, and what went wrong.
