Using AI to Draft FDA Submission Content: The Risks That Will Catch You at Review

May 25, 2026

Key Takeaways

• AI-generated documentation can contain fabricated citations, mischaracterized test data, and procedurally inaccurate content, errors that may not be obvious until FDA finds them.

• FDA reviewers check sources and cross-reference your submission against the underlying data; inaccuracies in one section raise questions about the accuracy of the entire submission.

• AI can support document organization and drafting scaffolding, but it cannot replace the expert judgment required to accurately characterize data, apply current regulatory requirements, and formulate credible regulatory arguments.

We reviewed a client's submission draft last month that cited three studies to support a key safety claim. Two of them didn't exist.

A hallucinated citation in a regulatory submission is not a typo. It is a credibility problem with the reviewer who decides your premarket submission's fate.

AI tools are increasingly used in early-stage device development, and the appeal is understandable. They are fast, accessible, and the outputs look polished. But there are specific failure modes that appear consistently in AI-drafted regulatory documentation and founders need to identify and eliminate them before anything reaches FDA.

When the Output Looks Right but Isn't: Three Failure Modes We're Seeing

Bench Test Reports: Plausible Language, Inaccurate Substance

One of the most common patterns we encounter is a company using AI to summarize a bench test or performance test report. The AI generates a clean, readable paragraph, but the summary mischaracterizes what was actually measured, invents literature references to support the performance claims, or draws conclusions that exceed what the data supports.

The practical implication here is significant: test data in a premarket submission is not a narrative exercise. FDA reviewers evaluate whether your testing methodology, acceptance criteria, and results are traceable to a recognized standard, ASTM, ISO, or an applicable FDA guidance, and whether your summary accurately reflects the raw data in the test report. If those two things do not align, the submission will not pass review. In some cases, it triggers broader questions about data integrity across the entire package.

AI does not understand what a bench test is measuring or where that data sits in your overall regulatory argument. It produces text that sounds like a test summary because it has been trained on text that sounds like test summaries. That is a fundamentally different capability than analyzing a test report.

QMS Procedures: Missing What Changed

A second failure mode involves using AI to draft Quality Management System (QMS) procedures. FDA's Quality Management System Regulation (QMSR), finalized in February 2024 under 21 CFR Part 820, replaced the legacy Quality System Regulation and aligned FDA's device quality requirements with ISO 13485:2016 with a required compliance date of February 2, 2026.

What we are seeing is AI-generated QMS procedures that reflect the prior regulatory structure and naming conventions, think design controls instead of design and development. AI tools were not trained on the QMSR as a live, current regulatory requirement in the way a regulatory professional applies it. The output uses familiar language, follows a recognizable structure, and passes a surface-level review. It falls apart under a real compliance assessment.

An important caveat: QMS gaps discovered during an FDA inspection or at the time of a submission review are not minor administrative corrections. They are findings that affect your program timeline, your ability to conduct clinical studies under an IDE, and your commercial readiness. They also point out a lack of quality understanding and training of your team.

Clinical Evaluation Conclusions: Missing the Point of the Study

A third pattern involves using AI to draft the conclusion of a clinical evidence summary or Clinical Evaluation Report. This is where the gap between AI-generated content and expert analysis is most consequential.

A credible clinical conclusion requires three things AI consistently cannot provide: an accurate characterization of what the study actually showed, an honest assessment of the study's limitations, and a defensible analysis of the state of the art in clinical practice for the relevant indication. These are not writing tasks. They are analytical judgments that require understanding the clinical context of the device, the evidentiary standard FDA will apply under 21 CFR 860.7 (reasonable assurance of safety and effectiveness), and the specific gaps between your study and that standard.

AI will produce a conclusion that sounds authoritative. What it cannot do is synthesize that language with an intimate understanding of your specific study design — why the endpoints were chosen, what the results actually mean in the context of current clinical practice, and where the data genuinely falls short of what FDA needs to see. Those judgments require someone who has worked with the study from the ground up. Pattern recognition is not clinical expertise.

The submission looks complete. The analysis is not.

Where AI Can and Cannot Add Value

To be direct: AI is not categorically off-limits for regulatory work. There are tasks where it genuinely helps.

AI can support:

• Organizing document structure and section templates

• Drafting boilerplate language for administrative sections

• Formatting and consistency checks across a large submission package

• Identifying content gaps against a document outline or regulation

AI cannot replace expert judgment for:

• Substantiating technical claims with traceable, verified evidence

• Accurately characterizing performance data against what was tested and what the results actually support

• Writing clinical conclusions that reflect the study's findings, limitations, and clinical context honestly

• Formulating regulatory arguments tailored to your specific device, indication, and submission type

Questions to Ask About Your Program

• Has a regulatory expert reviewed any AI-generated content in your submission, not just for structure, but for factual and regulatory accuracy?

• Can every citation in your submission be traced to a real, accessible source that says what your document claims it says?

• If AI was used to summarize test data, has an engineer or regulatory expert verified that the summary accurately reflects the raw data and acceptance criteria in the test report?

• Does your clinical evidence conclusion address the limitations of your data and the FDA standard your submission will be evaluated against?

If you are using AI tools to support your regulatory documentation and want an expert review before anything goes to FDA, we are happy to think through it with you. Reach out to FCG.