AI can draft large parts of a 510(k), but it cannot take responsibility for them: every 510(k) must include your statement that all its data and information are truthful and accurate. What holds up in FDA review is content you can trace to your own evidence and FDA's public records, checked by a qualified person before you submit.
Key takeaways
- Every 510(k) must include the submitter's statement that all data and information are truthful and accurate, under 21 CFR 807.87(l). AI changes who drafts, not who answers for it.
- We found no FDA guidance on using AI to draft a device submission, as of September 26, 2026. The existing review standards apply unchanged.
- FDA uses AI itself: its Elsa tool launched on June 2, 2025, and FDA says human experts verify its inputs, analysis and outputs.
- In a 2025 study of AI-written clinical notes, hallucinations appeared in 1.47% of 12,999 note sentences, about 1 in 68, and 3.45% of source sentences were left out.
- FDA cleared about 3,200 510(k)s in 2025, and the eSTAR template has changed 13 times since October 2023, so a tool with stale data will miss both.
Can AI write a 510(k)?
It can write a first draft of many sections, but not a submission you can sign unchecked. Every 510(k) must include a statement that the submitter believes, to the best of their knowledge, that all data and information are truthful and accurate, and that no material fact has been omitted [1]. That statement is yours, whoever or whatever drafted the text.
Match the task to the risk. A practical way to split the work:
| Tier | Examples | How to use AI |
|---|---|---|
| Low risk | Formatting, consistent terminology, summarizing documents you supply, first drafts of the device description from your specifications | Draft with AI, then review |
| Verify every fact | Predicate comparisons, standards lists, guidance references, labeling text | Draft with AI only from sources you supply, then check each fact against FDA's records |
| Expert authorship | Pathway choice, the substantial equivalence argument, interpretation of test results, responses to FDA | A qualified person writes it; AI can help with structure and wording |
The pattern is simple: the closer a section sits to a regulatory judgment, the less of it AI should write. Our 510(k) hub walks through each part of a submission.
What does FDA say about using AI to draft a submission?
We found no FDA guidance on using AI to draft a device submission, as of September 26, 2026. So the existing rules apply: the truthful and accurate statement, and FDA's normal review [1].
This article is about AI used to prepare a submission. If your device itself uses AI, see our FDA AI Device Tracker.
FDA separates writing help from AI that produces evidence. Its January 2025 draft guidance on AI for drug and biologic decisions leaves out AI used for "drafting/writing a regulatory submission", when that use does not affect patient safety, product quality or the reliability of study results [4]. It is not a device guidance, but it shows where FDA draws the line: AI that generates evidence gets scrutiny, and AI that helps you write does not change what the evidence must show.
FDA's reviewers use AI too. FDA launched Elsa, a generative AI tool for reading, writing and summarizing, on June 2, 2025, and says it does not train on data submitted by industry [2]. With Elsa 4.0, announced on May 6, 2026, FDA said its staff are involved at every stage, so that human experts verify all inputs, analysis and outputs [3]. That is the same standard to apply to your own AI drafts.
Where does AI help most in 510(k) work?
In search and structure, where the output is easy to check. These are the tasks where AI saves time, and where its output is easiest to check:
| Task | What AI does well | What you must check |
|---|---|---|
| Guidance search | Finds the FDA guidance documents relevant to your device | That each is current, and whether it is final or draft |
| Predicate research | Shortlists cleared devices with similar intended use and technology | Each 510(k) number and its cleared indications, in FDA's database |
| Standards mapping | Lists recognized consensus standards that may apply | Each standard's recognition status and edition |
| Substantial equivalence tables | Lays out a side-by-side comparison from data you supply | Every cell against the predicate's public summary |
| eSTAR drafting | Drafts sections in the order the eSTAR template expects | Answers against the current template version |
| Consistency checks | Flags differences between intended use, labeling and test reports | That each flag is real, and fixed at the source |
Our guides to choosing a predicate device and to the eSTAR template cover the human side of these tasks.
Where does AI-drafted content fail FDA review?
Where it cannot be traced, or is out of date. Five failure types stand out:
- Invented or wrong references. A 510(k) number that does not exist, a standard in the wrong edition, or a guidance title that is not quite right. Reviewers can check each of these in FDA's public 510(k) database and recognized standards database [7, 8].
- Outdated information. FDA cleared about 3,200 510(k)s in 2025 [6], and the eSTAR template has had 13 updates since it became mandatory on October 1, 2023 [9]. A model trained on older data will not know about new predicates, templates or guidance.
- Omissions. In the 2025 clinical-note study, the authors found omissions more common than hallucinations: 3.45% of source sentences were left out, against 1.47% of note sentences that were invented or wrong [5]. Hallucinations were more often serious, though: 44% were rated major, against 16.7% of omissions [5]. A 510(k) can fail on what is missing, not only on what is wrong.
- Inconsistency across sections. An intended use that shifts between the device description, the labeling and the test reports.
- Persuasive but unsupported claims. Fluent text that argues substantial equivalence beyond what your data shows.
The study is about clinical notes, not submissions, and its error rates changed as the authors refined prompts and workflows [5]. Its lesson still applies: at 1.47%, about 1 sentence in 68 contained a claim a clinician had to catch.
Errors that surface at screening cost time. Our FDA RTA checklist covers what gets a 510(k) placed on hold.
How do you check AI output before you submit?
Work from sources, not from memory, and verify every fact. A practical protocol:
- Supply the sources. Give the tool your specifications, test reports and the predicate's 510(k) summary. Do not ask it to find evidence for you.
- Trace every fact. Each number, standard and reference in the draft should point to a document you can open.
- Check FDA identifiers in FDA's databases. Confirm 510(k) numbers and predicate indications in the 510(k) database, and standards in the recognized standards database [7, 8].
- Check guidance status and date. Confirm each guidance in FDA's guidance search, and whether it is final or draft [10].
- Run a consistency pass. Read the intended use in every section side by side.
- Get expert sign-off. A qualified person approves each section before it goes into the eSTAR.
- Keep a record. Note which sections AI drafted, and who checked them.
How do you choose an AI tool for 510(k) work?
By whether you can verify what it produces. Ask five questions:
- Does it cite sources you can open, such as FDA database records and guidance documents?
- Is its FDA data current, including recent clearances and the current eSTAR version?
- Does it keep your data private, and not train on your documents?
- Does it fit your review process, so a person approves each section?
- Can you reach a regulatory expert when the answer needs judgment?
General-purpose AI assistants are useful for wording and summaries of documents you supply, but they may not know recent FDA records. Regulatory AI platforms aim to connect drafting to FDA's data. Consultants who use AI add expert accountability.
Complizen's Drafting Studio drafts sections such as the device description and the substantial equivalence summary from your device data, with citations, for your team to review.
Frequently asked questions
Can AI write my 510(k) submission?
AI can draft many sections, but you remain responsible for all of it. Every 510(k) must include your statement that all data and information are truthful and accurate. Use AI for first drafts and structure, and have a qualified person verify every fact and write the regulatory judgments.
Is it allowed to use AI to prepare an FDA submission?
We found no FDA guidance on AI-drafted device submissions, as of September 26, 2026, so the existing requirements apply, including the truthful and accurate statement. How a section was drafted does not change what it must show, or who answers for it.
Does FDA use AI to review submissions?
FDA uses Elsa, a generative AI tool it launched on June 2, 2025, for reading, writing and summarizing. At launch, FDA said it was already using Elsa to speed clinical protocol reviews and scientific evaluations. FDA says Elsa does not train on data submitted by industry, and that human experts verify all inputs, analysis and outputs.
What is the biggest risk of using AI for a 510(k)?
Content you cannot trace: invented or outdated references, omissions, and claims that go beyond your data. Reviewers can check 510(k) numbers, standards and guidance in FDA's public databases, so an error that would pass a quick read can still surface in review.
How often does AI make mistakes in medical writing?
It depends on the setup. In a 2025 study of AI-generated clinical notes, 1.47% of 12,999 note sentences contained hallucinations, and 3.45% of the source transcripts' sentences were left out. The study was about clinical notes, not regulatory submissions, but it shows why every sentence needs checking.
Can AI find my predicate device?
AI can shortlist candidate predicates with similar intended use and technology, which saves search time. You must then confirm each one in FDA's 510(k) database: that the number exists, its cleared indications, and that the comparison holds. The final choice is a regulatory judgment.
Should I use ChatGPT or a specialized regulatory AI tool?
General-purpose assistants work well for wording and for summarizing documents you supply. For FDA-specific facts, such as recent clearances, recognized standards and the current eSTAR version, prefer a tool that cites FDA's own records. Either way, verify every fact. Check whether the tool shows where each fact came from.
Which parts of a 510(k) should a person write?
The regulatory judgments: the pathway choice, the substantial equivalence argument, the interpretation of test results, and responses to FDA's questions. AI can help structure and word these sections, but a qualified person should decide what they say. Those sections carry the conclusions your truthful and accurate statement covers.
Can AI make my 510(k) fail FDA screening?
Not by itself, but errors it introduces can. Inaccurate answers, missing attachments and inconsistent information can put an eSTAR on hold at technical screening. Check AI-drafted answers against the current template version and your actual documents. FDA's screening checks the eSTAR for completeness before substantive review begins.
Is my data safe in a general-purpose AI tool?
It depends on the tool's terms. Before you share device data, check whether the tool trains on your inputs, where it stores data and who can access it. FDA's own Elsa does not train on industry data, which is a reasonable standard to ask for.
Will AI make 510(k)s faster?
AI can shorten research and drafting, but the FDA review itself does not change. Faster drafting helps only if the result is accurate, consistent and complete. A submission that goes on hold, or draws additional information requests, loses the time AI saved.
How should I document AI use in my 510(k) work?
Keep a simple record of which sections AI drafted, which sources it used, and who reviewed and approved each section. It helps your team trace any fact FDA asks about, and supports your truthful and accurate statement. Update the record when a section changes after review.
References
- eCFR. 21 CFR 807.87, Information required in a premarket notification submission, paragraph (l). Accessed September 26, 2026.
- FDA. FDA Launches Agency-Wide AI Tool to Optimize Performance for the American People. Press release, June 2, 2025.
- FDA. FDA Expands AI Capabilities and Completes Data Platform Consolidation. Press release, May 6, 2026.
- FDA. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. Draft guidance, January 2025.
- Asgari E, et al. A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. npj Digital Medicine 8, 274, May 13, 2025.
- openFDA. Device 510(k) API. 2025 decisions, accessed September 26, 2026.
- FDA. 510(k) Premarket Notification database. Accessed September 26, 2026.
- FDA. Recognized Consensus Standards: Medical Devices. Accessed September 26, 2026.
- FDA. eSTAR Program. Content current as of September 21, 2026; update count from our analysis of the template's version history.
- FDA. Search for FDA Guidance Documents. Content current as of September 24, 2026.