LumaBlog

How medical writers can use AI without risking fabricated references

Medical writers can use AI safely by keeping it in the drafting role and never treating it as a source. Every reference in the final document should come from a record you looked up yourself, and every claim should be checked against what that record actually says.

Used that way, AI can help with structure and wording without putting invented references under your name.

Why are AI-generated references a risk for medical writers?

Large language models can produce references that look complete and do not exist. A 2023 study by Walters and Wilder examined 636 citations in 84 literature reviews written by an earlier version and a newer version of a widely used chatbot. Of the earlier version's citations, 55 percent were fabricated, and 18 percent of the newer version's were. The reviews covered 42 topics from many fields, so the study was not specific to medicine.

In biomedicine, a 2026 audit by Topaz and colleagues in The Lancet reports that prior studies estimate 30 to 69 percent of references generated by large language models in biomedical contexts are fabricated. The same audit found 4,046 fabricated references across 2,810 papers, and 98.4 percent of affected papers had received no publisher action at the time of the audit.

For a writer, the lesson is direct: an error that leaves your desk may never be corrected, and your name stays on it.

What can AI do well in a medical writing workflow?

AI is useful for work where you can check the output against something you already have:

  • Outlining a document from your own notes or a brief.
  • Rewriting a passage for a different reading level, audience or tone.
  • Summarizing a paper you have open, so you can compare the summary line by line with the source.
  • Suggesting search terms and synonyms for a literature search you will run yourself.
  • Pointing out gaps or unclear steps in an argument.
  • Drafting plain-language versions of text you have already verified.

In every case, the facts come from you or from a source in front of you. The AI shapes the words.

What should AI never do in your workflow?

  • Supply references from memory. Any reference you did not find yourself stays unverified until you look it up.
  • Decide what a study found. Read the abstract or full text yourself.
  • Fill in numbers. Effect sizes, rates, sample sizes and doses must come from the source.
  • Strengthen a hedge. If a paper says a finding "may" hold, the draft should not say it does.

What does a safe workflow look like?

  1. Find the sources first. Search PubMed, the free search engine for biomedical literature maintained by the National Library of Medicine at the National Institutes of Health, and save each record with its PubMed ID (PMID) and digital object identifier (DOI).
  2. Give the AI only those sources, and ask it to draft from them without adding others.
  3. Tag every claim in the draft with its source, or with a note that it still needs one.
  4. Check every reference. Look up each PMID or DOI, confirm the authors, journal and year match, and read the abstract against the sentence. The full method is in how to check if a PubMed citation is real.
  5. Check every number against the source document.
  6. Do a final pass claim by claim instead of paragraph by paragraph, the idea behind claim-level verification.
  7. Log what you checked.

Giving the AI your own sources is not a guarantee. A model handed a paper can still misstate what it says, so the reference and number checks stay in place even when you supplied every source.

What should a verification log include?

A simple log turns your checking into a record that a reviewer, client or editor can follow. For each claim, note:

  • The claim as it appears in the document.
  • The source, with its PMID or DOI.
  • Where in the source the support appears, such as the abstract, a results table or a named section.
  • The outcome: supported, partly supported, not supported, or replaced.
  • Who checked it and when.

A log also saves time on revisions. When a reviewer questions a sentence, you can show what supports it without searching again.

What are the red flags in an AI draft?

Watch for these:

  • References you did not supply.
  • Titles that restate your claim almost word for word.
  • Reference entries with no PMID or DOI.
  • An identifier that opens a different paper than the one described.
  • Firm language where the source hedges.
  • Numbers with no source attached.
  • A real paper cited for a population or outcome it did not study, the subtler failure covered in why a real citation can still be the wrong citation.

For a shorter version to keep beside your desk, see a checklist for medical writers using AI drafting tools.

Where does an automated check fit?

An automated check can narrow the work, and you still sign off on it. Luma's demo takes a medical question, separates the answer into claims, searches PubMed for each one, links the supported claims to their sources and flags what it cannot verify.

Flagged claims tell you where to spend your reading time. You still open the sources, judge the evidence and decide what goes in the document.

Frequently asked questions

Can medical writers use AI to find references?

Use AI to suggest search terms, then find and confirm every reference yourself in PubMed. A reference that came from an AI tool's memory stays unverified until you have opened its record and read it.

Is it acceptable to use AI in medical writing?

That depends on the policies of your employer, your client and the journal or publisher involved. Check those policies before you start, and disclose AI use wherever they require it.

What is the quickest way to check an AI-generated reference?

Look up its PMID on PubMed or resolve its DOI, then confirm that the title, authors, journal and year match. If the record matches, read the abstract to confirm it supports the sentence.

Why keep a verification log?

A verification log shows reviewers and clients exactly how each claim was supported. It also saves you from repeating the search when a sentence is questioned during revision.

References

Check an AI medical answer

Paste an answer or ask a question. Luma splits it into claims, checks each one against published research, links what the evidence supports and flags what it cannot verify.

Try the demo

More from the blog

# How medical writers can use AI without risking fabricated references

Published October 14, 2026

Medical writers can use AI to draft but never as a source: verify every reference in PubMed, keep a verification log, and learn the red flags of fabrication.

Medical writers can use AI safely by keeping it in the drafting role and never treating it as a source. Every reference in the final document should come from a record you looked up yourself, and every claim should be checked against what that record actually says.

Used that way, AI can help with structure and wording without putting invented references under your name.

## Why are AI-generated references a risk for medical writers?

Large language models can produce references that look complete and do not exist. A 2023 study by Walters and Wilder examined 636 citations in 84 literature reviews written by an earlier version and a newer version of a widely used chatbot. Of the earlier version's citations, 55 percent were fabricated, and 18 percent of the newer version's were. The reviews covered 42 topics from many fields, so the study was not specific to medicine.

In biomedicine, a 2026 audit by Topaz and colleagues in The Lancet reports that prior studies estimate 30 to 69 percent of references generated by large language models in biomedical contexts are fabricated. The same audit found 4,046 fabricated references across 2,810 papers, and 98.4 percent of affected papers had received no publisher action at the time of the audit.

For a writer, the lesson is direct: an error that leaves your desk may never be corrected, and your name stays on it.

## What can AI do well in a medical writing workflow?

AI is useful for work where you can check the output against something you already have:

- Outlining a document from your own notes or a brief.
- Rewriting a passage for a different reading level, audience or tone.
- Summarizing a paper you have open, so you can compare the summary line by line with the source.
- Suggesting search terms and synonyms for a literature search you will run yourself.
- Pointing out gaps or unclear steps in an argument.
- Drafting plain-language versions of text you have already verified.

In every case, the facts come from you or from a source in front of you. The AI shapes the words.

## What should AI never do in your workflow?

- Supply references from memory. Any reference you did not find yourself stays unverified until you look it up.
- Decide what a study found. Read the abstract or full text yourself.
- Fill in numbers. Effect sizes, rates, sample sizes and doses must come from the source.
- Strengthen a hedge. If a paper says a finding "may" hold, the draft should not say it does.

## What does a safe workflow look like?

1. **Find the sources first.** Search PubMed, the free search engine for biomedical literature maintained by the National Library of Medicine at the National Institutes of Health, and save each record with its PubMed ID (PMID) and digital object identifier (DOI).
2. Give the AI only those sources, and ask it to draft from them without adding others.
3. Tag every claim in the draft with its source, or with a note that it still needs one.
4. Check every reference. Look up each PMID or DOI, confirm the authors, journal and year match, and read the abstract against the sentence. The full method is in [how to check if a PubMed citation is real](/blog/how-to-check-if-a-pubmed-citation-is-real).
5. Check every number against the source document.
6. Do a final pass claim by claim instead of paragraph by paragraph, the idea behind [claim-level verification](/blog/what-is-claim-level-verification).
7. Log what you checked.

Giving the AI your own sources is not a guarantee. A model handed a paper can still misstate what it says, so the reference and number checks stay in place even when you supplied every source.

## What should a verification log include?

A simple log turns your checking into a record that a reviewer, client or editor can follow. For each claim, note:

- The claim as it appears in the document.
- The source, with its PMID or DOI.
- Where in the source the support appears, such as the abstract, a results table or a named section.
- The outcome: supported, partly supported, not supported, or replaced.
- Who checked it and when.

A log also saves time on revisions. When a reviewer questions a sentence, you can show what supports it without searching again.

## What are the red flags in an AI draft?

Watch for these:

- References you did not supply.
- Titles that restate your claim almost word for word.
- Reference entries with no PMID or DOI.
- An identifier that opens a different paper than the one described.
- Firm language where the source hedges.
- Numbers with no source attached.
- A real paper cited for a population or outcome it did not study, the subtler failure covered in why a real citation can still be the wrong citation.

For a shorter version to keep beside your desk, see a checklist for medical writers using AI drafting tools.

## Where does an automated check fit?

An automated check can narrow the work, and you still sign off on it. Luma's [demo](/demo) takes a medical question, separates the answer into claims, searches PubMed for each one, links the supported claims to their sources and flags what it cannot verify.

Flagged claims tell you where to spend your reading time. You still open the sources, judge the evidence and decide what goes in the document.

## Frequently asked questions

### Can medical writers use AI to find references?

Use AI to suggest search terms, then find and confirm every reference yourself in PubMed. A reference that came from an AI tool's memory stays unverified until you have opened its record and read it.

### Is it acceptable to use AI in medical writing?

That depends on the policies of your employer, your client and the journal or publisher involved. Check those policies before you start, and disclose AI use wherever they require it.

### What is the quickest way to check an AI-generated reference?

Look up its PMID on PubMed or resolve its DOI, then confirm that the title, authors, journal and year match. If the record matches, read the abstract to confirm it supports the sentence.

### Why keep a verification log?

A verification log shows reviewers and clients exactly how each claim was supported. It also saves you from repeating the search when a sentence is questioned during revision.

## References

- Topaz M, Roguin N, Gupta P, Zhang Z, Peltonen LM. The Lancet. 2026;407(10541):1779-1781. doi:10.1016/S0140-6736(26)00603-3. PMID 42107362. https://pubmed.ncbi.nlm.nih.gov/42107362/
- Walters WH, Wilder EI. Scientific Reports. 2023;13(1):14045. doi:10.1038/s41598-023-41032-5. PMID 37679503. https://pubmed.ncbi.nlm.nih.gov/37679503/