# What is claim-level verification?

Published October 12, 2026

Claim-level verification splits an answer into its factual claims and checks each against evidence, so one wrong claim cannot hide in a mostly right answer.

Claim-level verification means breaking an answer into its individual factual claims and checking each one against evidence on its own. Instead of a single verdict on the whole answer, every statement gets its own verdict, so one wrong claim cannot hide inside an answer that is otherwise correct.

The approach matters most where small errors carry real cost, as in medicine. A paragraph that is right about a drug's use and wrong about one of its effects is still wrong in the place a reader might act on.

## What is a claim?

A claim is a single statement that can be true or false and can be checked against a source. Most sentences in an answer contain one or more.

Take a sentence saying that angiotensin-converting enzyme inhibitors (ACE inhibitors) are used to treat high blood pressure and can cause a dry cough. That sentence holds two claims:

- ACE inhibitors are used to treat high blood pressure.
- ACE inhibitors can cause a dry cough.

Each one can be checked separately, and each may need a different source.

Good splitting follows a few rules:

- One fact per claim. Compound sentences get split.
- Keep the qualifiers. Words such as "in adults", "may" and "in some patients" belong to the claim, because dropping them changes what has to be proved.
- Make each claim stand alone. Replace "it" or "this drug" with the actual name, so the claim makes sense without the sentence around it.

## Why check claims one at a time instead of the whole answer?

Judging a whole answer is like grading an essay: the reader forms an overall impression. If most of it is accurate and well written, it passes. Errors can sit in the specifics, such as a number, a population, a dose or a citation, and specifics are exactly what an overall impression skims over.

**A mostly right answer can be more dangerous than an obviously wrong one**, because its accuracy elsewhere earns the reader's trust. Fluent prose makes this worse, since a wrong statement reads exactly like a right one, as we explain in why fluent writing makes AI medical errors harder to spot.

Claim-level checking also handles partial answers. An answer might contain several claims that a source supports and one that nothing supports. A whole-answer verdict has to call that answer right or wrong. A claim-level check can say exactly which part is supported and which is not.

## How does claim-level verification work?

The process has the same basic shape in any system:

1. Split the answer into individual factual claims.
2. For each claim, search the literature for evidence, for example on PubMed, the free search engine for biomedical literature maintained by the National Library of Medicine.
3. Read what the search returns and judge whether it actually supports the claim.
4. Attach the source to each supported claim, and flag each claim that has no support.
5. Record a confidence level for each judgment, so a reviewer knows where to look first.

For a working example, see how Luma checks an AI medical answer, step by step.

## What counts as support for a claim?

Finding a paper on the same topic is not enough. A source supports a claim when it studied the same population, the same intervention or exposure, and the same outcome, and when its finding points the same way with at least the same firmness. A study that says a link "may exist" does not support a sentence saying one thing causes another.

The individual checks are laid out in how to tell whether a study actually supports a claim.

## Why does this matter for AI citations?

Citations in AI-generated text can fail in two ways: the source does not exist, or it exists and is cited wrongly. A 2023 study by Walters and Wilder examined 636 citations across 84 literature reviews generated by an earlier version and a newer version of a widely used chatbot, on 42 topics from many fields rather than medicine alone. Fabricated citations made up 55 percent of the earlier version's citations and 18 percent of the newer version's. Among the real citations, 43 percent of the earlier version's and 24 percent of the newer version's contained substantive errors.

A reviewer who only checks whether an answer has sources would pass every one of those. Checking each claim against its source catches both the invented reference and the real one cited for the wrong thing.

## What should happen to a claim that cannot be verified?

It should be flagged where the reader can see it. The tempting alternative, attaching the nearest related paper so every sentence looks sourced, is how misattributed citations get made. An honest flag shows the reader exactly where the evidence runs out. The design case for this is in why medical AI should be allowed to say "not supported".

An unverified claim is not necessarily false. It may be true but absent from the literature searched, phrased in a way the search missed, or a matter of clinical judgment that published research does not settle. The flag means "check this", and a person decides what happens next.

## What are the limits of claim-level verification?

- Splitting can strip context. A claim pulled out of a carefully hedged paragraph can look firmer than the author meant, so qualifiers have to travel with it.
- Not every sentence is a factual claim. Advice, definitions and judgments need a different kind of review.
- Evidence varies in strength. A claim supported by one small study stands on weaker ground than a claim supported by several large trials, so a supported verdict is a starting point for judgment.
- Search can miss things. A claim with no evidence found may still have evidence somewhere.

To see the process on a real question, Luma's [demo](/demo) separates an answer into claims, links the supported ones to their sources and flags what it cannot verify.

## Frequently asked questions

### How is claim-level verification different from ordinary fact-checking?

Claim-level verification is fact-checking done one statement at a time. Instead of rating an answer as a whole, it gives each factual claim its own check against evidence, so a single unsupported claim stays visible.

### Can an answer pass claim-level verification if one claim is unsupported?

The supported claims keep their sources, and the unsupported claim keeps its flag. The answer should not be presented as fully supported while any claim remains flagged.

### Does an unverified claim mean the claim is false?

No. It means the retrieved evidence did not support it. The claim may be true but missing from the literature searched, or it may need a person's judgment.

### Who uses claim-level verification?

Medical writers can use it to check drafts, researchers to check summaries, and teams building health AI products to show which statements in an answer have evidence behind them.

## References

- 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/