Unverifiable claim: what it means in AI accuracy scoring

An unverifiable claim is an assertion that the checker has no evidence for in either direction. The reference material neither supports it nor contradicts it.

Three outcomes are possible when a tool checks a claim against a set of known facts:

• Supported: a source confirms it.
• Contradicted: a source says something different. This is a false claim.
• Unverifiable: no source covers it.

The term "unsupported claim" is often used for both of the last two. That blurs the line this entry is about.

Accuracy tools usually break an answer into small checkable statements, often called atomic facts, and test each one against a reference set.

The benchmarks differ on what to do with a claim nobody can check:

• FActScore labels a claim "Not-supported" when the source "contradicts or lacks evidence for" it. Both cases land in the same bucket and count the same.
• RAGAS faithfulness divides supported claims by total claims, so an unsupported claim lowers the score whether or not it is false.
• SAFE sets some claims aside as irrelevant, but that covers claims off the question's topic. It does not address a specific fact the model asserts and nobody can check.
• Vectara's hallucination leaderboard excludes refused responses from its denominator, which is about a model declining to answer, not about unchecked assertions.

So no benchmark gives a ready answer for "the AI stated a number and we have nothing to compare it with." Separating unverifiable from false is a judgment call each tool makes for itself.

The choice changes your score and your to-do list.

If a tool counts every unverifiable claim as an error, a brand with few documented facts looks less accurate than one with many. The tool is measuring its own fact coverage as much as the engine's mistakes. If a tool ignores unverifiable claims, a fabricated headcount or founding year can slip through unflagged.

The next move depends on which kind of flag you have:

• Contradicted: find where the wrong fact comes from and correct it, on your site or on third-party pages. Our guide to stopping AI from making up brand information covers what is in your control.
• Unverifiable: the fact is probably missing. State it plainly on a page AI crawlers can fetch, in the HTML before JavaScript runs. A specific fact on a reachable page gives an engine a sentence it can quote, though nothing guarantees it will.

When you compare tools that alert you to false AI claims, ask how each one scores claims it cannot check.

"Unverifiable means false"

It means unknown. The claim could be accurate. Treating it as false punishes the answer for a gap in the checker's data. Read more in whether every flagged AI claim is actually false.

"Benchmarks have settled how to score it"

They haven't. FActScore and RAGAS merge it with contradicted claims. The benchmarks that exclude claims deal with different situations.

"A citation makes a claim verified"

A linked source may not say what the answer claims. See why AI citations sometimes don't support the claim.

• Factual accuracy: how often an AI answer's checkable claims match the facts.
• Atomic fact: the smallest standalone statement a checker can test.
• Contradicted claim: a claim a source shows to be wrong.
• Hallucination: a fabricated statement. Some tools use the word for unverifiable claims too, which overstates what they know.
• Cited vs retrieved sources: why a citation is a weak proof of support.