A citation is useful only when it supports the actual claim.
Fluency is not evidence
A generated answer can be clear and plausible while including incorrect claims. The tone of an answer does not establish whether it is supported.
Grounding can help
Providing relevant sources can give the model better information to work with. A groundedness check asks whether the response is supported by that material. It does not automatically establish that the material itself is correct.
Match review to the stakes
A small error in a brainstorming session is different from an error in an important decision. Interfaces should help users inspect evidence, recognize uncertainty, and seek appropriate review.
Check claims one at a time
FActScore evaluates support for atomic factual claims in long-form generation. Its useful lesson for readers is that an answer can mix correct and unsupported details. One accurate date or valid citation does not validate the rest of the paragraph. The study concerns its own tasks and models, not the accuracy of every current product.
Source: Min and colleagues · FActScore: evaluating factual precisionGrounding and truth are separate checks
An answer can correctly repeat an unreliable source, or incorrectly summarize an excellent one. First inspect whether the source is relevant, current and credible. Then check whether it actually supports the answer. A citation may be real while the claim attached to it is unsupported.
Uncertainty should help you act
Our recommendation is to make uncertainty specific: which detail is unresolved, why it matters and what evidence would settle it. A generic warning on every answer is easy to ignore. A precise note that two documents disagree about a deadline tells the user what to verify before relying on it.
A situation to think through
An assistant summarizes a grant announcement with eligibility, funding and a closing date. Open the official announcement and verify each field separately. An accurate funding amount should not reassure you about a deadline copied from last year’s program.
Questions to take with you
- Open sources for consequential claims rather than counting citations.
- Separate missing evidence, source disagreement and known error.
- Verify details that would change a decision before acting on them.
For more reading
The sources behind this page, with a reason to open each one. Practical examples and recommendations are our editorial interpretation.
- FActScore: evaluating factual precision
Evaluates support for individual factual claims rather than treating a long answer as entirely right or wrong.
- Granite Guardian documentation and examples
Versioned examples for risk, groundedness, tool-call checks and custom judging criteria.
- Generative AI Profile · NIST AI 600-1
A framework for identifying, measuring and managing generative AI risks across the system lifecycle.
Sources reviewed 13 September 2026. Product documentation can change. How we use evidence