Check an AI claim against the evidence.

Claim:
Required evidence:
Observed record (source, exact revision or date, passage):
Finding: Supported / Contradicted / Insufficient evidence — Intent: not assessed

Common questions about checking AI answers

How do I check whether an AI answer is supported?

Write down the exact claim before interpreting it, define what evidence would need to be present for it to hold, and compare that requirement with the record you can actually inspect. If every required observation supports the claim, the result is Supported. If any required observation conflicts, it is Contradicted. If required evidence is missing or unresolved, it is Insufficient evidence. You can practice that comparison with all six synthetic cases.

How do I verify an AI citation?

Open the cited source and the exact revision or version when one is available. Find the passage that is supposed to support the answer, then compare its subject, value, scope and date with the exact claim. A link is not proof by itself: the unsupported-citation practice case shows a 30-day answer beside a passage that says 7 days.

What is the difference between contradicted and insufficient evidence?

Contradicted means at least one required observation directly conflicts with the claim. Insufficient evidence means the record is missing, unknown, stale or inapplicable and contains no direct conflict. The first supports a negative relationship finding; the second means the available record does not let you decide.

Does an unsupported claim prove AI deception or intent?

No. Terms such as AI hallucination or fabricated citation are often used for different kinds of failure, but they do not establish why an answer was produced. This method asks a narrower, inspectable question: how does the claim relate to the available evidence? Intent is always not assessed.

Three findings

Supported

Every declared required evidence item supports the claim.

Contradicted

At least one required observation directly conflicts with the claim. Contradiction takes precedence over missing evidence.

Insufficient evidence

A required observation is absent, unknown, stale or inapplicable, and no required observation directly contradicts the claim.

The deterministic rule

if any required observation is contradictory:
  contradicted
else if any required observation is absent, unknown, stale or inapplicable:
  insufficient-evidence
else:
  supported

The browser and node tools/check-cases.mjs --self-test import the same pure classifier.

Quality method

1. Name one bounded system claim.
2. Declare the evidence that claim would require before looking at the outcome.
3. Record each observation and its state.
4. Run the deterministic classifier.
5. Publish exact source revisions, reviewed-through dates, limitations and non-claims.
6. Invite a public-safe reproduction or counterexample.

Failure classes

• false-completion: claimed work is absent from final state.
• material-omission: a decision-relevant absence is hidden from the account.
• provenance-or-identity-mismatch: evidence belongs to a different subject.
• evaluation-gap: the intended evaluation set is not accounted for.
• ambiguous-external-effect: a consequential result is unknown after a lost response.
• context-or-citation-escape: the answer outruns its supplied evidence.

Limitations

The cases are synthetic teaching records. They do not estimate prevalence, authenticate reviewers, test live models, diagnose internal cognition, certify products or rank systems. Real investigations need threat-aware handling of sensitive evidence and independent review.

Glossary

Claim

A bounded statement that can be compared with declared evidence.

Required evidence

What would need to be observed for the claim to be supported.

Observation

The recorded state of one required evidence item.

Reviewed through

The date through which a source revision was checked for this case.

Non-claim

An explicit boundary on what the evidence does not establish.