Your facts decide. Not a confidence score.
HowAITrust finds and prioritises answers that contradict what your company can prove, but never labels a claim false because a model or a heuristic disagrees.
- 1
Observation
Preserve the exact question, AI platform, market, answer text, cited URLs and time. An observation records what was said. It is not yet a verdict.
- 2
Detection
Compare the answer with your current approved facts. A finding is kept only when it quotes a sentence that is in the answer and names the fact it contradicts; anything else is dropped. Detection opens a suspect, never a verdict.
- 3
Priority
Deterministic triage from potential impact, claim type, repetition and missing sources. It answers “what should a person review first?”, not “is this true?”.
- 4
Human verdict
Accurate, inaccurate, outdated and ambiguous each need an approved fact and a public HTTPS source. The reviewer and time are recorded with every revision.
- 5
Correction at the source
Read the pages the AI cited and classify each one: repeats the claim, states both versions, states the fact, or does not bear on it. The correction plan targets the pages that repeat the claim.
- 6
Proof
Ask the same question on the same platform and market after the correction, then after 7 and 21 days. A case resolves only with a reviewed verdict, a correction record and a later answer.
What a resolved case does and does not prove
It proves that on this question, platform and market, a later answer no longer contradicts your approved fact after the recorded correction. It does not prove every AI answer everywhere is now correct: model output varies, so each case keeps the dates and the before and after text.
Evidence hierarchy
Prefer first-party policy, product, pricing, legal and leadership pages. For regulated or independently verifiable claims, keep the regulator, registry or standards source as well.