Resources / AI answer auditing & correction / Playbook
Classify and triage AI answer issues
Turn a questionable AI answer into a preserved, consistently classified case with consequence-based severity, ownership, and the right response path.
The short answer
Preserve the complete answer before diagnosing it. Then identify one disputed claim, verify the current evidence, assign the narrowest supported issue type, assess severity from likely consequence rather than tone, and route the case to the owner who can address the source or reporting path.
Issue type and severity are separate. A stale price can be high severity when it changes a purchase decision, while an obviously false but immaterial detail may be low severity. “Hallucination” is not a sufficient case classification because it does not identify what failed or what action follows.
Use one answer-capture log per observed run and one case-intake record per investigated issue.
Preserve evidence before triage
Save the exact prompt, relevant earlier turns, complete answer, visible citations, linked destinations, product and mode, date and time, market, language, and available account or personalization conditions. Preserve screenshots when presentation affects meaning, but keep selectable answer text and URLs as well.
Do not replace the original observation when repeating the prompt. A changed answer does not erase the reported one, and a repeated answer does not prove that every user receives it.
Classification decision tree
Questionable answer
├─ Can the exact output and context be reviewed?
│ ├─ No → Insufficient evidence. Request a complete capture; do not diagnose.
│ └─ Yes
├─ Does the concern identify a checkable material claim or source relationship?
│ ├─ No
│ │ ├─ Opinion, preference, criticism, or ranking → Not a factual correction case.
│ │ └─ Meaning remains unclear → Unverifiable; record what evidence is missing.
│ └─ Yes
├─ Are distinct people, organizations, products, or locations combined?
│ └─ Yes → Wrong entity or ambiguous identity.
├─ Does reliable current evidence contradict the claim?
│ ├─ Yes, no established formerly correct state → False; record history as unknown if needed.
│ ├─ Yes, it was formerly correct → Stale.
│ └─ Not directly
├─ Does an omission or qualification materially change the meaning?
│ ├─ Missing necessary fact or condition → Incomplete.
│ └─ Correct details create a materially wrong impression → Misleading.
├─ Is the primary failure in a displayed source relationship?
│ ├─ Destination fails → Broken citation.
│ ├─ Source is about another subject → Irrelevant citation.
│ ├─ Source does not establish the claim → Unsupported citation.
│ ├─ Answer changes the source's meaning → Misquotation or misrepresentation.
│ └─ Named source cannot be found or verified → Apparently fabricated reference.
├─ Is there evidence of deceptive identity or content?
│ └─ Yes → Suspected impersonation or manipulation; specialist escalation.
└─ Evidence remains insufficient or conflicting → Unverifiable or expert dispute.
Assign one primary class that determines the next action. When the answer substitutes or combines the wrong subject, use wrong entity as primary; record resulting factual or citation failures as secondary where useful. Add secondary classes only when they describe distinct failures, such as a stale claim linked to an unsupported citation.
Apply the labels consistently
| Class | Use when | Do not use when |
|---|---|---|
| False | Current reliable evidence directly contradicts a factual claim, and prior accuracy has not been established. | Evidence is merely absent, the matter is opinion, or established prior accuracy makes stale more precise. |
| Stale | A formerly accurate fact no longer reflects the effective date or current state. | Prior accuracy has not been established. |
| Incomplete | A missing condition or fact materially changes the reader’s decision. | The answer omits optional detail without changing meaning. |
| Misleading | Framing, comparison, or implication creates a materially wrong impression. | The team simply dislikes the wording or conclusion. |
| Wrong entity | The answer combines or substitutes distinct entities. | The identity is correct but a fact is stale. |
| Citation failure | A visible citation is broken, irrelevant, unsupported, misrepresented, or apparently fabricated. | There is no visible citation. Absence is a separate observation. |
| Suspected manipulation | Evidence indicates impersonation, counterfeit sources, deceptive redirects, or coordinated false material. | The answer is inaccurate without evidence of deception. |
| Unverifiable | Available evidence cannot resolve the claim or reproduce enough context. | Reliable evidence supports a narrower class. |
| Not a correction case | The concern is opinion, criticism, preference, ranking, or a legitimate unresolved dispute. | A checkable material fact is wrong. |
Avoid classifying an entire answer from one sentence. Quote the smallest passage that contains the issue, then state the corrected fact and evidence.
Determine severity independently
Assess the most credible likely consequence if a relevant person relies on the answer. Consider audience, reach where evidence exists, decision impact, subject sensitivity, persistence, reversibility, and evidence strength. Do not infer broad reach from one observation.
| Severity | Practical threshold | Response target |
|---|---|---|
| Critical | Credible imminent risk to health, safety, legal rights, security, financial assets, privacy, or essential operations. | Preserve evidence and alert the designated incident or specialist owner immediately. |
| High | Material harm to a consequential decision, active fraud or impersonation, significant reputational harm based on false fact, with recurrence affecting priority when it increases credible consequence. | Assign an owner promptly, verify evidence, and begin the appropriate correction or escalation. |
| Medium | Materially wrong information with limited or reversible impact, or an issue affecting a defined audience without urgent harm. | Correct through the normal owned workflow and schedule a recheck. |
| Low | Minor factual or citation defect unlikely to alter a meaningful decision. | Record, batch, and correct when practical. |
Escalate uncertainty itself when the possible harm is severe. A credible but not yet proven safety or fraud report may require urgent specialist review without prematurely labeling the factual cause as confirmed.
Same issue type, different severity
Hypothetical examples. A museum answer repeats a former mailing-box number in a historical description. The current visitor address is correct and no present decision depends on that detail: stale, low severity. An answer repeats an obsolete phone number for an urgent service, directing someone away from needed assistance: also stale, potentially critical if credible imminent harm is involved. In both cases, an archived record establishes the earlier fact; consequence and urgency determine the response.
If current records contradict a claimed founding date but its historical origin is unknown, record false with history unknown. Do not spend the investigation trying to prove that it was never correct before assigning the narrower actionable finding.
Triage workflow
1. Open and secure the case
Create the case ID, link the observation, remove unnecessary personal or confidential information, set access controls, and name the triage owner.
2. Write the disputed claim
Quote the exact passage and state the affected entity, audience, and decision. Split a report into separate cases when different claims require different evidence or owners.
3. Establish the current record
Use current authoritative evidence and record effective dates, versions, markets, and limitations. For an organization-controlled fact, link the entity fact record or a claim-and-source record. If qualified experts disagree, record the disagreement rather than choosing the preferred answer.
4. Classify and assess severity
Apply the decision tree, record the primary and any secondary class, state the evidence behind the label, and set severity separately. Mark both as provisional until the required reviewer approves them.
5. Choose the response path
| Evidence points to | First response |
|---|---|
| Incorrect first-party page | Correct the canonical page and its dependent records. |
| Conflicting structured data or feed | Reconcile the owning data source, visible page, and markup. |
| Entity ambiguity | Clarify identifiers and relationships; use the reconciliation log. |
| Incorrect independent source | Follow the publisher’s evidence-based correction process. |
| Answer-product output with no identified source | Use the product’s feedback or reporting path and retain the case ID. |
| Policy, privacy, intellectual-property, or legal issue | Use the applicable formal channel and specialist owner. |
| Impersonation, fraud, or security issue | Escalate to security, legal, fraud, or platform-trust teams. |
| Opinion or legitimate criticism | Document the decision not to pursue a factual correction. |
Correct the system that owns the fact before polishing a downstream display. A page edit cannot resolve an outdated feed that republishes the old value.
6. Hand off the correction and recheck
Give the owner the preserved observations, approved fact, classification, severity, and next action. Continue with the correction and recheck playbook for controlled changes, outside reporting, comparable rechecks, and closure. Keep submitted, accepted, source corrected, and later answer changed as separate states.
Rapid triage card
Case ID:
Observation ID and evidence location:
Exact disputed passage:
Affected entity:
Affected audience and decision:
Verified current fact:
Evidence and effective date:
Evidence limitations:
Primary class:
Secondary class, if needed:
Classification rationale:
Severity:
Likely consequence:
Urgency or response target:
Probable source or cause:
First response path:
Owner:
Specialist reviewer, if required:
Next action and due date:
Recheck date and conditions:
Quality check
Before acting, confirm that the team has preserved the complete output, isolated a material claim, used current evidence, distinguished issue type from severity, avoided “hallucination” as the only diagnosis, protected sensitive information, assigned an owner, chosen a response proportional to the evidence, and defined a comparable recheck.