Resources / AI answer auditing & correction / Template
AI answer audit case-intake template
Turn a reported or measured AI-answer problem into an owned case with preserved evidence, verification, classification, severity, and a defined investigation path.
What this template is for
Use this template when a captured AI answer may contain a material factual, identity, citation, safety, privacy, policy, or legal problem. It turns a report into a case that another person can verify, prioritize, investigate, and route without relying on a cropped image or paraphrased complaint.
Create one case for one disputed issue affecting one entity or closely related fact set. Link duplicate reports and repeated observations to that case, but preserve their different prompts, products, markets, dates, and outputs. Split the case when the facts, affected entities, severity, owners, or correction paths differ materially.
Intake does not prove that the answer is wrong or that a particular source caused it. The initial classification and severity remain provisional until a responsible reviewer checks the evidence.
Minimum intake, then investigation
To open a case, complete the case ID, observation link, exact disputed passage, affected entity and decision, owner, access restriction, provisional urgency, and next action with a due date. If the correct fact is not verified, state the gap and assign verification. A potentially serious report does not need a completed causal investigation before it reaches a responsible person.
Complete the proposed correct position, classification approval, source/cause assessment, action plan, duplicates, and case history as the investigation progresses. Keep uncertain fields marked unknown. The full form below supports that later work without making every field an intake prerequisite.
Filled hypothetical intake
| Required field | Fictional completed value |
|---|---|
| Case ID | HT-4-001: obsolete battery compatibility |
| Observation | OBS-041, linked full answer, prompt, citations, product/mode, market, language, account state, and timestamp |
| Disputed passage | “The current HT-4 uses the B12 battery.” |
| Entity and decision | Harbor Tools HT-4; which replacement battery to purchase |
| Known evidence | Operations specification revision 3 says B14, effective September 1; revision 2 records the former B12 value |
| Provisional classification | Stale, pending reviewer confirmation of the effective product version |
| Provisional severity | High: an incompatible purchase is a consequential error in this scenario; no safety harm is asserted |
| Owner and access | Catalog lead; internal working record, no personal customer data required |
| Next action | Verify the effective specification and inspect the cited page plus supplying catalog field by September 11 |
| Unknown | Whether the visible source caused the answer; broader prevalence |
See the completed correction and recheck example for how this intake develops into actions and later observations.
Copyable case-intake record
Copy the sections below into the approved case-management system. Restrict access when the evidence contains personal, confidential, health, legal, safety, fraud, or security information.
# AI answer audit case: [case ID] — [short issue name]
## Case control
- Case ID: [stable unique identifier]
- Received at: [YYYY-MM-DD hh:mm and time zone]
- Intake channel: [measurement handoff / employee / customer / partner / monitoring / other]
- Reporter or source: [person, team, system, anonymous, or withheld]
- Case owner: [person]
- Triage reviewer: [person]
- Current status: [new / verifying / investigating / correcting / reported externally / monitoring / resolved / closed]
- Initial response target: [YYYY-MM-DD hh:mm and time zone]
- Next review date: [YYYY-MM-DD]
- Access classification: [policy label]
- Related incident, legal, privacy, or security record: [ID or none]
## Issue definition
- Affected entity: [canonical name, entity type, and record ID]
- Disputed claim: [one exact factual statement or concise issue]
- Claim location in answer: [section, sentence, table, image, citation, or other location]
- Proposed correct fact or required treatment: [value, wording, removal, review, or unresolved]
- Affected audience: [who may encounter or act on it]
- Decision or action affected: [purchase, visit, application, safety action, reputation judgment, other]
- Potential consequence: [specific harm, confusion, loss, or other effect]
- Known scope: [products, prompts, markets, languages, dates, or unknown]
- Outside scope: [related issues not handled in this case]
## Preserved answer evidence
| Evidence ID | Observation ID | Product and mode | Prompt ID or exact prompt | Collected at | Market and language | Evidence artifacts | Material difference |
| --- | --- | --- | --- | --- | --- | --- | --- |
| EVD-001 | [answer-capture ID] | [displayed product and mode] | [ID or verbatim prompt] | [timestamp and zone] | [values] | [full text, screenshot, export, citations] | [original / repeated / different condition or output] |
- Evidence completeness: [complete / partial, with missing elements]
- Original answer preserved without edits: [yes / no, with reason]
- Conversation context preserved: [yes / no / not applicable]
- Visible citations and destination URLs preserved: [yes / no / none displayed]
- Sensitive information handling: [redaction, restricted location, access decision, or none]
- Reproduction status: [not attempted / reproduced / different output / product unavailable / error]
- Reproduction observation IDs: [IDs or none]
## Verification of the disputed claim
| Verification ID | Fact or question checked | Evidence or canonical record | Evidence location | Effective date or version | Result | Limits or disagreement | Reviewer and date |
| --- | --- | --- | --- | --- | --- | --- | --- |
| VER-001 | [precise fact] | [source, fact record, claim record, registry, specialist finding, or other] | [URL, field, page, section, or record ID] | [date or version] | [supports correction / supports answer / partial / conflicting / unresolved] | [scope, authority, uncertainty, or dispute] | [person, YYYY-MM-DD] |
## Provisional classification
- Primary issue type: [false / stale / incomplete / misleading / wrong entity / citation failure / suspected manipulation / unverifiable / not a correction case]
- Subtype, if useful: [ambiguous identity / broken citation / irrelevant citation / unsupported citation / misquotation / apparently fabricated reference / suspected impersonation / other / none]
- Secondary issue types: [values or none]
- Classification reason: [how the evidence fits the selected type]
- Fact, opinion, criticism, ranking, or unresolved expert dispute: [classification and reason]
- Evidence confidence: [high / medium / low, with reason]
- Classification reviewer and date: [person, YYYY-MM-DD]
## Severity and urgency
| Factor | Assessment | Evidence or reason |
| --- | --- | --- |
| Likely harm | [critical / high / medium / low / unknown] | [specific consequence] |
| Decision consequence | [critical / high / medium / low / none / unknown] | [decision affected] |
| Potential audience reach | [high / medium / low / unknown] | [available basis; do not invent exposure] |
| Persistence or recurrence | [repeated / observed once / unknown] | [observation IDs or limitation] |
| Subject sensitivity | [health / safety / finance / legal rights / security / fraud / discrimination / personal data / reputation / routine / other] | [context] |
| Evidence strength | [high / medium / low] | [reason] |
- Overall severity: [critical / high / medium / low]
- Urgency reason: [why the response target fits the evidence and consequence]
- Immediate protective action required: [action and owner or none]
- Specialist escalation required: [legal / security / privacy / safety / compliance / communications / executive / other / none]
- Escalation record and time: [reference or not applicable]
- Severity approved by and date: [person, YYYY-MM-DD]
## Initial source and cause assessment
| Candidate source or cause | Basis for considering it | Relationship to claim | Control | Confidence | Next check |
| --- | --- | --- | --- | --- | --- |
| [page, feed, profile, citation, identity ambiguity, delivery issue, manipulated source, or unknown] | [observed evidence] | [identified by product / cited / contains similar wording / inference / other] | [controlled / managed account / outside source / unknown] | [confirmed / likely / possible / unknown] | [specific check] |
- Product disclosed a source or generation path: [what was disclosed or no]
- Alternative explanations retained: [list]
- Probable cause statement: [carefully qualified conclusion or unknown]
## Ownership and initial action plan
| Action ID | Action | Purpose | Owner | Target date | Dependency | Status |
| --- | --- | --- | --- | --- | --- | --- |
| ACT-001 | [verify fact / inspect cited source / audit first-party record / correct controlled source / contact publisher / use product feedback / escalate / recheck / other] | [question or outcome] | [person] | [YYYY-MM-DD] | [record, reviewer, or none] | [planned / active / blocked / complete / declined] |
- Intended response level: [document only / first-party correction / publisher request / product feedback / policy report / specialist escalation / urgent incident response / pending verification]
- Controlled source owner: [person or team or none identified]
- Outside channel under consideration: [publisher or provider process or none]
- Comparable recheck conditions: [product, prompt, market, language, mode, and account conditions]
- Earliest useful recheck date: [YYYY-MM-DD and reason]
## Duplicate reports and linked records
| Record or report ID | Relationship | Material difference | Retained evidence | Reporter response needed |
| --- | --- | --- | --- | --- |
| [ID] | [duplicate report / related case / repeated observation / source conflict / other] | [different prompt, product, market, output, claimant, or none] | [artifact or record] | [action or none] |
## Intake decision
- Intake result: [accepted / needs evidence / merged / split / referred / not a correctable error / closed at intake]
- Decision reason: [evidence-based explanation]
- Blocking evidence or review still needed: [item, owner, and target date or none]
- Next accountable action: [one specific action]
- Decision owner and date: [person, YYYY-MM-DD]
- Stakeholder update: [what can be stated precisely at this stage]
## Case history
| Date and time | Status | Action or finding | Evidence or reference | Owner | Next step |
| --- | --- | --- | --- | --- | --- |
| [timestamp and zone] | [state] | [event] | [record] | [person] | [action and date] | # AI answer audit case: [case ID] — [short issue name] ## Case control - Case ID: [stable unique identifier] - Received at: [YYYY-MM-DD hh:mm and time zone] - Intake channel: [measurement handoff / employee / customer / partner / monitoring / other] - Reporter or source: [person, team, system, anonymous, or withheld] - Case owner: [person] - Triage reviewer: [person] - Current status: [new / verifying / investigating / correcting / reported externally / monitoring / resolved / closed] - Initial response target: [YYYY-MM-DD hh:mm and time zone] - Next review date: [YYYY-MM-DD] - Access classification: [policy label] - Related incident, legal, privacy, or security record: [ID or none] ## Issue definition - Affected entity: [canonical name, entity type, and record ID] - Disputed claim: [one exact factual statement or concise issue] - Claim location in answer: [section, sentence, table, image, citation, or other location] - Proposed correct fact or required treatment: [value, wording, removal, review, or unresolved] - Affected audience: [who may encounter or act on it] - Decision or action affected: [purchase, visit, application, safety action, reputation judgment, other] - Potential consequence: [specific harm, confusion, loss, or other effect] - Known scope: [products, prompts, markets, languages, dates, or unknown] - Outside scope: [related issues not handled in this case] ## Preserved answer evidence | Evidence ID | Observation ID | Product and mode | Prompt ID or exact prompt | Collected at | Market and language | Evidence artifacts | Material difference | | --- | --- | --- | --- | --- | --- | --- | --- | | EVD-001 | [answer-capture ID] | [displayed product and mode] | [ID or verbatim prompt] | [timestamp and zone] | [values] | [full text, screenshot, export, citations] | [original / repeated / different condition or output] | - Evidence completeness: [complete / partial, with missing elements] - Original answer preserved without edits: [yes / no, with reason] - Conversation context preserved: [yes / no / not applicable] - Visible citations and destination URLs preserved: [yes / no / none displayed] - Sensitive information handling: [redaction, restricted location, access decision, or none] - Reproduction status: [not attempted / reproduced / different output / product unavailable / error] - Reproduction observation IDs: [IDs or none] ## Verification of the disputed claim | Verification ID | Fact or question checked | Evidence or canonical record | Evidence location | Effective date or version | Result | Limits or disagreement | Reviewer and date | | --- | --- | --- | --- | --- | --- | --- | --- | | VER-001 | [precise fact] | [source, fact record, claim record, registry, specialist finding, or other] | [URL, field, page, section, or record ID] | [date or version] | [supports correction / supports answer / partial / conflicting / unresolved] | [scope, authority, uncertainty, or dispute] | [person, YYYY-MM-DD] | ## Provisional classification - Primary issue type: [false / stale / incomplete / misleading / wrong entity / citation failure / suspected manipulation / unverifiable / not a correction case] - Subtype, if useful: [ambiguous identity / broken citation / irrelevant citation / unsupported citation / misquotation / apparently fabricated reference / suspected impersonation / other / none] - Secondary issue types: [values or none] - Classification reason: [how the evidence fits the selected type] - Fact, opinion, criticism, ranking, or unresolved expert dispute: [classification and reason] - Evidence confidence: [high / medium / low, with reason] - Classification reviewer and date: [person, YYYY-MM-DD] ## Severity and urgency | Factor | Assessment | Evidence or reason | | --- | --- | --- | | Likely harm | [critical / high / medium / low / unknown] | [specific consequence] | | Decision consequence | [critical / high / medium / low / none / unknown] | [decision affected] | | Potential audience reach | [high / medium / low / unknown] | [available basis; do not invent exposure] | | Persistence or recurrence | [repeated / observed once / unknown] | [observation IDs or limitation] | | Subject sensitivity | [health / safety / finance / legal rights / security / fraud / discrimination / personal data / reputation / routine / other] | [context] | | Evidence strength | [high / medium / low] | [reason] | - Overall severity: [critical / high / medium / low] - Urgency reason: [why the response target fits the evidence and consequence] - Immediate protective action required: [action and owner or none] - Specialist escalation required: [legal / security / privacy / safety / compliance / communications / executive / other / none] - Escalation record and time: [reference or not applicable] - Severity approved by and date: [person, YYYY-MM-DD] ## Initial source and cause assessment | Candidate source or cause | Basis for considering it | Relationship to claim | Control | Confidence | Next check | | --- | --- | --- | --- | --- | --- | | [page, feed, profile, citation, identity ambiguity, delivery issue, manipulated source, or unknown] | [observed evidence] | [identified by product / cited / contains similar wording / inference / other] | [controlled / managed account / outside source / unknown] | [confirmed / likely / possible / unknown] | [specific check] | - Product disclosed a source or generation path: [what was disclosed or no] - Alternative explanations retained: [list] - Probable cause statement: [carefully qualified conclusion or unknown] ## Ownership and initial action plan | Action ID | Action | Purpose | Owner | Target date | Dependency | Status | | --- | --- | --- | --- | --- | --- | --- | | ACT-001 | [verify fact / inspect cited source / audit first-party record / correct controlled source / contact publisher / use product feedback / escalate / recheck / other] | [question or outcome] | [person] | [YYYY-MM-DD] | [record, reviewer, or none] | [planned / active / blocked / complete / declined] | - Intended response level: [document only / first-party correction / publisher request / product feedback / policy report / specialist escalation / urgent incident response / pending verification] - Controlled source owner: [person or team or none identified] - Outside channel under consideration: [publisher or provider process or none] - Comparable recheck conditions: [product, prompt, market, language, mode, and account conditions] - Earliest useful recheck date: [YYYY-MM-DD and reason] ## Duplicate reports and linked records | Record or report ID | Relationship | Material difference | Retained evidence | Reporter response needed | | --- | --- | --- | --- | --- | | [ID] | [duplicate report / related case / repeated observation / source conflict / other] | [different prompt, product, market, output, claimant, or none] | [artifact or record] | [action or none] | ## Intake decision - Intake result: [accepted / needs evidence / merged / split / referred / not a correctable error / closed at intake] - Decision reason: [evidence-based explanation] - Blocking evidence or review still needed: [item, owner, and target date or none] - Next accountable action: [one specific action] - Decision owner and date: [person, YYYY-MM-DD] - Stakeholder update: [what can be stated precisely at this stage] ## Case history | Date and time | Status | Action or finding | Evidence or reference | Owner | Next step | | --- | --- | --- | --- | --- | --- | | [timestamp and zone] | [state] | [event] | [record] | [person] | [action and date] |
AI answer audit case: [case ID] — [short issue name]
Case control
- Case ID: [stable unique identifier]
- Received at: [YYYY-MM-DD hh:mm and time zone]
- Intake channel: [measurement handoff / employee / customer / partner / monitoring / other]
- Reporter or source: [person, team, system, anonymous, or withheld]
- Case owner: [person]
- Triage reviewer: [person]
- Current status: [new / verifying / investigating / correcting / reported externally / monitoring / resolved / closed]
- Initial response target: [YYYY-MM-DD hh:mm and time zone]
- Next review date: [YYYY-MM-DD]
- Access classification: [policy label]
- Related incident, legal, privacy, or security record: [ID or none]
Issue definition
- Affected entity: [canonical name, entity type, and record ID]
- Disputed claim: [one exact factual statement or concise issue]
- Claim location in answer: [section, sentence, table, image, citation, or other location]
- Proposed correct fact or required treatment: [value, wording, removal, review, or unresolved]
- Affected audience: [who may encounter or act on it]
- Decision or action affected: [purchase, visit, application, safety action, reputation judgment, other]
- Potential consequence: [specific harm, confusion, loss, or other effect]
- Known scope: [products, prompts, markets, languages, dates, or unknown]
- Outside scope: [related issues not handled in this case]
Preserved answer evidence
| Evidence ID | Observation ID | Product and mode | Prompt ID or exact prompt | Collected at | Market and language | Evidence artifacts | Material difference |
|---|---|---|---|---|---|---|---|
| EVD-001 | [answer-capture ID] | [displayed product and mode] | [ID or verbatim prompt] | [timestamp and zone] | [values] | [full text, screenshot, export, citations] | [original / repeated / different condition or output] |
- Evidence completeness: [complete / partial, with missing elements]
- Original answer preserved without edits: [yes / no, with reason]
- Conversation context preserved: [yes / no / not applicable]
- Visible citations and destination URLs preserved: [yes / no / none displayed]
- Sensitive information handling: [redaction, restricted location, access decision, or none]
- Reproduction status: [not attempted / reproduced / different output / product unavailable / error]
- Reproduction observation IDs: [IDs or none]
Verification of the disputed claim
| Verification ID | Fact or question checked | Evidence or canonical record | Evidence location | Effective date or version | Result | Limits or disagreement | Reviewer and date |
|---|---|---|---|---|---|---|---|
| VER-001 | [precise fact] | [source, fact record, claim record, registry, specialist finding, or other] | [URL, field, page, section, or record ID] | [date or version] | [supports correction / supports answer / partial / conflicting / unresolved] | [scope, authority, uncertainty, or dispute] | [person, YYYY-MM-DD] |
Provisional classification
- Primary issue type: [false / stale / incomplete / misleading / wrong entity / citation failure / suspected manipulation / unverifiable / not a correction case]
- Subtype, if useful: [ambiguous identity / broken citation / irrelevant citation / unsupported citation / misquotation / apparently fabricated reference / suspected impersonation / other / none]
- Secondary issue types: [values or none]
- Classification reason: [how the evidence fits the selected type]
- Fact, opinion, criticism, ranking, or unresolved expert dispute: [classification and reason]
- Evidence confidence: [high / medium / low, with reason]
- Classification reviewer and date: [person, YYYY-MM-DD]
Severity and urgency
| Factor | Assessment | Evidence or reason |
|---|---|---|
| Likely harm | [critical / high / medium / low / unknown] | [specific consequence] |
| Decision consequence | [critical / high / medium / low / none / unknown] | [decision affected] |
| Potential audience reach | [high / medium / low / unknown] | [available basis; do not invent exposure] |
| Persistence or recurrence | [repeated / observed once / unknown] | [observation IDs or limitation] |
| Subject sensitivity | [health / safety / finance / legal rights / security / fraud / discrimination / personal data / reputation / routine / other] | [context] |
| Evidence strength | [high / medium / low] | [reason] |
- Overall severity: [critical / high / medium / low]
- Urgency reason: [why the response target fits the evidence and consequence]
- Immediate protective action required: [action and owner or none]
- Specialist escalation required: [legal / security / privacy / safety / compliance / communications / executive / other / none]
- Escalation record and time: [reference or not applicable]
- Severity approved by and date: [person, YYYY-MM-DD]
Initial source and cause assessment
| Candidate source or cause | Basis for considering it | Relationship to claim | Control | Confidence | Next check |
|---|---|---|---|---|---|
| [page, feed, profile, citation, identity ambiguity, delivery issue, manipulated source, or unknown] | [observed evidence] | [identified by product / cited / contains similar wording / inference / other] | [controlled / managed account / outside source / unknown] | [confirmed / likely / possible / unknown] | [specific check] |
- Product disclosed a source or generation path: [what was disclosed or no]
- Alternative explanations retained: [list]
- Probable cause statement: [carefully qualified conclusion or unknown]
Ownership and initial action plan
| Action ID | Action | Purpose | Owner | Target date | Dependency | Status |
|---|---|---|---|---|---|---|
| ACT-001 | [verify fact / inspect cited source / audit first-party record / correct controlled source / contact publisher / use product feedback / escalate / recheck / other] | [question or outcome] | [person] | [YYYY-MM-DD] | [record, reviewer, or none] | [planned / active / blocked / complete / declined] |
- Intended response level: [document only / first-party correction / publisher request / product feedback / policy report / specialist escalation / urgent incident response / pending verification]
- Controlled source owner: [person or team or none identified]
- Outside channel under consideration: [publisher or provider process or none]
- Comparable recheck conditions: [product, prompt, market, language, mode, and account conditions]
- Earliest useful recheck date: [YYYY-MM-DD and reason]
Duplicate reports and linked records
| Record or report ID | Relationship | Material difference | Retained evidence | Reporter response needed |
|---|---|---|---|---|
| [ID] | [duplicate report / related case / repeated observation / source conflict / other] | [different prompt, product, market, output, claimant, or none] | [artifact or record] | [action or none] |
Intake decision
- Intake result: [accepted / needs evidence / merged / split / referred / not a correctable error / closed at intake]
- Decision reason: [evidence-based explanation]
- Blocking evidence or review still needed: [item, owner, and target date or none]
- Next accountable action: [one specific action]
- Decision owner and date: [person, YYYY-MM-DD]
- Stakeholder update: [what can be stated precisely at this stage]
Case history
| Date and time | Status | Action or finding | Evidence or reference | Owner | Next step |
|---|---|---|---|---|---|
| [timestamp and zone] | [state] | [event] | [record] | [person] | [action and date] |
Accept a case only when it can be acted on
The intake reviewer should be able to answer these questions or assign the missing work:
- Which exact answer, entity, and claim are under review?
- Is the original evidence complete enough to understand the claim in context?
- What fact or treatment must be verified, and who is qualified to approve it?
- What consequence could the issue create for the affected audience?
- Does the case require restricted handling or immediate specialist escalation?
- Who owns the next action, and when will the case be reviewed again?
If evidence is incomplete, retain the report and request the missing prompt, full answer, product, date, and available citations. Do not fill gaps with assumptions or reject a potentially serious report solely because the reporter used the wrong terminology.
Set severity from consequence and evidence
Use the canonical severity thresholds, considering credible consequence, urgency, and evidence. Record the reason, not only the label. Follow a stricter applicable incident policy, and escalate a credible serious risk while verification continues. One answer capture does not establish audience reach.
Keep suspected causes separate from observed evidence
A visible citation can be inspected directly. An uncited page with similar wording is only a possible influence unless the product identifies it. A corrected source and a changed later answer are separate outcomes; neither proves the complete generation path.
Record source and cause statements as confirmed, likely, possible, or unknown, with the basis for that confidence. Keep alternative explanations open until the evidence rules them out.
Choose the next action without promising the outcome
Correct controlled pages, structured data, feeds, profiles, and records at their source. For outside content, use the publisher’s documented correction process. Use answer-product feedback for ordinary output problems and the appropriate policy, safety, privacy, intellectual-property, or legal channel for issues that meet that channel’s scope.
Interfaces and reporting categories change. Check the provider’s current instructions before submitting a report. Record the submission, evidence supplied, case number, response, and follow-up date. An acknowledged submission is not an accepted correction, and a corrected source does not guarantee that a future answer will change.
Validate the intake record
- Open every linked answer-capture record and compare it with the stored artifacts.
- Confirm that the disputed wording and affected entity are precise.
- Check the proposed correct fact against appropriate current evidence.
- Review the classification, severity factors, access restrictions, and escalation decision.
- Merge true duplicates without discarding materially different observations.
- Assign the next action, owner, target date, and comparable recheck conditions.
- Record the intake decision without describing an unverified report as a confirmed error.
Continue the audit
Return to the Auditing & Correcting guide for the case model and response principles, or use the AI answer auditing and correction checklist to trace sources, correct controlled information, choose outside reporting paths, recheck outputs, and close the case with evidence. Use the Entity Reconciliation Log when the issue exposes a conflict between an approved entity fact and a maintained or outside source.
Provider references
- Google Search Help: AI Overviews feedback
- OpenAI Help Center: Reporting content in ChatGPT and OpenAI platforms
- Microsoft Support: Privacy FAQ for Microsoft Copilot
Completion check
- Link to the original answer evidence and preserve every materially different observation separately.
- State the disputed claim, affected entity, corrected position, and decision consequence precisely.
- Separate verified facts from suspected sources, causes, and generation paths.
- Assign an owner, status, next action, target date, and access classification before accepting the case.