# AI Content Review Checklist and Output Log Template Controls so AI-generated content meets your accuracy and liability bar before it leaves the building. A content risk-tiering matrix with a review standard per tier, an output log, and a training scenario library of real hallucination cases. **Who this is for:** The owner of generative AI use in communications, marketing, legal, or regulatory work. Source playbook: https://aigovernance.com/playbook/ai-generated-content-and-hallucinations --- ## AI content risk tiering matrix _Use case to risk level to the review standard required before the content is used._ ### Template | Use case | Audience / exposure | Risk level | Required review standard | Grounding required | |---|---|---|---|---| | | internal / customer / public / regulator | Low / Moderate / High | scan / verify / independent rewrite | none / RAG with cited sources | ### Worked example | Use case | Exposure | Risk level | Review standard | Grounding | |---|---|---|---|---| | Internal meeting summaries | internal | Low | scan for obvious errors | none | | Marketing copy | public | Moderate | verify all claims, dates, figures | RAG for product facts | | Customer support replies | customer | Moderate | verify facts; approved-answer check | RAG over the knowledge base | | Regulatory filing text | regulator | High | independent rewrite by a qualified person; AI draft not used verbatim | RAG with cited primary sources | | Legal contract clauses | counterparty | High | lawyer drafts; AI for first pass only, every clause checked | none | ### Acceptance criteria - Every generative use case is placed in a tier with a named review standard. - High-risk uses do not permit AI output to be used verbatim. - The matrix states where retrieval-grounding with cited sources is mandatory. --- ## Review standard definitions _What "scan", "verify", and "independent rewrite" actually require, so the tier label means something._ ### Template | Standard | What the reviewer does | Time expectation | Sign-off | |---|---|---|---| | Scan | Read for plausibility and obvious errors; fix or reject | minutes | reviewer initials on the log | | Verify | Check every factual claim, date, figure, name, and citation against a source; annotate each | proportional to claim count | reviewer + source list attached | | Independent rewrite | Subject-matter expert writes the final version; AI output is reference only | full authoring time | named author owns the content | ### Worked example > Customer support reply, Moderate tier, "verify" standard. | Claim checked | Source | Result | |---|---|---| | Refund window | policy doc | correct | | Product compatibility | spec sheet | correct | | Named contact | staff directory | correct | | "48-hour escalation guarantee" | not in policy | hallucinated, removed | Source list attached to the log; reviewer initials recorded. ### Acceptance criteria - Each standard has a concrete definition a reviewer can follow. - "Verify" requires a source for each claim, attached to the record. - Reviewers are trained on which standard applies to their use case. --- ## AI-generated content log _A record for AI content used in significant decisions or external communications._ ### Template | Date | Use case | Model / version | Prompt or source summary | Grounding sources | Reviewer | Review standard applied | Changes made | Published / used where | |---|---|---|---|---|---|---|---|---| | | | | | | | scan / verify / rewrite | | | ### Worked example | Date | Use case | Model / version | Prompt summary | Grounding sources | Reviewer | Standard | Changes | Used where | |---|---|---|---|---|---|---|---|---| | 2026-09-03 | Support reply | copilot v1.6 | "explain refund eligibility for order X" | KB articles 12, 44 | u-2201 | verify | removed invented 48h guarantee | ticket 88213 | | 2026-09-04 | Blog post | ext API | "draft 600 words on AI inventory basics" | none | u-1180 | verify | corrected 2 stats, added 3 citations | /blog/ai-inventory-basics | ### Acceptance criteria - Content used externally or in significant decisions is logged with its model version and reviewer. - The log records what the reviewer changed, so hallucination patterns are visible over time. - Grounding sources are listed where the use case required them. --- ## Training scenario library _Real hallucination examples with the correct handling, for staff training._ ### Template > 6-10 short scenarios. Each: the AI output, what is wrong, how a trained person should handle it. **Scenario template:** - **Context:** what the person asked the AI for - **AI output (excerpt):** - **The problem:** the specific fabrication or error - **Correct handling:** what to check, what to change, whether to use it at all - **Rule it illustrates:** ### Worked example **Scenario 3 (citations):** - **Context:** analyst asked the AI to "summarize the case law on AI vendor liability with citations". - **AI output:** cited "Meridian Corp v. DataForge, 2024" with a plausible summary. - **The problem:** the case does not exist. The citation format and reasoning looked correct. - **Correct handling:** never use an AI-provided legal citation without pulling the primary source. Here, no source exists, so the paragraph is deleted and flagged to Legal. - **Rule it illustrates:** AI output is a draft. Every citation, statute, date, and figure is verified against a primary source before use. **Scenario 5 (confident numbers):** - **Context:** marketing asked for "our market share vs competitors". - **AI output:** gave specific percentages with a confident tone. - **The problem:** the model has no access to that data and invented it. - **Correct handling:** treat any specific figure the model was not given as fabricated. Pull the real numbers from the analytics team. - **Rule it illustrates:** the model cannot know facts it was not given; specificity is not evidence. ### Acceptance criteria - Scenarios are drawn from real incidents or realistic near-misses, not invented. - Each scenario ends with the rule it teaches. - The library is used in onboarding and refreshed as new failure patterns appear. --- ## Governance controls this kit produces evidence for - **SAF-001**: The risk matrix, review standards, and log are the hallucination detection and mitigation process. - **SAF-002**: The "verify" and "rewrite" standards plus the log are the AI output validation record. - **MGV-010**: The High-tier review standard is pre-publication verification for high-stakes claims. - **MGV-008**: The content log supports AI-generated deliverable disclosure and citation standards. - **ALC-001**: The content log extends decision and output logging to generative use.