Do-it-yourself game plans · No. 13 / 22
How should you handle incorrect AI answers about your company?
Last updated
Maintain an accuracy register and fix important errors without waiting for the experiment to finish. Do not preserve a harmful factual error merely to protect a comparison group.
Recommended register fields:
| Field | What to record |
|---|---|
| Claim | Exact checkable statement, with answer date and surface |
| Reference evidence | Approved source, version, date and evidence owner |
| Classification | Supported, contradicted, outdated or not verifiable |
| Recurrence | Which observations contained the claim |
| Severity | Proposed high, medium or low priority based on buying impact |
| Likely correction route | Owned page, identifiable third-party page or platform feedback |
| Owner and action | Responsible person, requested change and due date |
| Closure evidence | Published correction and subsequent rechecks—not just a submitted request |
Escalate sensitive security, legal, health or financial claims to an appropriate specialist. Distinguish “we cannot verify this” from “this is false”.
A source correction and an assistant-answer correction are different milestones. Record both rather than promising that one will automatically cause the other.