How AI Name Answers Can Misstate Your Past

A short model answer can omit context or merge same-name people. The safest response is identity resolution and source verification, not more allegation-heavy prompting.

What Is Actually Known

Some general-purpose tools can summarize public information. Search-backed tools may retrieve current indexed pages; closed-book models do not necessarily search the live web. Informal use by employers, landlords, or others is difficult to measure, so it should not be advertised as a standard or universal screening practice.

Where Errors Enter

• Old pages omit a dismissal, acquittal, sealing, expungement, or correction.
• Mugshot mirrors and broker profiles copy stale data.
• Generic teaser text is mistaken for a factual record.
• Two people with the same name are blended.
• A model invents a source or unsupported connection.

Do Not Prime the Answer

Older reputation-audit advice often recommended multiple questions about criminal history, lawsuits, trustworthiness, and risk. That is a poor measurement design: it introduces the allegation into the prompt and can pollute repeated observations.

If an optional sample is justified, use one neutral identity prompt with full name and known location. Ask whether the model can confidently distinguish the exact person. Treat ambiguity as insufficient identity. Keep search-backed sampling separately authorized and labeled.

Routine Monitoring Should Be Passive

Track fixed Google/search queries, image results, and verified source pages. An index recurrence is a reason for source review, not proof that a removal failed. An index absence is not source-site clearance.

FixMyRecord's optional LLM sampler is currently production-disabled. If later released, sampling will remain sparse, explicitly consented, privacy-gated, and recorded with the exact model, prompt version, timestamp, and mode. It will carry zero score and cannot authorize outreach.

Realistic Next Steps

• Verify the person and original source before labeling harm.
• Use real broker opt-out, correction, publisher-policy, search-removal, or legal routes.
• For mugshots or court pages, gather official disposition or expungement documents when available.
• Build accurate positive profiles and original content, then measure fixed search queries without guaranteeing suppression.

What Good Evidence Looks Like

Good evidence preserves the query, returned model, date, source URL, identity basis, caveats, and what was independently verified. It distinguishes observation from causation and never claims private model-memory deletion.