Restore · 05
AI Answer Correction
A model can state something false about you with complete composure, in fluent prose, with no indication that it is guessing. There is no complaints department. The correction has to happen upstream, in the sources the answer was assembled from.
This is the newest problem in the practice and the one clients find most disorienting. A search result is a document with an author and a publisher, meaning someone to write to. An AI answer is a synthesis. It has no author, no publication date, no correction policy, and no address. It simply asserts, and it asserts in the register of settled fact.
The errors take recognizable forms. Conflation: your record merged with a namesake's, sometimes across countries. Staleness: a matter that resolved years ago presented in the present tense. Invention: a plausible detail with no source at all, generated because the pattern fitted. And omission, which is often the most damaging: a substantial career summarized in one sentence about its single worst week.
None of these can be corrected by contacting the model. What can be corrected is the material the model draws on: the pages, databases, profiles, and structured records that constitute the available evidence about you. Where those are wrong, thin, or contradictory, the model fills the gap by inference. Where they are accurate, consistent, and well-structured, it has considerably less room to invent.
The verification discipline matters as much as the correction. Answers vary by model, by version, by phrasing, and by day. A single satisfying response proves nothing. We test across systems with multiple phrasings and document the distribution rather than the best case.
When it applies
The situations this is the right tool for.
A model states something factually wrong
A role you never held, a company you never ran, an outcome that never occurred, delivered with complete confidence.
Your record is merged with a namesake's
The most common failure mode. Two people with one name, one of whom has a problem, and the model has not distinguished them.
A resolved matter appears unresolved
The allegation is in the sources; the dismissal or settlement is not, or is far less visible.
The answer is thin where it should be substantial
Decades of work reduced to one line, because the accessible sources are sparse and the model works with what exists.
Different systems give contradictory answers
A reliable signal that the underlying source material is inconsistent, which is a fixable condition.
A model declines to answer at all
Read by a counterparty as evasion. Usually it indicates insufficient authoritative material rather than anything adverse.
Process
How the work runs.
Multi-system audit
The same questions across major assistants and AI search surfaces, with several phrasings each, captured verbatim and dated. Answers vary; the distribution is the finding.
Source tracing
Where each claim plausibly originates. Cited sources are the starting point, not the whole answer, since much of what a model asserts is uncited.
Upstream correction
Fixing the sources themselves: factual corrections, structured data, disambiguation between you and a namesake, and adding the authoritative material that was missing.
Entity disambiguation
Making it structurally unambiguous which person or organization you are, which is the single most effective intervention against conflation.
Re-audit on a schedule
Because systems change under you. A correction verified once is a correction verified once.
Before you engage
Where the limits are.
No one controls what a model outputs. Not us, and not the companies that build them. Anyone claiming a guaranteed AI outcome is describing something that does not exist.
Training data is also periodic. A model trained before a correction may continue to reflect the older picture until it is retrained, and that schedule is not ours. Retrieval-based systems update faster; foundation knowledge lags.
We do not attempt prompt injection, poisoned inputs, or manufactured source material. They are detectable, they degrade the information environment, and a client caught doing it has a far worse problem than the one they started with.
Questions we are asked
- Can you make ChatGPT stop saying something about me?
- Not directly, and no one can. What we can do is correct and strengthen the sources it draws on, which measurably changes what it says over time. We report what actually changed rather than what we hoped would.
- How long until an AI answer changes?
- Retrieval-based systems that search live can reflect corrected sources within weeks. Answers drawn from training data may not change until the model is retrained, on a schedule the developer controls and does not publish.
- Which systems do you check?
- The major assistants and AI search surfaces, with multiple phrasings per question, because a model that answers well to one wording frequently answers badly to another. The audit records the range, not the best result.
- Is there any way to make an AI system forget?
- Not reliably, and claims to the contrary are ahead of the technology. Some jurisdictions are developing erasure obligations for AI systems, and the mechanisms are immature. We track this closely and will tell you plainly what is available today.
- What does this cost?
- The fee is stated in a written proposal before any work begins, scoped to what the assessment finds. How that is arrived at is set out in full on the How We Work page.
Related
Mechanisms that often run alongside this one.
Is this your situation?
One conversation, in confidence, with an honest reading of whether this mechanism is the right one, including when it is not.