Not every false claim an AI model makes about your organisation was published somewhere by a human. Some are invented — fabricated by the model's own statistical machinery from patterns across related content. These hallucinations are the hardest errors to correct, because there is no source to fix. Deep Narrative Intelligence was built to handle both species of error. Here is how it tells them apart — and why the distinction changes everything about correction.
Two species of error
When a model says something false about you, the error falls into one of two categories:
- Source-based errors — the model retrieved or learned a claim that exists on the web. A blog post, a news article, a forum thread contains the false information, and the model repeated it. These are traceable: you can find the source, correct it, and the model's next retrieval or training cycle will encounter the correction.
- Model-generated errors — the model fabricated the claim through statistical pattern-matching. No human published it. No web page contains it. The model constructed it by inferring from related content — companies in your sector had a scandal, so it "stands to reason" that you did too. These are hallucinations, and they are categorically different to correct.
DNI's first job is to distinguish between the two, because the correction strategy is entirely different. Correcting a source-based error means fixing the source. Correcting a hallucination means breaking the pattern the model used to generate it — which requires publishing authoritative content that contradicts the fabricated claim, so future training cycles ingest the contradiction instead of the inference.
Entity confusion: when the model blends you with someone else
The most common model-generated error is entity confusion — the model conflates your organisation with another that shares a name, a sector, or a geographical region. A "SetRecord Ltd" in London gets blended with a "Set Record" music label in New York. A bank in Paris gets confused with a different bank with a similar name in Lyon. The model doesn't distinguish; it synthesises a single answer from fragments of both.
Entity confusion is dangerous because it is entirely model-generated. No source says "the Paris bank and the Lyon bank are the same entity" — the model inferred it. And the error is sticky: every time the model sees both entities in related contexts, it reinforces the confusion, making the blend deeper with each training cycle.
Don't know what AI says about you today?
A SetRecord audit queries supported AI models about your organisation and measures the gap with verifiable facts — within five business days.
Phantom controversies and invented figures
The second species of hallucination is the phantom controversy: the model generates a claim that your organisation was involved in a scandal, a lawsuit, or a regulatory action — not because any source says so, but because the statistical pattern across your sector makes it plausible. "Companies in the energy sector were fined for emissions in 2023" becomes "the company was fined for emissions in 2023" — even if your specific company was not fined and no source says it was.
Invented figures follow the same logic. The model has seen revenue numbers for companies in your sector and your size range, and it generates a plausible number for you — one that appears nowhere in your financials. A journalist or investor asking an AI about your revenue receives that fabricated figure and may cite it as fact.
Why hallucinations persist longer
Source-based errors can be corrected in days to weeks (web layer) or in one training cycle (training layer). Hallucinations are harder, because there is no single source to correct — the error is distributed across the model's weights, not stored in a retrievable document.
The correction strategy for hallucinations is therefore different: instead of correcting a source, DNI publishes new authoritative content that directly contradicts the hallucinated claim. If the model hallucinates a fine, we publish a verified statement of clean compliance. If the model invents a revenue figure, we publish the audited figure in a structured, machine-readable format. The goal is to ensure that the next training cycle ingests the correct fact alongside the pattern that generated the hallucination — so the model learns that the pattern does not apply to your specific organisation.
The DNI approach to hallucinations
Here is how DNI handles model-generated errors specifically:
- Detection — the claim is flagged as a likely hallucination when reverse-searching the phrasing produces no matching web source. If no human published it, the model generated it.
- Classification — is it entity confusion, a phantom controversy, or an invented figure? Each type requires a different correction content.
- Correction content — publish authoritative, structured, dated content that directly contradicts the hallucination. For entity confusion, this means content that clearly distinguishes your organisation from the namesake. For phantom controversies, a verified statement of clean record.
- Patience — hallucinations may take two training cycles to correct, not one, because the model needs to unlearn the pattern, not just ingest a new source. DNI's six-month programme is calibrated to this reality.
This is why a method that only traces provenance — and stops there — cannot fix hallucinations. Provenance tracing finds source-based errors. But when the error has no source, provenance tracing returns empty, and the method stalls. Deep Narrative Intelligence goes further: it identifies the empty result as a hallucination, classifies the type, and deploys a correction strategy designed for model-generated errors, not source-based ones. That difference is the difference between monitoring and actually setting the record straight.