AI disinformation

Brand hallucinations: anatomy of an AI factual error

July 24, 2026 7 min read
Kamel Malek Director of SetRecord.ai

Ask an AI assistant enough questions about any organisation and, sooner or later, it will state something as fact that has no basis in reality — a product line that was never launched, an executive who never held the title, a lawsuit that was never filed. These are not glitches in the traditional sense. They are a predictable output of how generative models fill gaps in what they actually know.

What counts as a brand hallucination

A hallucination, in the technical sense used across the field, is a statement generated with the same fluency and confidence as a verified fact, but which is not supported by any retrievable, checkable source. For a brand, this typically shows up as a specific, checkable detail — a date, a name, a number, an event — that sounds exactly like the kind of thing the organisation would say about itself, but that the organisation never said.

That precision is what makes brand hallucinations distinct from vague AI errors. A model rarely hedges with "possibly" when it hallucinates a detail; it states it the way it states a verified headquarters address, because internally, both are just the most probable next tokens given the prompt.

Hallucinated details are also often inconsistent across sessions in a way verified facts are not: ask the same question five times and a fabricated employee count might shift each time, while a genuinely indexed figure stays fixed. That inconsistency is diagnostic, but it requires deliberately re-querying to observe — a single answer, taken at face value, gives no signal that anything is wrong.

Why models fabricate rather than say "I don't know"

Generative models are trained to produce a plausible, complete answer to nearly any question they are asked. When retrieval turns up thin or contradictory material about a narrow topic — a mid-sized company, a regional executive, a discontinued product — the model still has to generate something. It falls back on statistical pattern-completion: filling the gap with what a similar sentence, about a similar entity, would typically contain. The result reads as specific and confident precisely because the underlying mechanism does not distinguish between "recalled" and "generated to fit the pattern." The same dynamic explains why hallucination rates spike for very recent events: a model's retrieval layer has had less time to accumulate corroborating material, so pattern-completion has more room to operate.

This is also why the same question, asked twice, can produce two different fabricated details. The model is not retrieving a stored falsehood — it is regenerating a plausible one each time, unless a retrieval hit locks the answer to a specific, verifiable source.

Four patterns we see most often

The specific content varies by industry and geography, but structurally, the errors we log fall into four recurring categories:

  • Phantom products — a plausible-sounding product or service extrapolated from the organisation's known category, described with invented specifications.
  • Invented personnel — an executive title attached to a real name from an old filing, or a fabricated name attached to a real title, blending two half-remembered facts.
  • Fictional litigation — a lawsuit or regulatory action generated by pattern-matching against how similar companies in the same sector have been covered.
  • Merged entities — facts, numbers or events belonging to a similarly named company blended into the profile of the one actually being asked about.

These categories are not mutually exclusive — a single answer will often combine an invented executive with a phantom product they are said to have launched, because the model is completing one coherent-sounding narrative rather than fact-checking each element independently.

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.

Request an audit

The thin-retrieval problem

Hallucination rates are not evenly distributed. They rise sharply for organisations with a small, inconsistent, or outdated public footprint — exactly the mid-market companies, regional institutions and recently rebranded entities least likely to be closely tracked by major outlets. A well-documented multinational gives a model enough verified material to draw from across most questions; a smaller organisation gives it just enough fragments to pattern-match convincingly, and just few enough to leave large gaps the model has to fill in on its own. Paradoxically, the organisations with the least media attention are often the most exposed.

Our audits consistently find this threshold effect in practice: organisations with only a handful of substantive, independent mentions across the sources a model is likely to retrieve show measurably higher hallucination rates on specific-fact questions than better-documented peers, regardless of how large or well-run the organisation actually is.

Anatomy of a typical fabrication

A representative pattern: a company that spun off a division several years ago gets asked about "its current subsidiaries." Because older filings mentioning the now-divested subsidiary are still indexed, and few newer sources correct the record explicitly, the model reports the subsidiary as current — sometimes adding invented detail about its size or activities to make the answer feel complete. Nobody planted this error; it emerged from the gap between what changed and what is still indexed.

A second, equally common pattern involves executive titles: a regional manager promoted internally without a press release gets described by a model as still holding their previous title, simply because no indexed source has caught up with the change — the model is not wrong about the person, only about the timeline.

What a hallucination actually costs

The damage is rarely a single embarrassing answer. A fabricated executive name can surface in due-diligence research ahead of a deal. An invented product can confuse customers comparing your catalogue against what an assistant describes. A phantom lawsuit, repeated across enough sessions, can start appearing in genuinely independent commentary that treats the AI-generated claim as a starting point worth investigating — closing the loop between hallucination and reputational harm.

None of this requires bad faith on any side. A journalist, an analyst or a curious customer who repeats what an assistant told them is not lying, only relaying an answer that read as authoritative — which is precisely what makes the cost hard to contain once it starts compounding.

Correcting a hallucination

Because a hallucination has no single source to retract, the fix is not a correction request sent to one publisher. It starts with documenting the exact prompts and models that produce the error, then closing the retrieval gap that made the fabrication possible — usually by publishing or updating an authoritative, specific, dated source covering exactly the fact that was missing. Documenting which specific prompt phrasing triggers the error matters too, since a model can answer one version of a question accurately and hallucinate on a simple rephrasing of the same question. Our audits typically find that hallucination rates on a given topic drop measurably once an unambiguous, well-structured source exists where none did before; the model stops needing to fabricate because it has something concrete to retrieve.

As with any correction, persistence matters more than the first fix — see our methodology for how we track whether an error keeps recurring after the record has been set straight.

See what AI says about you

A SetRecord audit checks what supported AI models say against your verifiable facts — within five business days.