Direct answer
An AI can assert inaccurate things about your brand — a false attribute, a mix-up with a competitor, outdated information — and do it with the same confidence as when it's right. It's structural: the AI generates what's probable, not what's true. Unlike an online review, you can't directly "reply" to an AI's assertion. The first useful step isn't to correct in a panic, it's to measure: how often the error appears, in which modes, across which models, and with what contradictions. You then act on facts, not on an anecdote.
The problem
One day, you ask an AI what it thinks of your brand. It attributes to you a product you don't make. Or it pins a competitor's flaw on you. Or it states false information — confidently, in flawless prose.
First instinct: panic. "The AI is spouting nonsense about us, we have to fix it right away." Bad instinct, for two reasons: you don't yet know whether it's systematic or anecdotal, and in any case you can't "reply" to an AI the way you'd reply to a Google review.
The idea to grasp
To understand why an AI says false things, you have to understand what it does. It doesn't consult a database of truths: it predicts the most probable continuation, word after word. When the probable matches the true, you get a fact. When it diverges, you get an error — and the AI doesn't see the difference.
An image captures it well: imagine an employee who lied on their résumé to land the job. You hand them a task they can't do. They can't say "I don't know" — their résumé claims otherwise. So they fake it: they produce something plausible, with enough confidence that no one asks any questions. The AI, by design, is that employee: it was taught to always answer, never to admit a gap.
Hence three types of errors about a brand:
- Pure invention: the AI doesn't know, so it fills the gap with something plausible (a product, an executive, a date that don't exist).
- Confusion: it attributes to you the characteristics of a neighboring brand that it blends with you in its statistics.
- Outdated information: it repeats old data (a former positioning, an out-of-date fact) because that's what dominates in its memory.
Crucial point: these errors aren't necessarily constant. They can appear in one mode and not another, at one provider and not another, on one wording and not another. Before reacting, you need to know how widespread it is. An error that shows up one time in twenty doesn't call for the same response as a systematic one.
What you hear everywhere
"The AI said something false, we have to correct it." You don't correct an AI the way you reply to a comment. There's no "report" button. Acting requires deep work (on sources, presence, consistency) — and first of all, knowing whether the error is isolated or systematic.
"It's a hallucination, it'll fix itself." Maybe, maybe not. Models evolve, but nothing guarantees that this error will disappear. Waiting for it passively means putting up with it.
"Just write to them to set the record straight." There's no reliable channel to "correct" what an AI says about you. What influences the models is what circulates on the web and what they hold in memory — not a message to support.
My stance: only the facts. Faced with an AI that says false things, the first rational act isn't indignation, it's measurement: which error, how often, where, with what contradictions.
My take: measure the error before reacting
From here on, the register changes: we describe the instrument.
Faced with a problematic AI assertion, measuring lets you qualify the problem:
- Error frequency: across many queries, how many reproduce it? (anecdotal vs systematic).
- Location: in which mode (web / memory), at which providers, on which wordings.
- Contradictions: does the AI say the opposite depending on the question? An internal inconsistency is itself information.
- Misattributed traits: spotting a competitor's characteristics pinned on you (or the reverse).
- Dating and sealing: having a time-stamped record, useful if the matter becomes sensitive (dispute, communication).
The line not to cross. mAIr measures the instability and contradictions in what AIs say. It does not decide what's true or false: there's no absolute reference truth that a tool could arbitrate. mAIr is a contradiction detector, not a lie detector. It says "the AI asserts X here and non-X there," or "the AI attributes such-and-such a trait to you" — it doesn't judge who's right. This restraint is deliberate: it's what makes it a measurement, not an opinion.