跳过主要内容
全部术语
How Engines Retrieve & Rank Sources更新于

AI hallucination

又称: LLM hallucination, confabulation, AI making things up

一句话解释

An AI hallucination is a statement produced by a language model that is fluent, confident and false. It happens because models predict likely text rather than retrieve verified facts, so a plausible-sounding answer is generated with the same confidence as a correct one.

Why fluency is the trap

A hallucination does not look like an error. It reads exactly like a correct answer: same tone, same structure, same confidence, frequently with a citation attached. There is no signal in the output distinguishing a fact the model retrieved from one it produced.

For a brand this matters because errors propagate silently. Nobody reports the assistant that quoted the wrong price. The buyer just believes it.

The brand-specific version

Hallucinations about businesses tend to fall into a few shapes:

  • Stale facts — pricing, features or availability from an old page the engine keeps reading.
  • Attribute transfer — a competitor's feature or limitation attached to you, usually from a comparison article.
  • Confabulated specifics — a plausible integration, certification or policy that does not exist.

The important observation is that most of these are not invention. They are faithful reproductions of a bad source, which makes them tractable.

Fixing it

Find the source first. Ask the engine to cite what it used, then read those pages. The correction is almost always at the source rather than in the model.

Then make the accurate version easy to retrieve: current, specific, and stated plainly on a page the engines already read. Then re-check on a schedule — answers vary run to run, so one improved answer is not evidence the correction landed.

常见问题

Why do AI models hallucinate?
Because their objective is to produce likely text, not true text. When the model has no reliable information, the most likely-looking continuation is still generated — and a fabricated citation or specification looks exactly like a real one to the system producing it.
What do you do if an AI says something false about your business?
Find the source. Hallucinations about brands usually trace to a stale or wrong page the engine keeps reading, not to invention. Correct the source, make the accurate version prominent on pages engines retrieve, and re-check on a schedule until the answer moves.
Does grounding fix hallucination?
It reduces it substantially. Answers built from retrieved sources are far more likely to be correct and are checkable against the citation. It does not eliminate the problem — a model can still misread a source or attribute a claim to a document that does not support it.

相关术语

  • Grounding (AI)

    Grounding is the practice of connecting a language model's answer to specific external sources retrieved at answer time, rather than relying on what the model memorised during training. A grounded answer can cite where each claim came from, which is what makes citations — and therefore AI visibility — possible at all.

  • Large language model (LLM)

    A large language model is a neural network trained on very large amounts of text to predict likely continuations of language. That single capability produces answering, summarising and writing. It is the component underneath ChatGPT, Claude, Gemini and the AI features inside search engines.

  • Brand sentiment

    Brand sentiment is the emotional tone of what is said about a brand — positive, negative or neutral — measured across mentions and aggregated into a trend. In AI answers it extends beyond tone to accuracy: how an engine characterises your product matters as much as whether it sounds favourable.

  • Brand monitoring

    Brand monitoring is the systematic tracking of where and how a brand is mentioned across the web — press, social platforms, forums, reviews and increasingly AI answers. Its purpose is to turn scattered mentions into something countable, so changes in perception and visibility are noticed rather than guessed at.

相关指南