← Back to concept explainers

Concept Fable

Hallucination

Conclusion first:Hallucination is information generated by a model that sounds plausible but is actually wrong, unsupported by evidence, or fabricated. It commonly occurs when a model tries to fill in missing facts, answers with overconfidence, confuses similar information, or lacks reliable context. For enterprise Agents, the danger of hallucination is not just giving a wrong answer, but that erroneous content may enter decisions, customer communications, contracts, code, or business systems.

Hallucination · LLM Behavior

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

Told like a mapmaker who saw it with his own eyes

In Yuanxing Town there was a mapmaker who was a very good talker. When travelers asked him how to get to North Mountain, he could describe the rivers, woods, post stations and mountain passes, and it sounded as if he had walked the route many times himself. One day a merchant caravan asked him to draw the route to the silver mine. The mapmaker spread out paper and drew a stone bridge, three pine groves and an old well. He said, after crossing the bridge head west, and when you see the well turn right, and you will reach the silver mine. The caravan followed his directions and ended up in a swamp. There was no stone bridge there, and no old well.

The caravan came back and questioned him. The mapmaker was surprised and said he had not intentionally deceived anyone; he had merely inferred that the north was similar based on the appearance of several southern roads. Stone bridges are common, old wells are common, and this route sounded reasonable.

After hearing this, the old guide spread out the old map: the road you drew did not come from exact observation, it was only what you thought should exist. It looks like a real map, but it cannot lead anyone.

The mapmaker argued that his inferences were indeed useful in many cases. The old guide nodded: in casual chat, reasonable inference may be harmless; when a caravan sets out with goods, time, and risk, an error that sounds true is more dangerous than saying "I don't know."

Later, the town set three rules for the mapmaker. Roads walked with one's own eyes were drawn with solid lines, roads heard from others with dashed lines, and merely guessed roads could not be drawn as roads; key routes had to be checked against old maps and road signs; when uncertain, one had to say so clearly.

The mayor first told the mapmaker to speak more slowly and cautiously. But slowness could not make a nonexistent stone bridge appear, and caution could not find the traveler a road. What really needed to change was that beside every route there had to be labels for personally tested, heard, and inferred—the three could not be drawn with the same ink line.

Later, when travelers came again to ask the way, the mapmaker first laid out the old maps, the road steles, and the guides' accounts. Where there was no evidence, he would rather leave a blank than fill the page with handsome mountain lines. The caravans found that although the blanks looked ugly, they were more reliable than a confident false road.

The map lost its legendary flavor but saved more caravans. A reliable map does not rush to fill in every blank; it knows where the blanks still are.

Reveal

This story is about: hallucination

Hallucination is information generated by a model that sounds plausible but is actually wrong, unsupported by evidence, or fabricated. It commonly occurs when a model tries to fill in missing facts, answers with overconfidence, confuses similar information, or lacks reliable context. For enterprise Agents, the danger of hallucination is not just giving a wrong answer, but that erroneous content may enter decisions, customer communications, contracts, code, or business systems.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Mapmaker: generative models, skilled at organizing coherent content based on patterns
  • Stone bridge, pine forest, old well: details that sound concrete and credible but are not supported by evidence
  • The caravan walks into the swamp: hallucinated content is used in real actions and causes losses
  • Inferring the north from the south road: the model completes based on patterns, not based on factual verification
  • Like a real map but not a map: fluency and specificity do not equal truth
  • Solid lines, dashed lines, and what cannot be drawn as a road: distinguishing known facts, reported information and speculation
  • Check old maps and road signs: use retrieval, citations, fact-checking, and tool verification to reduce hallucinations
  • Knowing where the blanks still are: a reliable system must be able to express uncertainty

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

groundingfactualityRAGuncertaintyevaluation
京ICP备2026014272号