---
id: con-hallucination
schema: concept/1
layer: garden
status: established
name: Hallucination
aka: ["making things up", "confabulation", "fabrication", "AI errors", "fake citations"]
summary: >
  When AI states something false or invented, such as a statistic, quote, source or policy detail,
  with the same confidence as something true.
first_seen: "2022"
relations:
  builds_on: [con-large-language-model]
  evolved_from: []
  related: [con-sycophancy, con-jagged-frontier]
  parallels:
    - idea: "confabulation"
      field: "psychology"
      note: "People with memory gaps sometimes fill them with detailed, sincere, invented stories."
    - idea: "a confident witness who misremembers"
      field: "law"
      note: "Confidence and accuracy are different things; courts learned to look for corroboration."
implementations: []
competencies: [cmp-check-the-result]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "The best-known AI failure; anchors why checking against something outside the model matters."
---

Language models produce likely-sounding text. Most of the time, likely and true line up. Sometimes they don't, and the AI produces a made-up statistic, a quote nobody said, a court case that doesn't exist, or a policy rule that sounds right for your organisation but isn't. It doesn't flag these.

Hallucinations are most common with specific details: numbers, names, dates, citations, and anything niche or very recent. They are rarer, though still possible, when the AI works from material you gave it.

The tone of an answer tells you nothing about its accuracy. What helps is knowing which kinds of claims need checking and having a quick way to check them.

## Example

A communications lead asks for "three statistics on volunteer retention with sources." One source is real, one is real but says something different, and one doesn't exist. She keeps the real one and replaces the others with figures from a report she trusts.

## How it connects

- A direct result of how a **large language model** predicts text.
- Often confused with **sycophancy**, where the AI bends facts toward what you want to hear.
- **Verification** is the habit that catches it; the **jagged frontier** explains why it appears on some tasks and not others.

## Try it

Ask your AI tool a factual question about your own field where you already know the answer. Write the correct answer down first, then compare. Repeat with a very specific detail, such as a date or a figure.