AI hallucination
YZ halüsinasyonu
D2
An AI hallucination is a colloquial term for a generative-AI output that is erroneous, false, or inconsistent with its prompt or context.
Review status: 2026-11-26
Technical explanation
NIST calls the phenomenon confabulation and notes that statistically generated outputs can be accurate and consistent in some cases but can also be wrong.
Conceptual boundaries
Hallucination is not evidence of intention or conscious perception, and it does not mean that every model output is false.
Provider-neutral example
If a system confidently invents a report citation that a reviewer cannot locate, the reviewer should treat the claim as unsupported rather than rely on it.
Limitations
The terminology is colloquial and can be contested, so a workflow should evaluate a specific output against relevant evidence instead of treating the label as a complete diagnosis.
Related concepts
Atomic claims and evidence
1.1NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
- Source
- NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
- Source role
- Authoritative source
- Exact locator
- Confabulation pp. 5-6: definition and colloquial terms
- Supported claim
- NIST uses confabulation for GAI systems that generate and confidently present erroneous or false content, including outputs that diverge from prompts or contradict earlier statements in the same context, and notes that these are colloquially called hallucinations.
- Last verification
- Review due
- Scope limitation
- This is NIST’s risk terminology for GAI output and does not establish a clinical or intentional state in a system.
2.1NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
- Source
- NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
- Source role
- Authoritative source
- Exact locator
- Confabulation pp. 5-6: statistical prediction can be accurate or erroneous
- Supported claim
- Statistical next-token prediction can produce factually accurate and consistent outputs, but it can also produce erroneous outputs.
- Last verification
- Review due
- Scope limitation
- This is a general risk statement and does not determine the accuracy of a particular response.
3.1NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
- Source
- NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
- Source role
- Authoritative source
- Exact locator
- Confabulation pp. 5-6: confabulation and colloquial hallucination terminology
- Supported claim
- NIST calls this output phenomenon confabulation and notes that hallucination is a colloquial label; the cited PLOS opinion argues that the label is metaphorical.
- Last verification
- Review due
- Scope limitation
- This records terminology framing and does not make a claim about LLM consciousness or mechanism.
3.2Smith, Greaves, and Panch, Hallucination or Confabulation? Neuroanatomy as metaphor in Large Language Models
- Source
- Smith, Greaves, and Panch, Hallucination or Confabulation? Neuroanatomy as metaphor in Large Language Models
- Source role
- Supplementary source
- Exact locator
- Published Opinion: paragraph beginning ‘Employing the term’
- Supported claim
- NIST calls this output phenomenon confabulation and notes that hallucination is a colloquial label; the cited PLOS opinion argues that the label is metaphorical.
- Last verification
- Review due
- Scope limitation
- This records terminology framing and does not make a claim about LLM consciousness or mechanism.