DataTerminology.comNEW
Also known as: Confabulation
When generative models produce plausible but incorrect outputs not supported by training data or external reality.
DESCRIPTION
Hallucinations occur when models confidently output fabricated facts, citations, or content. They arise from limited knowledge, ambiguous prompts, or misaligned objectives. Mitigation strategies include grounding with retrieval, calibration, prompt constraints, and model improvements. Governance requires documenting risks, monitoring, and user guidance. High-stakes domains need human oversight and verification. Ongoing research targets truthfulness, citation quality, and robustness against adversarial prompts.
ALTERNATIVE DEFINITIONS
Wikipedia
In artificial intelligence (AI), a hallucination or confabulation is a confident response by an AI that does not seem to be justified by its training data.
Highlights unjustified confidence relative to training data.
IBM
AI hallucinations are outputs from generative models that are nonsensical or not based on real-world data.
Emphasizes nonsensical outputs and lack of real-world grounding.
RELATED TERMS
Grounding Retrieval-Augmented Generation Fact Checking Calibration Uncertainty Quantification Knowledge Cutoff Prompt Engineering Factuality Evaluation
COMPETING TERMS
Grounded Generation
KNOWLEDGE AREAS
TARGET AUDIENCE
AI Engineer Data Scientist AI Product Manager Chief Data Officer
EXTERNAL SOURCES
- Wikipedia - Hallucination (artificial intelligence)
- arXiv - Survey on LLM hallucinations
- OpenAI - Strategies to reduce hallucinations
- Google DeepMind - Responsible AI guidance
TOP QUESTIONS
Why do Hallucinations happen in LLMs?
Models optimize next-token prediction and can produce plausible but incorrect content, especially under ambiguity or gaps in knowledge.
Category: Overview
How can Hallucination be reduced?
Ground responses with retrieval, require citations, tune prompts, calibrate confidence, and use guardrails and post-processing checks.
Category: Requirements
What is the biggest misunderstanding about Hallucination?
That larger models eliminate it. Bigger models reduce rates but cannot guarantee perfect factuality without grounding and verification.
Category: Misunderstanding
What are the risks of Hallucination in production?
Misinformation, safety incidents, legal liabilities, and loss of user trust—especially in regulated domains.
Category: Risks
How do you evaluate Hallucination?
Use factuality benchmarks, human review, citation accuracy, and error analysis across domains and query types.
Category: Testing
What is currently hot in Hallucination mitigation?
Retrieval augmentation, tool-use with verification, and self-consistency or debate-style prompting.
Category: Topic
How should teams manage Hallucination with governance?
Document limits, monitor outputs, set escalation paths, and ensure human oversight for high-stakes decisions.
Category: Management
What role does prompt engineering play in Hallucination?
Structured prompts, instructions to cite sources, and constraints reduce fabrication and improve adherence.
Category: Opinion
What is an analogy for Hallucination?
It’s like confidently guessing answers to an open-book exam without opening the book.
Category: Example