Glossary

Hallucination

When a language model generates confident, fluent output that is factually wrong or unsupported by its source material.

Also called: AI hallucination, confabulation

Hallucination is the term for a language model producing output that sounds plausible and is stated with confidence but is factually incorrect, internally inconsistent, or not actually supported by the documents it was given. It can range from a fabricated statistic to an invented citation or a plausible-sounding but nonexistent API method.

It happens because a large language model is trained to predict likely continuations of text, not to verify facts; when it lacks reliable information on a topic, it still generates a statistically plausible answer rather than saying it does not know. retrieval-augmented generation reduces this by grounding answers in retrieved source text, but does not eliminate it: a model can still ignore, misquote, or extrapolate beyond what was retrieved.

Hallucination matters because it undermines trust in any system that presents LLM output as fact, particularly in regulated or high-stakes domains. Practitioners mitigate it with retrieval grounding, structured prompts that ask the model to cite sources, guardrails that check outputs against source text, and LLM evaluation pipelines, including LLM-as-a-judge checks, that specifically measure factual accuracy rather than fluency.

Last reviewed September 22, 2026

In the index now

Related terms

Large language model (LLM) A neural network trained on vast amounts of text to predict and generate language, the technology behind most modern AI assistants. Retrieval-augmented generation (RAG) An architecture that retrieves relevant documents at query time and feeds them to a language model so its answers are grounded in sources. LLM evaluation Systematically testing a language model's outputs for quality, accuracy, safety, and consistency before and after deployment. Guardrails Rules, filters, or checks placed around a language model to keep its inputs and outputs within acceptable, safe bounds. LLM-as-a-judge Using a language model to score or compare the outputs of another model, as a scalable substitute for human review.

Related tools

Evidently AI Open-source and cloud model-monitoring tool that checks deployed ML/LLM systems for drift, data quality, and performance issues. Fiddler AI Hosted AI observability and guardrails platform for monitoring deployed ML models and LLM/agent applications. Patronus AI Evaluation-focused platform offering pre-built and custom evaluator models to score LLM outputs for accuracy, safety, and quality.

Related guides

How to choose an LLM observability tool Pick an LLM tool by whether you need tracing, scoring or a gateway first — most teams eventually need all three, rarely from one vendor. How to evaluate LLM applications Offline evals, LLM-as-judge, human review and production monitoring, aimed at the whole pipeline rather than the model alone. How to use AI for data analysis safely Text-to-SQL and AI data assistants can speed up analysis or produce fluent wrong answers — the difference is what sits between the model and your data.