A hallucination is when a large language model produces confident, fluent text that is factually wrong or unsupported — inventing a citation, a…
Understand why a model can be fluent, confident, and wrong at once.
A hallucination is model output that sounds authoritative but is false or unsupported — a made-up statistic, a fake quote, a nonexistent function, or a citation to a paper that was never written. The danger is the tone: hallucinations are delivered with the same fluent confidence as correct answers.
The cause is fundamental. A language model predicts the most likely next token given its input; it optimizes for plausible-sounding text, not verified truth. When it lacks the fact, it fills the gap with something that fits the pattern — which is often wrong.
Hallucination is not deception and not a simple bug — it is a side effect of how these models work. It can be reduced a great deal, but no current model is fully immune, so design systems that assume it can happen.
Recognize the common forms so you know what to watch for.
Hallucinations cluster in predictable places. Fabricated facts and figures: invented dates, numbers, or events. Fake sources: plausible-looking citations, URLs, or quotes that don't exist. Invented specifics: nonexistent library functions or API parameters in generated code. And unfaithful answers in retrieval systems, where the model strays from the provided context.
Risk rises when the question is about niche, recent, or private topics the model didn't learn well, when it is pushed to be specific, and when there is no source to check against.
See how connecting the model to real sources makes claims checkable.
Grounding means giving the model real information to answer from and holding it to that evidence, instead of relying on its memory. The main tool is retrieval-augmented generation (RAG): fetch relevant passages and put them in the prompt, then instruct the model to answer only from them and to cite which passage supports each claim.
Grounding cuts hallucination two ways: it supplies facts the model might not know, and it makes answers checkable — a reader (or an automated check) can verify each claim against the cited source. Telling the model to say 'I don't know' when the sources don't cover the question is what stops it from filling the gap.
Layer practical defenses and know what grounding cannot fix.
Beyond RAG, several practices help. Require citations and verify them. Ask the model to express uncertainty and to answer only from provided context. Use retrieval for anything factual, recent, or private. For high-stakes outputs, add a verification step — a second check that each claim is supported — and keep a human in the loop.
Lower-level tuning (better instructions, sometimes lower temperature for factual tasks) helps at the margins, but grounding and verification are the load-bearing defenses.
Grounding reduces hallucination but doesn't eliminate it: the model can still misread a source, cite the wrong passage, or stray from provided context (an unfaithful answer). And if retrieval fetches the wrong documents, the model is grounded in bad evidence. Evaluate faithfulness on real questions, verify citations, and never treat a fluent, cited answer as automatically true.
A hallucination is confident, fluent output that is false or unsupported, caused by the model predicting plausible text rather than verified truth. It clusters around niche, recent, or private topics and pressure to be specific. Grounding is the main defense: retrieval-augmented generation supplies real sources and makes claims checkable, especially when the model is told to answer only from context and cite it. But grounding isn't a cure — verify citations, evaluate faithfulness, and keep humans in the loop for high stakes.
You are building an assistant that answers medical policy questions. Describe how you would ground it to reduce hallucination, what you would require in every answer to make claims checkable, and why you would still not fully trust a fluent, cited response.
What is a hallucination in a large language model?
Hallucinations are plausible-sounding but false or unsupported outputs, produced because the model predicts likely text rather than verified truth.
Why do language models hallucinate?
Hallucination is a side effect of next-token prediction: lacking a fact, the model generates something that fits the pattern, which is often false.
How does grounding reduce hallucination?
Grounding supplies facts and makes answers verifiable against cited sources, turning a closed-book task into an open-book one.
What is a limitation of grounding?
Grounding depends on good retrieval and faithful use of context; verify citations and evaluate faithfulness rather than trusting a fluent, cited answer.