Hallucinations: fluent text without support
A response can be well formed and plausible while inventing a fact, source, or event.
In this lesson
Language plausibility and factual support are separate properties.
Name the failure precisely
“Hallucination” is an informal umbrella term. It can refer to an unsupported claim, a contradiction of supplied evidence, an invented citation, or a false report of a tool action. Being specific helps: “the cited paper does not contain this result” is more actionable than simply labeling an answer hallucinatory.
The prediction objective does not inherently distinguish a true claim from a common misconception. TruthfulQA, introduced in 2021, tested whether models reproduced false beliefs in questions designed around misconceptions. Its historical results describe the tested models and benchmark, not a fixed error rate for all current systems.
Why confident wording is weak evidence
A probability distribution over next tokens concerns continuations under a model. It is not automatically a calibrated probability that an entire statement is true. A sentence can have predictable wording because it follows a familiar pattern, even when the names, dates, or causal claims inside it are invented.
In an original example, a model completes “The paper was published in…” with a plausible year and journal. The grammar can be excellent while both details are wrong. Lowering temperature may make the same unsupported completion more repeatable. Repeatability and correctness are different properties.
Confidence and support are separate checks.
How it sounds
- Fluent wording
- Confident tone
- Plausible detail
What supports it
- Traceable evidence
- Source agrees
- Claim checked
Reduce the opportunity for unsupported claims
Relevant retrieval can supply evidence; a calculator can check arithmetic; a database can provide a current record. These measures help only when the right tool is used and its result is represented faithfully. A citation without support is still a failure, and a tool result can be stale or irrelevant.
An answer can also state the boundary of available evidence: “The provided document does not give a date.” That is useful behavior when the question cannot be answered from the context. Evaluation should reward appropriate uncertainty rather than treating every unanswered question as worse than a confident guess.
Test the claim, not the tone
For a factual response, break the text into checkable claims. Identify the evidence for each, inspect whether dates and units match, and distinguish what the source states from what the answer infers. This is a practical reading strategy, not a claim that every sentence needs an external citation.
For a system evaluation, include questions with false premises, missing information, contradictory documents, and invented references. Record whether the system corrects the premise, asks for clarification, or fabricates. A useful assistant should remain useful under uncertainty; it should not conceal missing evidence behind polished prose.
A citation that exists can still fail
An answer cites a real study about short English summaries to support a claim about all languages and long documents. The source is genuine, but the claim exceeds its evidence. Checking only whether the URL opens would miss the error.
Confident language or low temperature makes an answer factual.
Tone and decoding settings do not verify claims. Evidence and task-specific checks are needed.
Take it with you
- Describe the exact kind of unsupported claim.
- Token probability is not a truth probability.
- Evaluate abstention and evidence use alongside answer accuracy.
Check your understanding.
Choose an answer, then see the explanation. You can try again.
Practice only. Your answer is not stored or sent anywhere.
Evaluation: measuring the behavior you need
Continue when you’re ready. Marking this lesson as read is optional.
Sources & further reading
Primary references for this lesson. The explanation and examples above are original.
- TruthfulQA: Measuring How Models Mimic Human Falsehoods ↗Stephanie Lin, Jacob Hilton and Owain Evans · 2021-09-08 · checked 13 Nov 2026