Hallucination rate is the proportion of model outputs, or of claims within those outputs, that are false, fabricated or unsupported by the provided source or by verifiable facts. In model evaluation for artificial intelligence, it is one of the main metrics used to judge whether a language model can be trusted for factual answers, summaries and retrieval-based tasks.
The figure is always relative to a definition, a dataset and a judging method. A model can show a low hallucination rate on short summaries of supplied documents and a much higher one on open questions about obscure facts, so a number without its test conditions says little.
How hallucination rate is calculated
Hallucination rate is calculated by dividing the number of hallucinated units by the total number of evaluated units. The unit can be a whole response, a sentence or an atomic claim, and that choice changes the result substantially.
Response-level scoring marks an answer as hallucinated if it contains at least one unsupported statement. Claim-level scoring breaks each answer into individual factual statements and checks each one, which gives a finer and usually lower percentage for long answers.
| Unit | Counted as hallucinated when | Effect on the rate |
|---|---|---|
| Response level | Any part of the answer is unsupported | Higher for long answers |
| Sentence level | The sentence contains an unsupported claim | Moderate granularity |
| Atomic claim level | The single factual statement is unsupported | Finest, often lower per claim |
| Citation level | A cited source does not support the claim | Specific to grounded answers |
Many benchmarks also track the abstention rate, meaning how often the model declines to answer. A model that refuses frequently can show a low hallucination rate while being far less useful, so the two figures should be reported together.
Types of hallucination being measured
Evaluations distinguish intrinsic and extrinsic hallucinations, and the distinction determines what counts as an error. Intrinsic errors contradict the given source, while extrinsic errors add information that the source cannot confirm.
| Type | Definition | Typical task |
|---|---|---|
| Intrinsic | Output contradicts the supplied input | Summarization, translation |
| Extrinsic | Output adds content not verifiable from the input | Summarization, retrieval answers |
| Factual | Output conflicts with established world knowledge | Open-domain question answering |
| Citation fabrication | References or quotes that do not exist | Research and legal assistance |
| Instruction drift | Output ignores constraints and invents details | Agent and tool use |
In grounded settings such as retrieval-augmented generation, extrinsic content may be true in the real world but still counts as a hallucination, because the system was required to answer only from the retrieved documents.
Methods for detecting hallucinations
Hallucinations are detected by human annotation, automated judges or reference-based checks, and each method has a different cost and reliability profile. Large evaluations usually combine automated scoring with a human-audited sample.
| Method | Strength | Weakness |
|---|---|---|
| Expert human review | Highest accuracy on nuanced claims | Slow, costly, needs agreement checks |
| LLM-as-a-judge | Fast and scalable | Judge model can itself err or be biased |
| Entailment models | Consistent source-to-claim checks | Struggle with paraphrase and numbers |
| Reference matching | Objective for short factual answers | Fails on open-ended text |
| Self-consistency sampling | No reference needed | Consistent errors go undetected |
Whatever the method, reliable studies report inter-annotator agreement or judge-to-human agreement. Without that, a reported hallucination rate may reflect the judge’s errors as much as the model’s.
Factors that change the measured rate
The measured hallucination rate depends heavily on task type, question difficulty, decoding settings and grounding. Comparing figures across different benchmarks is therefore misleading unless the conditions match.
| Factor | Tends to raise the rate | Tends to lower the rate |
|---|---|---|
| Knowledge type | Rare entities, recent events | Common, well-documented facts |
| Grounding | No source documents supplied | Relevant retrieved context |
| Answer length | Long, detailed outputs | Short, constrained outputs |
| Sampling temperature | High temperature | Low or greedy decoding |
| Abstention policy | Model pressured to always answer | Model allowed to say it is unsure |
Training incentives also matter. Benchmarks that reward any answer over no answer encourage confident guessing, while calibration-aware scoring that penalizes wrong answers more than abstentions pushes models toward admitting uncertainty.
Reducing the hallucination rate in practice
Teams reduce the hallucination rate by combining grounding, decoding control, training changes and output verification, because no single technique removes fabrication entirely. The appropriate mix depends on whether the application needs open knowledge or answers drawn from a fixed corpus.
- Retrieval grounding supplies relevant documents so the model can quote rather than recall.
- Citation requirements force each claim to point to a passage that a checker can verify.
- Lower sampling temperature reduces creative but unsupported continuations.
- Preference training for abstention teaches the model to state uncertainty instead of guessing.
- Post-generation verification uses a second model or rule set to flag unsupported claims before display.
Each intervention should be validated on the same fixed evaluation set before and after the change. Otherwise, an apparent improvement may come from easier prompts or a different judge rather than from genuinely more faithful outputs.
Limitations and common misreadings
A single hallucination rate is not a universal property of a model. It is a measurement on one distribution of prompts, and the same model can vary by an order of magnitude between a summarization benchmark and an obscure-fact quiz.
- Benchmark contamination can make a model look more factual than it is on unseen questions.
- Low rate with high abstention may hide poor coverage rather than better knowledge.
- Judge drift occurs when the automated judge model changes between evaluations.
- Severity is ignored by a plain rate: a wrong date and a fabricated legal precedent count the same.
For deployment decisions, teams therefore pair the rate with domain-specific test sets, severity weighting and error analysis of real failures rather than relying on a public leaderboard figure alone.
In plain words
Hallucination rate tells you how often an AI makes things up. If a chatbot answers 100 questions and 7 answers contain invented facts, its hallucination rate on that test is 7 percent.
The number is only meaningful with its context. Easy questions give low rates, hard or obscure questions give high ones. A good result also requires that the model does not simply refuse to answer whenever it is unsure.
A real-world example
A legal services firm tests an AI assistant that summarizes court decisions for its lawyers. The evaluation team prepares 300 decisions with expert-written reference summaries and breaks each model summary into individual factual claims, checked by two paralegals against the source text.
The first version shows about 9 percent of claims unsupported, many of them invented case citations. After the team adds retrieval of the full decision text and instructs the model to quote only from it, the rate falls to under 2 percent of claims, with abstentions rising slightly. The firm approves the tool for internal drafting but keeps mandatory human review for any citation used in a filing.