# What are LLM hallucinations?

> An LLM hallucination is fluent output that is false or unsupported by its input. Types, causes, and what reduces them, from the main surveys and papers.

Source: https://www.honeyhive.ai/learn/llm-hallucinations (updated 2026-11-08)

An LLM hallucination is output that reads as fluent and confident but is false, or that conflicts with the input the model was given. Anthropic's docs [describe](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) it as text that is factually incorrect or inconsistent with the given context. A [2023 survey](https://arxiv.org/abs/2311.05232) calls the content "plausible yet nonfactual", which is the part that makes it hard to catch: the wrong answer looks like a right one.

## Types of hallucination

The survey by Huang and colleagues, accepted by ACM Transactions on Information Systems, [sorts](https://arxiv.org/html/2311.05232) hallucinations into two groups.

| Group | Subtype | What goes wrong |
|---|---|---|
| Factuality | Factual contradiction | Real-world facts are stated incorrectly (a wrong entity or a wrong relation between entities) |
| Factuality | Factual fabrication | Claims that cannot be checked against established knowledge, including overclaims |
| Faithfulness | Instruction inconsistency | The output departs from what the user asked for |
| Faithfulness | Context inconsistency | The output contradicts the context it was given, such as a retrieved document |
| Faithfulness | Logical inconsistency | The output contradicts itself, often in multi-step reasoning |

The split matters for what you do about it. Factuality errors are about the model's knowledge of the world. Faithfulness errors can happen even when the right facts are in the prompt, which is why they show up in [RAG pipelines](/learn/rag-pipeline) that retrieve good documents and still get an answer that ignores them.

Hallucination was studied well before chat models. An [earlier survey](https://arxiv.org/abs/2202.03629) by Ji and colleagues, first posted in February 2022, covers it in summarization, dialogue, question answering, data-to-text, translation and visual-language generation.

## Why models hallucinate

The 2023 survey groups causes by where they enter the pipeline:

- **Data.** Misinformation and bias in pretraining data, facts the model never learned (its knowledge boundaries), and weak alignment data.
- **Training.** Pretraining, supervised fine-tuning and RLHF can each contribute.
- **Inference.** Imperfect decoding, over-confidence and reasoning failures.

A September 2025 paper by Kalai, Nachum, Vempala and Zhang goes further on the statistical side. It [argues](https://arxiv.org/abs/2509.04664) that hallucinations arise as ordinary errors in binary classification: if a model cannot tell false statements from true ones, pretraining produces errors through normal statistical pressure. For arbitrary facts, the paper [states](https://arxiv.org/html/2509.04664) that the post-pretraining hallucination rate should be at least the fraction of training facts that appear only once. If 20% of birthday facts appear once, it expects base models to hallucinate on at least 20% of birthday facts. In the paper's own example, a model asked for one author's birthday, and told to answer only if it knew, gave three different wrong dates.

The same paper blames evaluation for keeping the problem alive. Models are tuned to score well on tests, and on a test that gives zero for "I don't know" and zero for a wrong answer, guessing is the better strategy. The authors [propose](https://arxiv.org/html/2509.04664) that benchmarks state a confidence target in the instructions, for example "Answer only if you are >t confident, since mistakes are penalized t/(1-t) points", and that mainstream benchmarks adopt it. They also reject the claim that hallucinations are inevitable: a model that answers only a fixed set of known questions and otherwise says "I don't know" would not hallucinate, though it would not be very useful.

These are research arguments, and the surveys name causes without ranking them.

## What reduces hallucinations

Anthropic's [guide](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) lists the following techniques. It is written for Claude, but the prompts work the same way with other models.

| Technique | How it works |
|---|---|
| Permit "I don't know" | Tell the model it may admit uncertainty or missing information; Anthropic says this can drastically reduce false information |
| Quote first | For long documents (over 20k tokens), have the model extract word-for-word quotes before it answers, and base the answer on them |
| Cite and retract | Have the model find a supporting quote for each claim, and remove any claim it cannot support |
| Restrict knowledge | Instruct the model to use only the provided documents, not its general knowledge |
| Best-of-N comparison | Run the same prompt several times; inconsistent answers may signal a hallucination |
| Reasoning review | Read the summarized reasoning when an answer looks wrong, to find faulty logic or assumptions |
| Iterative refinement | Feed the output back with a request to verify or expand earlier statements |

Anthropic [adds](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) that these techniques reduce hallucinations significantly but do not eliminate them, and that critical information should always be validated. For the full set of layers around the prompt, see [how to reduce hallucinations in LLM apps](/learn/reduce-llm-hallucinations). Grounding the model in retrieved documents is covered in [RAG pipelines](/learn/rag-pipeline) and [chunking strategies](/learn/chunking-strategies); reasoning prompts are in [chain-of-thought prompting](/learn/chain-of-thought-prompting).

The Kalai paper adds a point for anyone building evals: if your own scoring gives no credit for abstaining, you are rewarding the behavior you are trying to remove.

## Measuring hallucinations in your own app

No benchmark number tells you how often your app hallucinates, because it depends on your prompts, your documents and your users' questions. Measure it directly (the methods are compared in [hallucination detection](/learn/hallucination-detection)):

1. Collect real questions, with the source passage or reference answer for each, into a dataset.
2. Score each output for the two failure families above: is every claim supported by the provided context (faithfulness), and is it true (factuality)? Use code checks where you can, an [LLM judge](/learn/llm-as-a-judge) for the rest, and human review to check the judge.
3. Count abstentions separately from wrong answers, so a change that makes the model say "I don't know" more often is not scored as a regression or hidden as a win.
4. Repeat on sampled production traffic, since real inputs drift from the dataset.

See [LLM evaluation](/learn/llm-evaluation) for the wider method and [LLM observability](/learn/llm-observability) for tracing production calls.

In HoneyHive, [LLM evaluators](https://docs.honeyhive.ai/v2/evaluators/llm.md) score traces and experiment outputs with a custom prompt and rubric, [online evaluations](https://docs.honeyhive.ai/v2/monitoring/onlineevals.md) run code or LLM judges on production traces and surface quality drift on the dashboard, and [annotation queues](https://docs.honeyhive.ai/v2/evaluation/annotation-queues.md) collect expert labels that become ground truth for those evaluators.

## FAQ

### What is an LLM hallucination?

Text a model generates that is plausible but false, or that conflicts with the context it was given. Anthropic's docs [define it](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) as output that is factually incorrect or inconsistent with the given context.

### What are the types of LLM hallucinations?

A [2023 survey](https://arxiv.org/abs/2311.05232) splits them into factuality hallucinations (conflict with real-world facts) and faithfulness hallucinations (conflict with the user's input or with the output itself).

### Why do LLMs hallucinate?

Kalai and colleagues [argue](https://arxiv.org/abs/2509.04664) that pretraining produces them as ordinary statistical errors, and that benchmarks keep them alive by rewarding guesses over abstaining.

### Can hallucinations be eliminated?

Not in a general-purpose model today. Anthropic [says](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) its techniques reduce hallucinations significantly but do not eliminate them, so critical output still needs validation.

### How do you measure hallucinations in an LLM app?

Score outputs against the source material or a reference answer, with human review, code checks or an [LLM judge](/learn/llm-as-a-judge), on a fixed dataset before release and on sampled production traffic after.