AI hallucinations are answers that sound confident but are false or unsupported: a court case that never happened, a statistic nobody measured, a software package that does not exist. They happen because a language model predicts plausible text instead of looking facts up. Its training also rewards a confident guess over “I don’t know.”
You cannot switch hallucinations off, but you can make them rarer and much easier to catch. Here is why they happen, according to recent research, and nine habits that catch them.
- A hallucination is a fluent, confident answer that is false or not supported by any source.
- Models predict likely text. When they do not know something, a plausible guess reads exactly like a fact.
- Training and benchmarks reward guessing: “I don’t know” scores zero, while a lucky guess scores a point.
- The risk is highest for specifics: names, numbers, dates, quotes, citations, links and package names.
- Catch them by grounding the model in sources, letting it admit doubt, and checking specifics before you use them.
What is an AI hallucination?
Some researchers prefer “confabulation,” the medical term for filling a memory gap with a confident invention. A 2024 study in Nature used it for the most arbitrary errors: answers that change when you simply ask again.
They come in a few recognizable shapes.
| Type | What it looks like | Example |
|---|---|---|
| Invented fact | A confident detail with no source behind it | A wrong founding year, price or birthday |
| Fake citation | A source that does not exist, or does not say what is claimed | A court case or paper made up to support a point |
| Unfaithful summary | A summary that adds, drops or reverses what the text said | “The contract allows early exit” when it forbids it |
| Invented code | A function, option or package that does not exist | An install command for a library nobody published |
| False premise | Going along with a wrong assumption in your question | Explaining why a product has a feature it never had |
The unfaithful summary is the odd one out. The model had the right text in front of it and still strayed, so the fix is quotes, not more sources.
Why does AI hallucinate?
The simplest cause is missing information. A model knows nothing after its training cutoff, and nothing about your files unless you share them. Ask anyway, and it fills the gap. But models also hallucinate about things they could know. Research points to four reasons.
It predicts text instead of looking things up
A language model learns patterns from its training text, then writes the most plausible continuation of your prompt. For well-known facts, plausible and true usually match. For rare facts, they drift apart. Our plain-English guide to how LLMs work covers this prediction process in full.
A September 2025 paper by researchers at OpenAI and Georgia Tech shows why. An arbitrary fact, like a person’s birthday, has no pattern to learn. If 20% of birthday facts appear exactly once in the training data, the authors write, expect a base model to get at least 20% of birthday questions wrong.
Training rewards a confident guess
The same paper argues that the bigger problem is grading. Most benchmarks score answers as right or wrong, and “I don’t know” earns nothing. A model that always guesses beats an honest one on the leaderboard, so models learn to guess.
The authors examined ten influential benchmarks. Nine gave no credit at all for admitting uncertainty. Models, they write, “are optimized to be good test-takers.” Their example of a typical bluff is “September 30” instead of “sometime in autumn.”
The “I know this” signal misfires
Anthropic’s researchers looked inside their Claude model to see how it decides whether to answer. By default, a circuit tells the model it cannot answer. Recognizing a familiar name switches that circuit off.
Hallucinations happen when the recognition fires without the facts behind it. A name feels familiar, the brake releases, and the model confabulates the rest. The researchers triggered this on purpose, making Claude claim that an unknown man played chess.
It wants to agree with you
Assistants are tuned with human feedback, and people tend to prefer answers that match their views. A 2023 study found that five leading assistants “consistently exhibit sycophancy,” telling users what they want to hear. Build a question on a false premise, and the answer may build on it too.
Do reasoning models hallucinate less?
Not automatically. Reasoning models work through a problem before they answer, which helps with logic and math. But more reasoning can also mean more claims, and every claim is a chance to be wrong.
In April 2025, OpenAI reported that its o3 reasoning model hallucinated on 33% of questions in PersonQA, a test of facts about people. The older o1 did so on 16%, and the smaller o4-mini on 48%. OpenAI wrote that o3 “tends to make more claims overall,” and that more research was needed.
Anthropic’s docs describe its latest models as “less prone to hallucinations” when working with code, yet say no technique eliminates them. Treat any published rate as specific to one test, and check a model on your own questions.
Where hallucinations hide
A model can get the gist right and invent the details. Spend your checking time on specifics, and relax a little where an error would be obvious or harmless.
| Always check | Usually safe after a quick read |
|---|---|
| Numbers, prices, dates and statistics | Rewording or tightening your own text |
| Names of people, products, laws and court cases | Brainstorming and first drafts |
| Quotes, citations and links | Explanations of well-known concepts |
| Anything recent or niche | A summary of a short text you can reread |
| Function names, options and package names | Code that you run and test |
Skipping the left column has a documented cost:
9 ways to catch AI hallucinations
Think of these as three checkpoints: before you ask, while the model answers, and before you use the result. Match the effort to the stakes. A birthday message needs none of this. A contract summary needs most of it.
Before you ask
Give it the source. Paste the document, attach the file or turn on web search. A model answering from text in front of it has much less to invent. This is the idea behind retrieval-augmented generation.
Let it say “I don’t know.” Say so outright. Anthropic’s guidance says this one instruction “can drastically reduce false information.” Our prompt engineering guide covers more habits like it.
Ask neutral questions. “Why did the launch fail?” assumes it failed. “How did the launch go, and what went wrong, if anything?” leaves the model room to disagree.
This prompt combines all three. Its scoring line borrows the OpenAI paper’s fix: make a wrong answer cost more than a blank.
Answer using only the document below. Rules: - Support each claim with a short quote from the document. - If the document does not answer the question, reply "Not in the document." - Scoring: a correct answer earns 1 point, "I don't know" earns 0, and a wrong answer loses 9 points. Only answer if you are more than 90% sure. Document: [paste the document] Question: [your question]
While it answers
Quote first, answer second. For long documents, ask for the exact supporting quotes before the answer. Anthropic recommends this for documents over about 20,000 tokens, because it anchors the answer in real text.
Ask for sources, then open them. A citation only helps if it exists and says what the model claims. Click every link, and search every case name, paper title and quote. Fake citations are formatted exactly like real ones.
Ask twice. Regenerate, or ask again in a fresh chat. If the answers disagree on a fact, treat that fact as unknown. The 2024 Nature study turned this into a detection method, because models tend to confabulate differently each time. Agreement is not proof, but disagreement is a strong warning.
Before you use it
Check the claims one by one. List every factual claim, then verify each one separately, ideally in a fresh session so the original answer cannot sway the check. A 2023 method called chain-of-verification uses this pattern and reduced hallucinations across several tasks.
Confirm specifics at the source. Numbers, dates, prices, laws and quotes deserve a primary source: the official page, the paper, the filing. A second model makes a useful skeptic, but models share blind spots, so agreement between two is not proof.
Run it. For code, running it and its tests is the fastest fact-check there is. Before you install a package a model suggests, confirm it exists and is the one you expect, because attackers register the names models invent. The trick is called slopsquatting.
Below is an answer from an AI model. Do not assume it is correct. 1. List every factual claim in it: names, numbers, dates, quotes and sources. 2. For each claim, write one question that would verify it. 3. Answer each question on its own, without relying on the original answer. 4. Mark each claim as supported, contradicted or unverifiable, and say why. Answer to check: [paste the answer]
A 60-second check before you trust an answer
FAQ
Can AI hallucinations be eliminated?
Not with today’s models. The OpenAI researchers argue they are not inevitable, since a model can always say “I don’t know,” but that means answering fewer questions. Grounding, better scoring and checks make them rarer, not impossible.
How common are AI hallucinations?
It depends on the model, the task and the test. In OpenAI’s April 2025 tests, o3 hallucinated on 51% of the short fact questions in its SimpleQA benchmark. Answers grounded in a document you provide go wrong less often, but still do.
Does RAG stop hallucinations?
It reduces them but does not stop them. The legal research tools in the Stanford study above used retrieval and still hallucinated 17% to 33% of the time. You still need quotes and checks.
- Hallucinations are confident guesses, a side effect of how models are trained and graded.
- Give the model sources and explicit permission to say “I don’t know.”
- Check specifics claim by claim, and run any code before you trust it.
Next, learn how to debug with AI step by step, or see when reasoning models are worth the wait.
- Why language models hallucinate, Kalai, Nachum, Vempala and Zhang (OpenAI and Georgia Tech), September 2025
- Tracing the thoughts of a large language model, Anthropic, March 2025
- Towards understanding sycophancy in language models, Sharma et al., October 2023
- OpenAI o3 and o4-mini system card, OpenAI, April 2025
- Detecting hallucinations in large language models using semantic entropy, Farquhar et al., Nature, 2024
- Chain-of-verification reduces hallucination in large language models, Dhuliawala et al., September 2023
- Hallucination-free? Assessing the reliability of leading AI legal research tools, Magesh et al., Stanford, May 2024
- We have a package for you! A comprehensive analysis of package hallucinations by code generating LLMs, Spracklen et al., USENIX Security 2025
- AI Hallucination Cases database, Damien Charlotin, updated September 21, 2026
- Reduce hallucinations, Anthropic docs, accessed September 2026
- Prompting best practices, Anthropic docs, accessed September 2026
