AI tools can write a clear, confident answer that is completely wrong. That failure has a name, and anyone who builds with or relies on large language models needs to understand it.
This guide explains what AI hallucinations are, why they happen, how often they occur and what you can do to reduce them.
What Is an AI Hallucination?
An AI hallucination is an answer from an AI model that sounds plausible and confident but is false, made up or not supported by any source. In the words of OpenAI's hallucination research, "Hallucinations are plausible but false statements generated by language models."
Hallucinations show up as invented facts, fake quotes, citations to papers that don't exist or details a source never mentioned.
A hallucination is not a rare glitch. It is a normal side effect of how today's models generate text, which is why the causes matter.
Why Do AI Models Hallucinate?
AI models hallucinate for two broad reasons. Some causes come from how the model is built and trained. Others come from the data the model receives at the moment it answers.
Most explainers cover only the first group. If you build AI products, the second group is often the one you can control.
Models Predict Likely Words, Not Verified Facts
A large language model (LLM) is a model trained on huge amounts of text to predict which words are likely to come next. It produces an answer by repeating that prediction, word by word.
That process has no built-in fact-check step. The model produces what sounds right based on patterns, even when the pattern points to a false detail.
Training Rewards Guessing Over "I Don't Know"
Research on why language models hallucinate puts it directly: "language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty." Most tests give credit for a right answer and none for "I don't know," so a wrong guess scores the same as an honest abstention.
The same paper shows that facts appearing only once in training data are especially likely to be hallucinated. Models learn to give an answer even when they shouldn't.
Knowledge Cutoffs Make Answers Go Stale
A knowledge cutoff is the date after which a model has no training data. Once training ends, the model's internal knowledge stays frozen unless it is given new information.
This creates three different situations that look the same to a user:
- Never learned it: the fact was rare or missing from the training data.
- Learned it, but it changed: the fact was true at training time and is now out of date, such as a price, a policy or a company's CEO.
- Wasn't given it: the answer depends on private or recent information the model never received.
Each situation needs a different fix. Retraining won't help with yesterday's news, and better prompts won't help when the facts were never supplied.
Bad or Outdated Retrieved Context
Retrieval-augmented generation (RAG) is a method that finds relevant documents and feeds them to the model when it answers. RAG helps with stale knowledge, but it adds a new risk: the model now depends on what it retrieves.
If retrieved pages are outdated, the model can follow them into a wrong answer. A stale-document poisoning study found that "outdated retrieval flips 30% of Llama and 37% of Qwen answers even without instructions to trust the document." In those cases, the model would have been right without the retrieved text.
Noisy context causes similar problems. Raw HTML full of menus, ads and scripts buries the facts a model needs, which is why teams convert pages into clean, LLM-ready web data before passing them to a model.
Types of AI Hallucinations
AI hallucinations fall into a few common types. Two terms help describe them: an intrinsic hallucination contradicts the source the model was given, and an extrinsic hallucination adds claims the source can't confirm.
The table below links each type to its likely cause and a practical first fix.
| Type | What It Looks Like | Likely Cause | First Fix |
|---|---|---|---|
| Made-up facts | A confident but false date, number or name | Rare fact in training data; pressure to guess | Allow "I don't know"; ground in sources |
| Outdated facts | An old price, policy or job title stated as current | Knowledge cutoff or stale retrieved pages | Retrieve fresh sources; check document dates |
| Fabricated citations | A paper, case or URL that doesn't exist | Model generates a plausible-looking reference | Confirm every cited source exists |
| Misused real sources | A real source cited for a claim it doesn't make | Weak link between claim and source | Re-read the cited page for the exact claim |
| Contradicting the source (intrinsic) | A summary that reverses what a document says | Model misreads or overrides the context | Compare output against the input text |
| Unsupported details (extrinsic) | Extra "facts" a source never mentions | Model fills gaps from memory | Restrict answers to the supplied context |
| Reasoning errors | Correct facts combined into a wrong conclusion | Multi-step reasoning breaks down | Break tasks into checked steps |
Real-World Examples of AI Hallucinations
Hallucinations already cause real damage in law, research and business. The first two examples below come from tracked, current data rather than one-off anecdotes.
- Courts: Lawyers and people representing themselves have filed briefs citing cases that don't exist. An AI hallucination cases database maintained by researcher Damien Charlotin lists "2184 cases identified so far" as of October 10, 2026, and that count only includes documented court decisions.
- Scientific research: Fake references are spreading through academic papers. An audit of 111 million references produced "a conservative estimate of 146,932 hallucinated citations in 2025 alone" across four research repositories.
- Customer-facing chatbots: Consider a support bot. A bot without access to the current policy page can state a refund rule, price or policy the company never offered. The model fills the gap instead of finding the real source.
How Often Do AI Models Hallucinate?
How often AI models hallucinate depends on the task. A model summarizing a document it was given hallucinates far less than a model answering from memory alone.
| Task | What Was Measured | Reported Range | Source and Date |
|---|---|---|---|
| Summarizing a supplied document | Summaries that add unsupported content | About 1.8% to 24.2% across listed models | Vectara leaderboard, May 11, 2026 |
| Answering questions from memory (closed-book) | Wrong answers among questions the model could not answer correctly | 22% to 94% across 26 models | Stanford AI Index 2026, Apr 2026 |
The Vectara hallucination leaderboard tests grounded summarization and reports rates of about 1.8% to 24.2% across the models it lists. On closed-book questions, the Stanford AI Index 2026 reports that "Across 26 models, hallucination rates range from 22% to 94%."
These numbers measure different things, so they can't be compared directly. Be cautious with any claim that "AI hallucinates X% of the time" unless it names the task, the benchmark and the date.
Why RAG and AI Agents Still Hallucinate
RAG and AI agents reduce some hallucinations, but they don't remove them. They move much of the risk from what the model remembers to what the system retrieves.
An AI agent is a system that plans steps, calls tools and acts on the results, such as searching the web, reading pages and filling in a report. Understanding how agentic search works shows why this matters: an agent that accepts one wrong fact early can carry it into later steps.
Citations help only when the cited page actually supports the claim. A citation to an outdated or unrelated page looks trustworthy but proves nothing. Our guide to reducing hallucinations in grounded responses covers this failure mode in more detail.
How to Reduce AI Hallucinations
You can't eliminate AI hallucinations, but you can make them rarer and easier to catch. The steps below target the data a model receives, the freedom to admit uncertainty and the checks applied to its output.
Ground Answers in Fresh, Clean Sources
Grounding means giving the model current, relevant sources at answer time so it doesn't rely on memory alone. This directly addresses outdated facts and missing information.
- Retrieve at request time: fetch current pages when the question is asked, for example through a live web search API.
- Send clean text: pass Markdown or structured JSON instead of raw HTML so the model sees the facts, not the page clutter.
- Keep dates attached: include publish or update dates with each document so stale sources can be spotted and ranked lower.
Let the Model Say "I Don't Know"
A model that can abstain makes fewer confident errors. OpenAI reported that on the SimpleQA benchmark, gpt-5-thinking-mini abstained on 52% of questions with a 26% error rate, while o4-mini abstained on 1% with a 75% error rate. Both models had similar accuracy.
You can build the same idea into your product. Prompt the model to answer only from supplied sources, and treat "not found" as a valid result. For example, Olostep's Answers endpoint returns sourced answers with NOT_FOUND values when a field can't be verified from web sources, instead of filling it with a guess.
Check Claims Against Their Sources
Checking output against sources catches the hallucinations that grounding misses. Use this checklist for any answer that matters:
- Confirm the source exists. Open every cited URL, paper or case.
- Confirm the source says it. Re-read the cited page for the exact claim. In automated pipelines, you can scrape the cited page and compare its text with the claim.
- Check the date. Make sure the source is current enough for the question.
- Flag high-stakes answers. Send legal, medical and financial answers to a human reviewer.
- Track errors over time. Keep a small test set of known questions and re-run it when you change models or prompts.
Frequently Asked Questions
Can AI Hallucinations Be Fixed Completely?
No. OpenAI's researchers note that some real-world questions are inherently unanswerable, so accuracy can't reach 100%, but grounding, abstention and source checks can reduce confident errors.
Why Does ChatGPT Hallucinate?
ChatGPT predicts likely text rather than looking up verified facts, and its training rewards giving an answer over admitting uncertainty. Gaps or outdated information in what it knows or retrieves make this worse.
Do Newer AI Models Hallucinate Less?
Not necessarily. Hallucination rates vary widely by task and model, so test any model on your own use case before relying on it.
Does RAG Stop AI Hallucinations?
RAG can reduce hallucinations caused by missing or outdated knowledge, but it doesn't stop them. In the stale-document poisoning study cited above, outdated retrieved documents flipped answers that models had gotten right without retrieval.
Does a Citation Mean an AI Answer Is Correct?
No. A citation only helps if the source exists, is current and actually supports the claim, so open and check it.
How Can I Spot an AI Hallucination?
Look for specific details with no source, citations you can't find and facts that conflict with a current official page. Be extra careful when an answer is confident about recent events.
Is an AI Hallucination the Same as AI Bias?
No. A hallucination is a false or unsupported statement, while bias is a systematic slant in outputs, though both can appear in the same answer.