AI Hallucinations Explained: Why Chatbots Make Things Up
Why AI chatbots state false facts with confidence, the common types of hallucination, and practical checks that catch errors before you rely on them.
Short answer
An AI hallucination is a confident answer from a chatbot that is false or unsupported, such as an invented citation or a wrong date. It happens because language models generate likely-sounding text instead of looking facts up, so you should verify anything that matters against a primary source.
Key takeaways
- Hallucinations come from how language models work: they predict plausible text and have no built-in fact checker.
- Fake citations, wrong numbers, and invented product features are the most common forms.
- Asking for sources, checking them yourself, and giving the model the document you care about all reduce the risk.
You ask a chatbot for three studies on a topic and get three neat citations, complete with authors and journal names. Two of them do not exist. That is an AI hallucination, and it is one of the most common ways people get burned by otherwise useful tools. This post explains why it happens, what it tends to look like, and how to catch it before it ends up in your report, code, or email.
What is an AI hallucination?
An AI hallucination is output from a generative AI model that is stated with confidence but is false, made up, or not supported by any source.
The word is informal. NIST, the US standards agency, uses "confabulation" in its Generative AI Profile (NIST AI 600-1) to describe the same thing: content that sounds authoritative yet is wrong. The defining feature is the confidence. A model rarely flags these answers as guesses, and the wording looks exactly like its correct answers.
Hallucinations are different from simple outdated information. If a model says a phone launched last year when it launched this year, that may be a training cutoff issue. If it invents a phone model that never existed, that is a hallucination.
Why do language models make things up?
Language models generate the most plausible next words based on patterns in their training data, and nothing in that process checks whether the result is true.
A large language model writes one token at a time, choosing likely continuations (see what a large language model is for the mechanics). For common facts, the most likely continuation is usually correct because the training data repeated it many times. For rare facts, such as a specific court case, a niche paper, or a small company's founding year, the model has weak signals and fills the gap with something that fits the pattern.
A 2025 paper titled "Why Language Models Hallucinate" adds a second reason. Its authors argue that many benchmarks score answers as simply right or wrong, with no credit for saying "I'm not sure." Under that scoring, guessing beats abstaining, so models are tuned to be confident test-takers. They propose changing how existing benchmarks are scored so honest uncertainty is not penalized.
What do hallucinations usually look like?
Most hallucinations fall into a handful of recognizable patterns, and invented sources and precise-looking numbers are the most frequent.
| Type | Example | Why it slips through |
|---|---|---|
| Fake citation | A paper title, author, and DOI that do not exist | Format looks exactly like a real reference |
| Wrong number | A battery capacity or release date that is slightly off | Plausible values are hard to spot without checking |
| Invented feature | A settings menu or app option that was never shipped | Mixes real product names with made-up details |
| False attribution | A quote assigned to the wrong person | Famous names make it feel credible |
| Bad code | A function or library method that does not exist | Code compiles in the reader's head, fails when run |
| Misread source | A summary that claims a page says something it does not | The link is real, so people stop checking |
The last row matters more as chatbots gain web search, because a real link next to a false claim looks convincing.
When are hallucinations most likely?
Hallucinations are most likely when you ask about obscure, recent, or highly specific facts, or when you push the model to give an answer it cannot support.
Watch for these situations:
- Questions about events after the model's training cutoff, when no search tool is used.
- Requests for exact figures, such as prices, percentages, or specifications.
- Niche topics with little public writing, including small businesses and local rules.
- Long conversations, where early instructions or facts drop out of the context window.
- Prompts that assume a false premise, like asking why a feature was removed when it never existed.
The false-premise case is easy to fall into. If you ask "Why did the company discontinue X?" the model may produce a reason rather than tell you X was never discontinued.
How can you catch and reduce hallucinations?
Treat a chatbot answer as a draft, ask for sources, and check every claim that matters against a primary source you open yourself.
A practical routine:
- Give the model the material. Paste the document, manual, or policy you care about and ask it to answer only from that text.
- Ask for sources and open them. Confirm the page exists and says what the summary claims.
- Rephrase and ask again. If two phrasings produce different numbers, neither is reliable yet.
- Ask the model to list what it is unsure about. This does not guarantee honesty, but it often surfaces weak points.
- Run code before trusting it, and check that every imported library and function actually exists.
- For health, legal, or financial decisions, go to an official source or a qualified professional.
Running a model locally does not change any of this; small open-weight models can hallucinate more often than large hosted ones on factual questions. If you are curious anyway, how to run AI models locally covers the hardware side.
Will hallucinations ever go away?
As of November 2026, no model is free of hallucinations, though retrieval, better training, and scoring that rewards uncertainty can lower how often they happen.
Connecting models to search engines or company documents (often called retrieval-augmented generation) helps with current or private facts. Training changes that reward a model for saying "I don't know" address the incentive problem described above. Neither removes the core issue that the model generates text rather than retrieving verified records.
For now, the safest habit is to match your checking to the stakes. A brainstormed list of gift ideas needs no verification. A number you plan to publish, a quote you plan to attribute, or a command you plan to run on a server does. One more point: what you paste into a chatbot to ground its answers may be stored, so check your AI chatbot privacy settings before sharing sensitive documents.
Frequently asked questions
Why do chatbots hallucinate instead of saying 'I don't know'?
Research published in 2025 argues that training and evaluation often reward guessing over admitting uncertainty, much like a multiple-choice test with no penalty for wrong answers. A model tuned to score well on those tests learns that a confident guess pays off more often than a refusal.
Does web search stop hallucinations?
It reduces them for current facts, because the model can quote retrieved pages. It does not eliminate them: the model can still misread a page, mix two sources, or cite a link that does not say what the summary claims.
Are newer models free of hallucinations?
No current model is free of them. Error rates vary by model and by task, and they are usually highest for obscure topics, precise numbers, and recent events.
What is the difference between a hallucination and a confabulation?
They describe the same behavior. NIST's Generative AI Profile uses the term confabulation for confidently stated but false or erroneous content produced by a generative AI system.
Sources
- Why Language Models Hallucinate (arXiv)
- NIST AI 600-1: Artificial Intelligence Risk Management Framework, Generative AI Profile
- Introduction to Large Language Models (Google for Developers)
Research-based: Technorat does not claim hands-on testing. See how we research. Spotted an error? contact@technorat.com