Hallucination
In AI, a hallucination is an answer that sounds fluent and confident but is false, unsupported or invented. It can be a wrong date, a fabricated quotation, a function that does not exist or a source that was never published.
Last updated November 8, 2026
Readable on your devices only.
Your conversation is readable on your device, and on your other devices if you sync them. Nowhere else.
Unreadable to us.
Your message passes through without being written down; your chats are only ever stored encrypted.
The model doesn’t know you.
It gets the question, never your email, your account or your IP address.
Nothing is left along the way.
The reply comes back and is stored here, encrypted; even when you sync, your key never reaches us in the clear.
How it works
Hallucinations come from how language models work. A model learns to produce text that is statistically plausible given what came before; it does not look facts up, and it has no built-in sense of what it does not know. When the answer is missing from its training, or only weakly represented, the most plausible continuation is still a fluent sentence, so it fills the gap with something that reads well. Vague or leading questions, obscure topics, requests for exact figures, footnotes, case citations or phone numbers, and very long inputs make this more likely.
They take several forms. A model can invent a fact, attribute a quote to the wrong person, cite a study, a court ruling or a web address that does not exist, mix up two similar towns, drugs or companies, or contradict a document it was given. In code, a typical case is a call to a library function with a plausible name that the library does not have. The tone gives no clue, since a made-up answer is usually written as smoothly as a correct one.
Nothing removes the problem completely, but several habits reduce it. Give the model the source material and ask it to answer only from that. Ask for quotations or references you can open. Use a reasoning model for multi-step questions. Compare answers across two models, or have a second one review the first. Lower the randomness for factual lookups, and prefer a short, scoped question to a sprawling one. And treat anything with legal, medical, financial or safety weight as something to confirm with a reliable source. Researchers distinguish intrinsic cases, where the output contradicts the supplied material, from extrinsic ones, where it adds claims that cannot be confirmed either way; summarization, translation and question answering each show both kinds.
In NeverNo
The page about fact-checking AI answers describes how NeverNo’s cross-check works: a model from another family rereads the answer, quotes what looks wrong and proposes a fix, with two modes, one that rereads without the web and one that looks claims up on the web and cites sources. Claude Sonnet 5.5, Anthropic’s balanced model, is one of the anonymized models you can use for either side of that exchange; it has a context window of 200,000 tokens and an adjustable reasoning effort.
Anonymized means the maker, Anthropic, reads the text without knowing who you are, and its own rules apply. We do not claim that any model hallucinates more or less than another. The sound approach is the same for all of them: ask for sources, open them, and keep a human check on what matters.
Worth remembering
A hallucination is a fluent answer that is not true, produced because the model predicts plausible text.
It can be an invented fact, quotation, source or function, and it is rarely flagged by the tone.
Giving the model the source text and asking it to answer only from it lowers the risk.
A second opinion from another model or a source you can open is the most reliable check.
Warning signs: oddly precise numbers, quotations with no source, dead links, and cheerful agreement with a false premise slipped into your question.
Frequently asked questions
Why do AI models make things up?
Because they generate the most plausible next words rather than retrieving verified facts. When knowledge is thin, a plausible but wrong completion can still come out sounding confident.
Can hallucinations be eliminated?
Not at present. They can be reduced by supplying source documents, asking for references, using reasoning models and adding a review step, but important claims still deserve a human check.
How can I tell if an answer is invented?
Look for details that are suspiciously exact, such as page numbers, names and links, then try to find them. Ask the model how it knows, request a quotation, or put the same question to a different model and compare. Spreadsheets, citations managers and official registries make quick referees for dates, statutes, dosages and version numbers.
Ask your first question
No account and no card. Write in any of the 10 languages NeverNo speaks.
Open the chat