Glossary · AI & automation
AI hallucination meaning
An AI hallucination is a confident answer from an AI model that is false or unsupported. Why language models hallucinate, common types, and how to catch them.
Why does AI hallucinate?
A large language model predicts the next words that fit the pattern of the conversation. When it has the fact, the likely words are usually right. When it does not, the likely words are still fluent, so it fills the gap with something that sounds right. It has no built-in sense of "I don't know" unless it is designed and instructed to have one. NIST's generative AI profile calls this "confabulation", which is a fairer word: the model is not seeing things, it is filling gaps.
What are the common types?
| Type | Example |
|---|---|
| Invented fact | A product feature the brand does not offer |
| Wrong number | A price, date or quantity that is close but wrong |
| Fake source | A citation, law section or link that does not exist |
| Unsupported claim | "Best-selling in India" with nothing behind it |
| Misread source | The document says one thing, the summary says another |
A worked example in marketing content (illustrative)
A spice brand asks an AI tool for a Diwali post about its garam masala. The draft says "no added salt, loved by 10,000 families, now in 500 stores". The pack lists salt in the ingredients, the brand has never counted families and it sells in far fewer stores. Every one of those phrases is a hallucination, and the first could also breach labelling and advertising rules. The fix is to generate from the product's own catalogue data and have a person approve each post. That is how an AI social media content generator should work: brief from the catalogue, human edit before publishing. All details here are illustrative.
How do you reduce hallucinations?
- Ground the model. Give it the facts (catalogue, price list, policy) and tell it to use only those.
- Ban categories of claim. No statistics, awards, certifications or comparisons unless supplied.
- Ask for citations on anything factual, and check that they exist.
- Allow "I don't know". A callback or "not in the file" beats a guess.
- Keep a person on the approve button for anything public, financial or legal. See what to check before an AI-made post goes live.
What are the common confusions?
- Hallucination vs bias. Bias is a skewed pattern across many answers; a hallucination is a specific false statement.
- Hallucination vs creativity. Invented wording in a caption is fine; invented facts in it are not.
- "Newer models don't hallucinate." They do so less often, not never.
A clear brand voice guide also helps, because it tells the model what it may and may not say.
Related terms
Read next
Where this shows up in our work. Content AI: Posts, carousels, festival creatives and videos from your catalogue, in your voice — scheduled and published everywhere. Book a 30-minute working session with an engineer.
Questions
What is an AI hallucination in simple words?
It is when an AI tool states something false or made up as if it were true, such as a wrong price, a fake statistic or a source that does not exist, usually in a confident and fluent way.
Why do chatbots make things up?
Language models generate the most likely-sounding words rather than look up facts. When they lack the right information, they fill the gap with plausible text unless they are tied to trusted sources and allowed to say they do not know.
Can AI hallucinations be completely eliminated?
Not with current models. Grounding answers in your own data, banning unsupported claims, requiring citations and having a person approve important outputs reduce them greatly, but checks are still needed.