"Hallucination" is the term for an AI stating something false with the same confidence as something true. It's the single biggest reason teams hesitate to put AI in front of customers — and understanding what causes it makes it much less scary.
Where hallucinations come from
A language model's job is to predict plausible next words, not to check facts against a source. Ask it something it has no real basis for, and it doesn't fail loudly — it fills the gap with something that sounds right. There's no built-in mechanism forcing it to say "I don't actually know this."
What "grounded" means in practice
A grounded answer is one the model generated from specific content it was handed for that question — a retrieved passage from your docs, not its general training data. This is the retrieval-augmented generation (RAG) pattern: find the relevant source material first, then answer from it. See our RAG explainer for the full picture.
Grounding doesn't make hallucination impossible — a model can still misread a retrieved passage — but it collapses the failure mode from "invented an answer from nothing" to "misinterpreted a real source," which is a much smaller and easier problem to catch.
How to actually tell the difference
- Ask where it came from. A grounded system can point back to the document or passage an answer was based on. A hallucinated answer has no real source to point to, even if the chatbot invents one.
- Watch what happens outside your docs. Ask a grounded chatbot something your documentation genuinely doesn't cover. If it says so instead of guessing, that's the grounding working as intended — see why "I don't know" is a feature, not a bug.
- Check if answers change when your docs change. A grounded chatbot's answers shift when you update the source document. One relying purely on general knowledge won't budge no matter what you edit.
Why this is the whole point
Customers don't need an AI that sounds confident — they need one that's right, or honest when it isn't. Grounding is the mechanism that makes that possible, and it's the difference between a chatbot you can put in front of customers and one you can't.
Related reading
- How chunking and embeddings work — the mechanism that makes grounding possible.
- Prompt injection and source-grounded bots — what happens when someone tries to break grounding on purpose.
- We tried to jailbreak our own chatbot — grounding put to the test for a week straight.
- 5 things everyone gets wrong about AI chatbots for support — why sounding human matters less than being honest.
- Tokens and context windows, explained without the jargon — the mechanical limits grounding has to work within.
- Anatomy of one chatbot conversation — grounding and honest fallbacks shown in a real exchange.