Ask a generic AI chatbot about your refund policy and it might give a confident, well-written, completely fabricated answer — quoting a 60-day window you have never offered. The customer believes it. Now you have a support problem, a trust problem, and possibly a financial one, all created by the tool you installed to help.
This is hallucination, and it is the single biggest reason businesses hesitate to put AI in front of customers. The good news: it is a solvable engineering problem, and the solution has a name — grounding.
Why language models make things up
A large language model is, at its core, a very sophisticated predictor of the next word. It was trained to produce text that is plausible, not text that is true. Those are not the same thing. When you ask it a question, it generates the most likely-sounding answer based on patterns in its training data — and it has no built-in knowledge of your business, your pricing, or your policies.
Worse, most models are tuned to always give an answer. Faced with a question it cannot actually answer, a raw model rarely says ‘I don’t know.’ It fills the gap with something that reads correctly and is factually invented. Confidence and accuracy are completely decoupled.
Why a bigger model does not fix it
It is tempting to assume the next, larger model will simply stop hallucinating. Bigger models hallucinate less on general knowledge, but they still have no way of knowing your facts — the ones that live in your help centre, your contracts, and your product docs. No amount of general training teaches a model your exact cancellation terms. That knowledge has to be supplied at answer time.
Grounding: answers from your content, not the model’s imagination
Grounding — often implemented as retrieval-augmented generation, or RAG — flips the process around. Instead of asking the model to answer from memory, you make it answer from your documents. The flow looks like this:
- Ingest: your knowledge base — help articles, PDFs, policy docs, product pages — is broken into passages and indexed.
- Retrieve: when a customer asks a question, the system finds the passages most relevant to it.
- Constrain: the model is instructed to answer only using those retrieved passages, and to cite them.
- Refuse gracefully: if nothing relevant is found, the assistant says so and hands off to a human, instead of inventing an answer.
The difference in behaviour is night and day. A grounded assistant asked about your refund policy retrieves your actual policy and answers from it. Asked about something you have never documented, it declines rather than guesses.
Grounding is necessary, but design still matters
Grounding dramatically reduces hallucination, but it is not a magic switch. Three things determine whether it actually works:
- Knowledge base quality. The model can only be as accurate as the content you feed it. If your documentation is thin, outdated, or contradictory, grounding faithfully serves up thin, outdated, contradictory answers.
- Retrieval accuracy. If the system fetches the wrong passage, the model answers from the wrong context. Good retrieval is as important as a good model.
- Out-of-scope guardrails. The assistant needs an explicit, well-designed path for ‘I don’t know’ — a graceful handoff, not a shrug.
How to test whether your chatbot is actually grounded
Do not take a vendor’s word for it. Stress-test it before your customers do:
- Ask out-of-scope questions. Things you have deliberately never documented. A grounded bot declines; an ungrounded one invents.
- Ask adversarial questions. Try to lead it into agreeing to a policy you do not have. See if it holds the line.
- Ask about stale information. Update a document and confirm the answer changes to match.
- Check for sources. A trustworthy answer can point to where it came from.
The bottom line
An AI chatbot that occasionally invents facts is not a smaller version of a good chatbot — it is a liability wearing a helpful face. For any business putting AI in front of customers, grounding is not an optional feature; it is the difference between an assistant you can trust and one you have to apologise for.
This is the principle Monology is built on: chatbots answer strictly from your verified knowledge base, and a built-in knowledge-base health check scores your content for gaps, clarity, and retrieval quality — so you can find the weak spots before a customer does.
