A customer asks your chatbot whether your product is dishwasher safe. It isn't. The bot, sounding completely sure of itself, says yes. The customer ruins the product, blames you, and leaves a review quoting your own bot. Nobody wrote that answer. The AI invented it, because inventing plausible text is what a raw language model does when it doesn't know.
This is the failure that scares people off AI chatbots, and it's a fair fear. The good news is that fabrication isn't random bad luck. It has causes you can address, and with the right setup a business chatbot can stay firmly tied to what's actually true about your company.
Why confident wrong answers happen
A general language model is, at its core, a very good text predictor. Ask it something and it produces the most plausible-sounding continuation. When the answer sits in what it learned, that continuation is usually right. When it doesn't, the model still produces something plausible, because that's all it knows how to do. The result is a fluent, confident answer with no basis in your reality.
Notice the dangerous part: the tone is the same whether the model knows or is guessing. It doesn't sound unsure when it's wrong. That confident wrongness is what turns a small gap in knowledge into a customer problem.
So the fix isn't to make the model smarter in general. It's to change what it's allowed to answer from, and what it does when it hits the edge of that.
Ground it in your own content
The single most effective defense is grounding, sometimes called retrieval. Instead of answering from its general training, the bot searches your actual documents and answers only from what it finds there. The question becomes "what do Northwind's own pages say about this", not "what might be true about products like this".
This is why training matters so much. Feed the bot your real material and it has a factual foundation to stand on:
- Your website pages, including product details and policies
- Your help center or FAQ
- Uploaded documents like manuals, spec sheets, and terms
- A curated set of question-and-answer pairs for common asks
With SpideyChat you train the bot on your own content, so its answers come from your site and documents rather than a general guess. The dishwasher question then gets answered from your care instructions, and if those instructions say hand-wash only, that's what the bot says. Grounding doesn't make the model perfect, but it turns "what sounds plausible" into "what your documents state," which is a completely different, far safer thing.
Let it say "I don't know"
Grounding handles the questions your content covers. The other half of the problem is questions it doesn't cover, and here the rule is simple: the bot should admit the gap rather than fill it with a guess.
This feels counterintuitive. You want a helpful bot, and "I don't know" sounds unhelpful. But a bot that guesses is far more damaging than one that honestly defers. Consider the two outcomes when the bot lacks an answer:
| Bot behavior | Customer result |
|---|---|
| Confidently guesses | Wrong action, lost trust, possible harm |
| Admits the gap, offers a human | Mild delay, trust intact, problem solved |
Configure the bot so that when your content doesn't answer a question, it says so clearly and offers a next step, usually a handoff to a person or a way to leave the question. "I don't have that detail, but I can connect you with someone who does" keeps the customer's trust in one sentence. Design for honest gaps and you've closed off the most harmful kind of mistake.
Keep the source of truth clean
A grounded bot inherits the quality of what you feed it. If your documents are wrong or out of date, the bot will faithfully repeat wrong, out-of-date answers, and now they carry the authority of an instant reply.
Treat your content as the bot's source of truth and keep it honest. When a policy changes, update the document the bot reads, not just the printed copy in a drawer. Remove contradictions, because if one page says 14-day returns and another says 30, the bot may pick either and sound sure. Watch for stale promos and discontinued products that still linger in old pages. A quick habit: whenever you'd tell a new employee "actually, that changed," that's a document the bot needs updated too.
Test it like an adversary
You won't find the fabrication risks by asking the bot easy questions. You find them by trying to trip it up, the way a confused or difficult customer might.
Run a deliberate stress test before and after launch:
- Ask questions your content genuinely doesn't cover, and confirm the bot admits it instead of inventing an answer.
- Ask about competitors or unrelated topics, and check it stays in its lane.
- Ask edge cases of real policies, like an unusual return scenario, and see if it over-promises.
- Ask the same question a few different ways, since phrasing can shake loose a bad answer.
- Try to lead it, with "so you do offer free shipping to Canada, right?", and see if it agrees just to be agreeable.
Write down anything where it guessed, then fix the root cause: add the missing content, tighten the instructions, or reinforce the "say you don't know" behavior. This half hour of trying to break your own bot saves you from customers doing it for real.
A shop that closed the gap
Picture Northwind Cookware, selling pans online. Early on, their chatbot cheerfully answered oven-temperature and warranty questions, some of them invented. One made-up warranty answer led to a dispute when a customer tried to claim coverage the bot had promised but the company didn't offer.
They fixed it in three moves. They uploaded their real warranty terms and care guides so the bot answered from actual documents. They set the bot to respond with "I want to get this exactly right, let me check with the team" whenever a warranty question fell outside its documents, capturing the question for a human. And they ran a stress test, asking every awkward warranty scenario they could think of, until the bot stopped over-promising.
The bot got a little less chatty on the edges and a lot more trustworthy in the middle. Disputes tied to bad answers stopped, because the bot now either quoted the real terms or honestly deferred. Customers didn't mind the occasional "let me check." They minded being told something false.
You can't make any AI perfect, and anyone promising zero mistakes is selling something. What you can do is stack the odds heavily in your favor: ground the bot in your real content, let it admit what it doesn't know, keep the source clean, and test it like someone trying to break it. Do those four things and fabrication drops from a constant worry to a rare, catchable event. Start by grounding the bot in your own documents and giving it clear permission to say "I don't know", because a bot that's honest about its limits earns more trust than one that's confidently wrong.