How do you stop an AI assistant from making things up?
You cannot make it impossible, but you can reduce it. Give the assistant only information you have approved, tell it in writing what it may and may not answer, let it say "I don't know", hand the conversation to a person when it is unsure, test it on real customer questions before launch, and read actual conversations afterward. Treat each mistake you find as a reason to fix the information or the limits.
What the vendor guidance says
Anthropic's guide for developers who use its Claude model says even advanced models can sometimes produce text that is factually incorrect. It lists several techniques to reduce this, and says they do not eliminate it entirely. Its advice: "Always validate critical information, especially for high-stakes decisions".
The guide lists allowing the model to admit uncertainty, and says this "can drastically reduce false information". It also lists telling the model to use only the documents provided rather than its general knowledge, and running the same question several times and comparing, since inconsistent answers could indicate made-up content. The six things below mix ideas from the guide with general advice for owners.
Six things an owner can do
- Approved information only. Write down your services, hours, areas and policies in plain words, and ask for an assistant that answers from that and nothing else. Keep it current.
- Permission to say "I don't know". Decide what it says instead, such as offering a phone number or passing the question to a person.
- Written limits. List what it must not answer: prices not shown on your site, availability, quotes, refunds, and health, legal or money advice.
- Handoff to a person. It should pass the conversation, with a record, when it is unsure, when the customer asks for a person, or when the topic is on the limits list.
- Testing on real questions. Collect a few dozen from past emails, calls and messages, including ones whose honest answer is "we don't do that". Ask the same question several times, and rerun the set after every change.
- Reading conversations. Read a sample each week at first, note every wrong answer, and fix the information or the limit behind it.
Where to start
The first three decide what the assistant knows and may say. The last three catch what slips through. If you can only do a few, a good start is approved information, a written limit list and a handoff, then reading conversations often. The related answer on website chatbots gives a short version of the limits, handoff and weekly-reading steps for a chatbot on your site.
What none of this does
None of these steps makes wrong answers impossible, and a careful owner plans for that. Keep a person on anything a customer may hold you to. The related answer on whether AI assistants get facts wrong reports a 2025 study of assistants' answers to news questions, and the related answer on what to ask before hiring someone to build AI includes testing on your real examples and who checks output.
Sources
Related answers
- Do AI assistants get facts about businesses wrong?
- Should a small business use a chatbot on its website?
- What should you ask before hiring someone to build AI for your business?
AI assistants and chatbots
A good assistant answers the question a customer asks at nine at night, in your words, from your information, and brings in a person when it should. We build that, and we test it on your real questions before a customer sees it.