Why do AI chatbots make things up? The causes, and what reduces them

An AI chatbot makes things up because a language model writes the answer that sounds most likely, whether or not it is true. If your content lacks the answer, or search hands the model the wrong page, it fills the gap. Answering only from your content, saying “I don't know” and handing over to a person reduce it.

Below is what OpenAI, Anthropic and Google say about the problem in their own documentation, read on 13 October 2026, and the steps that help most in a chatbot that answers customers.

What is an AI hallucination?

OpenAI defines hallucinations as “plausible but false statements generated by language models” in a September 2025 research post. Google Cloud's explainer calls them “incorrect or misleading results” and notes they can extend to “fabricating links to web pages that never existed.” In a website chatbot they look ordinary: a returns window a week too long, a service you stopped offering, last year's price, a phone number with two digits swapped.

Why do AI chatbots make things up?

OpenAI's post gives two reasons. One is how models learn. In pretraining, a model predicts the next word across huge amounts of text, and there are no “true/false” labels attached to each statement. Regular patterns such as spelling come out right, but “arbitrary low-frequency facts”, like a pet's birthday, “cannot be predicted from patterns alone”. Your Saturday opening hours are that kind of fact.

The other is how models are scored. OpenAI compares it to a multiple-choice test: a wild guess might be right, while a blank scores zero, so models graded only on accuracy “are encouraged to guess rather than say ‘I don't know.’” Its own example from the SimpleQA test shows the trade. An older model, o4-mini, was right 24% of the time and wrong 75% of the time. A newer one, gpt-5-thinking-mini, was right 22% of the time and wrong 26%, because it declined to answer more than half the questions.

Where wrong answers come from in a business chatbot

A business chatbot should answer from your content, not from what the model remembers. That moves the problem without removing it. OpenAI's guide to optimizing accuracy names two places it breaks. In retrieval, “you can supply the wrong context, so the model can't possibly answer, or you can supply too much irrelevant context, which drowns out the real information and causes hallucinations.” And the model “can also get the right context and do the wrong thing with it.” In practice the causes look like this:

• The answer is not in your content. Nobody wrote down the holiday hours, so the model guesses.
• The content is out of date. An old price list is still online, or the chatbot last read your site months ago.
• The search found a nearby page. Asked about delivery to Canada, it pulls your UK delivery page and answers from that.
• General knowledge fills a gap. “Most shops accept returns within 30 days” quietly becomes your policy.
• A figure gets worked out. The model adds up a total, converts a currency or applies a discount by itself.

How to reduce AI chatbot hallucinations

None of the three companies' guides offers a single switch. Their advice, applied to a chatbot that talks to customers, comes down to six habits.

1. Answer only from your content. Anthropic's guide to reducing hallucinations lists external knowledge restriction: tell the model to use only the documents provided and not its general knowledge. Google's grounding overview, updated on 9 October 2026, says connecting a model to your data “tethers their output to these data and reduces the chances of inventing content.” Our explainer on RAG chatbots shows how that search through your content works.
2. Make “I don't know” an acceptable answer. Anthropic's first technique is to give the model explicit permission to admit uncertainty, which it says “can drastically reduce false information.” OpenAI's post says “errors are worse than abstentions.” Pair the “I don't know” with a way forward, such as passing the chat to a person.
3. Show sources. Anthropic suggests making a response “auditable” by citing quotes and sources for each claim. Google lists auditability among grounding's benefits, through links to sources. A visitor who can open the page an answer came from can check it, and so can you.
4. Keep prices and figures out of free text. Keep one current page with your prices and take old ones down. Tell the chatbot to quote a figure only when it appears word for word in your content, and leave anything calculated, such as a quote, a total or a discount, to a person or to your own system.
5. Hand off, then read what it missed. Questions it could not answer, and answers a visitor marked as unhelpful, are your list of content to write. Our post on when a chatbot should hand off to a person covers how a good hand-over works.
6. Test with real questions. OpenAI's accuracy guide treats a set of 20 or more questions with known correct answers as the baseline to measure against. Use questions your customers actually send, including the awkward ones, and run them again after every change to your content.

Can you stop a chatbot from hallucinating completely?

No. Anthropic's guide ends by saying its techniques “significantly reduce hallucinations” but “don't eliminate them entirely”, and asks you to validate critical information. OpenAI's post says accuracy “will never reach 100%” because some questions cannot be answered, yet hallucinations are not inevitable, “because language models can abstain when uncertain.” The realistic goal is a chatbot that is right when it answers and honest when it cannot, with a person behind it for the rest.

How MagizAI handles wrong answers

MagizAI's chatbot answers from your website pages, files and written answers. Its default instructions tell the model to answer only from that content, from tool results and from the conversation, and when the answer is not there, to say so and offer to connect the visitor with your team. A figures rule says every price, plan name, limit, percentage and date it states has to appear word for word in the content it was given. That rule is an instruction to the model, not a check in code, so one current price page still matters.

When the search finds little to support an answer, an extra instruction tells the model not to guess. A setting called “Answer only from trained knowledge” stops it falling back on the model's general knowledge; it is off for a new chatbot and on in the clinic template. Up to three sources appear under an answer, linked when the source is a web page, unless you switch that off. Questions answered with nothing behind them, and answers a visitor rated with a thumbs-down, go to a knowledge gaps list. A visitor who asks for a person moves to your team inbox with the whole conversation. The training guide and the handoff guide show where each part lives.