AI Hallucination
Last updated:
What an AI hallucination is
The term describes output that is not supported by the model’s input or by reality: a wrong date, a feature that does not exist, a quote nobody said, a link that leads nowhere. What makes hallucinations risky is their tone. The model writes a wrong answer in the same fluent, confident style as a correct one.
Why language models hallucinate
Several causes usually work together:
- Next-token prediction: a model is trained to produce plausible text, not to check facts. When information is missing, a plausible guess is often what comes out.
- Gaps in training data: the model may never have seen your business, or it learned information that is now outdated.
- Missed retrieval: in a chatbot, if the search step does not find the right passage, the model may fill the gap from general knowledge.
- Leading questions: a question that assumes something false, such as a discount that does not exist, can pull the model along.
Example
A visitor asks a hotel chatbot: “Is breakfast included in the weekend package?” The knowledge base describes the package but says nothing about breakfast. A poorly configured chatbot replies “Yes, a breakfast buffet is included”, because many weekend packages include one. A grounded chatbot says the package description does not mention breakfast and suggests asking the hotel directly.
Why hallucinations matter for business chatbots
On a business website, an invented price, return window or opening time is not a harmless slip. Customers act on it. In a 2024 case, a Canadian tribunal held Air Canada liable for incorrect fare information its website chatbot had given a customer. Hallucinations cannot be eliminated completely, but they can be reduced:
- Answer from retrieved company content instead of the model’s memory.
- Instruct the model to say when its sources do not contain the answer.
- Keep sources current and remove contradictory or outdated pages.
- Add exact Q&A pairs for questions where the wording matters.
- Test with questions the content does not answer, and review chat logs after launch.
How intoCHAT reduces hallucinations
For every message, intoCHAT retrieves relevant excerpts from your content with hybrid search and passes them to the model with built-in rules. Business facts such as prices, policies and contact details must come from your configuration, your sources or a successful action result. When the evidence is missing or conflicting, the agent says what it cannot confirm instead of guessing. Q&A answers take priority over other sources, and you can test the agent in the playground and read every conversation in the chat logs.