SANIK Blog
Does an AI chatbot make things up? How to prevent hallucinations
The key principles for the safety of AI answers: the knowledge base, limits, handover to a human and tests before deployment.
1. When AI starts making things up
Most often when there is no clear data or when we ask it about a topic that is not in the knowledge base. Without safeguards, the model tries to be helpful at all costs.
2. How to limit the risk
The basis is a closed knowledge base and a rule: answer only from the facts available. When data is missing, the bot should say plainly that it has no information and offer contact with a human.
3. The testing process before launch
Before launch it is worth going through a set of control questions: prices, dates, policies, tricky questions and questions outside the scope. Every wrong answer goes back for a fix in the prompts or the data sources.
4. Safe escalation to a human
The bot should be able to hand over the conversation when a sensitive or non-standard topic appears. This is not a weakness of the system, but a mechanism that protects the company's reputation.