Hallucination

A customer asks about the notice period, and the AI says three months. The answer sounds confident, is politely worded and is wrong. Answers like this are called hallucinations. For customer service, they are the most important risk of generative AI.

#Why language models hallucinate

A language model produces the most likely continuation of a text. It does not check whether a statement is true. When information is missing, it fills the gap with something that sounds plausible.

This is not a defect the next update will fix; it follows from how the technology works. That is why safeguards belong in how you deploy it, not in hoping for a better model.

#Typical forms

• Invented facts. A deadline, a fee or an opening time that does not exist.
• Invented sources. A reference to a page or document that does not exist.
• Wrong attribution. The terms of one tariff are attributed to another.
• Outdated statements. What used to apply is presented as current.

#How to limit the risk

What works is a combination of measures. None of them is enough on its own. Most of them work not on the model, but on the sources, the approvals and the question of who is accountable for an answer.

#Limiting hallucinations in Univerx

In Univerx, generative answers rely on your knowledge base in Univerx Knowledge. Agents accept suggestions from Univerx Assist deliberately, and AI agents act only within the scope you approve in Agent Studio. Our principles are set out under Responsible AI.