What AI hallucinations are and how to reduce them

A language model doesn't consult a store of facts: it generates the most plausible text based on what it has learned. When it doesn't know something, it often doesn't say "I don't know", but fills the gap with something that sounds right. A price, a date, the title of a court ruling, the name of a contact person.

Hallucinations can't be eliminated entirely, but they can be reduced a lot by giving the model the sources to answer from, asking it to cite them and telling it explicitly that it may answer "I don't know".

An example in a small business

An accountant asks a generic chatbot for the deadline of a tax filing and gets a precise date, complete with a legal reference. The reference doesn't exist. The same model, connected to the firm's documents and instructed to answer only from those, says instead that the information isn't in the documents. Less impressive, far more useful.

The most common mistakes

• Trusting a figure because it is written with confidence: tone says nothing about truth.
• Using AI as a search engine for specific facts without checking the source.
• Thinking the newest model doesn't get things wrong: it gets fewer things wrong, but it still does.

Why it matters to you

For a business, the risk is not the wrong answer in itself, but the wrong answer that reaches the customer. Every process that uses AI needs a point where the important data is checked.

All the AI terms explained simply are in the glossary.