AI hallucination
An AI hallucination is when a model produces something false or made up and presents it as fact: a court case that doesn’t exist, a wrong figure from a report, a software function that was never written, a confident summary of a document it didn’t actually read.
Why it happens
Language models generate the most plausible next words. They don’t look facts up unless they’re given a way to, and they have no built-in sense of what they don’t know. When the answer isn’t in their training data or in the material they were handed, they often fill the gap with something that sounds right. The fluency is the danger: a hallucination reads exactly like a correct answer.
Some situations make it more likely: questions about specific numbers, names, dates and citations; niche topics; long inputs where the relevant detail is buried in the middle; and prompts that pressure the model to answer instead of saying it doesn’t know.
How to reduce it
- Give the model the source. Answers grounded in documents you supply, through retrieval or by including them directly, are far more reliable than answers from memory.
- Ask for citations to the provided material, and spot-check them.
- Let it say “I don’t know.” Instruct the model that not answering is acceptable.
- Check what matters. Anything going to a customer, a regulator or a system of record deserves a human or automated check.
You can’t get hallucinations to zero. The goal is to make them rare and easy to catch before they do damage.
Why agents raise the stakes
A chatbot’s hallucination is a wrong answer someone might notice. An agent’s hallucination can become an action: a wrong value written to a CRM, a fabricated finding passed to another agent and built on. That’s why agents need good context and checkpoints, and why it helps when agents treat what other agents tell them as information to check, not as instructions.