AI hallucination
When an AI model states something that isn't in the data it was given, such as a total it added up wrong, a rounded figure or a name it invented, in the same confident tone as a true answer.
In a business system the dangerous kind is a number that is nearly right. The way to rule it out is to keep the model away from the numbers, as described in how an AI assistant answers from your records; the way to catch it is a grounding check, which you can run yourself with the number check.
Talk to us about your systems.
Tell us what's held together with workarounds right now. We reply to every message ourselves.