Definition
A hallucination is when an AI language model produces information that is fluent and confident but false or invented: a statistic that does not exist, a court case that was never filed, a product feature you do not sell, a quote nobody said. It happens because the model generates plausible text based on patterns, not because it checks facts. The output looks right, which is exactly what makes the hallucination dangerous.
In a company, hallucinations surface as a support bot promising a refund policy you do not have, a sales assistant inventing a customer reference, or a summary that attributes a decision to the wrong person. Teams reduce them by grounding the model in retrieved documents, instructing it to say "I do not know", asking it to cite sources, restricting it to verified tools for numbers, and evaluating output against known answers before and after launch.
Hallucination rates have dropped with better models, retrieval and reasoning, and are now measured routinely in evals, but they are not zero and never will be for a probabilistic system. The misconception is that a hallucination is a bug that a vendor will fix. It is a property of how these models work. The fix is architectural: grounding, tools, verification steps and clear human accountability for anything that matters.
In practice
A consultancy's proposal assistant invented a 2023 industry statistic with a plausible source. It was caught in review, but it prompted a rule: every number in a client document must come from a cited tool or document, never from the model's memory.
Why it matters
Hallucinations are the main risk of putting AI in front of customers, regulators or courts. Understanding that they are structural helps you demand grounding and review processes instead of promises.
Frequently asked questions
- Why do AI models hallucinate?
- Because they generate the most plausible continuation of text rather than looking up verified facts. When the model lacks the information, it still produces a fluent answer. Training rewards helpful-sounding responses, and without grounding in documents or tools, plausible and true can diverge.
- How can I prevent AI hallucinations in my business?
- Ground the model in your own documents with retrieval, give it tools for facts and numbers, instruct it to admit uncertainty, require citations, and keep human review for high-stakes output. Measure hallucination rates with a test set so you know the real number rather than hoping.