AI-generated information that is fabricated, factually incorrect, or unsupported by source documents.
In AI discussions, hallucination refers to generated information that is fabricated, factually wrong or unsupported by the evidence being used. An answer can sound fluent and specific while containing an invented detail. Confident presentation is not a dependable indicator of correctness.
Distinguish several problems during review. The model might invent a fact, misread a valid source or repeat inaccurate information from a source. These cases can require different repairs. A response that includes citations still needs checking because a linked page may not support the associated statement.
Accurate business information is useful, but consistent addresses or structured data do not eliminate all AI errors. Nor does every incorrect answer prove that conflicting listings caused it. Inspect the actual claim and available evidence before assigning a cause or promising a fix.
Takeaway: Verify consequential details before using AI output in customer-facing work. Treat unsupported names, prices, qualifications and availability as issues to investigate, not facts to polish and publish.
Example
An assistant describes a retailer as offering weekend installation even though the available service page only mentions delivery. That unsupported addition should be corrected against the actual offer. Adding a citation to the delivery page does not make installation a supported claim.
What to check
Ask for inspectable sources where relevant and check them directly. Keep uncertainty visible when evidence is missing. Review errors by type so that source maintenance, retrieval improvements and editorial checks address the right problem.
For more detail, see Enabling Large Language Models to Generate Text with Citations.