Glossary
An AI hallucination is an output that presents unsupported or incorrect information as though it were valid. In product-data work, this can mean an invented specification, an incorrect compatibility claim or a source citation that does not support the value. Fluent wording and apparent confidence do not establish that the output is true.
A model asked to fill a missing protection rating might choose a rating common among similar products. The answer may look reasonable and fit the expected format, but it is still unsupported for the exact item. A citation to a broad category page would not resolve that gap.
Other errors can be more subtle: combining specifications from two variants, using an obsolete datasheet or treating an optional accessory as a standard included component. These outputs can survive a basic format check because the values are plausible within the category.
Require a reliable source where the attribute needs factual evidence, and confirm that the source describes the right product and context. Separate extraction from inference. If no sufficient source is available, preserve the missing or unresolved state rather than making the record appear complete.
Retrieval, validation and review can reduce errors, but none is a universal guarantee. An AI model checking another model may repeat the same assumption. Use independent product evidence and representative evaluation cases, with review rules proportionate to the consequences of publishing an incorrect value.
Discuss the product information and controls your AI use case needs.