A hallucination is when an AI model produces information that sounds plausible but is false or unsupported. It might invent a discount code, cite a policy that does not exist or describe a product feature the item lacks.
It happens because language models generate likely-sounding text rather than looking facts up. When the model lacks the relevant information, or the prompt invites a guess, it can still produce fluent output. Newer models and better techniques reduce the rate, but no current model is free of it.
In retail and hospitality, the risk shows up in customer-facing assistants: wrong allergen details, incorrect opening hours, promises of refunds or delivery dates you cannot honour. These can create legal exposure under Australian Consumer Law if a customer relies on a misleading statement made on your behalf.
Reduce the risk by grounding the assistant in approved content with retrieval, instructing it to say it does not know when unsure, keeping it away from topics like allergies and legal advice, and handing over to a person for anything high-stakes. Review conversation logs regularly and test with trick questions.
Recognising hallucination is the foundation of responsible AI use: treat output as a draft or a controlled answer, not as automatic truth.