Recently some Australian shoppers got more than they bargained for when they chatted with Woolworths’ AI assistant, Olive.
Instead of sticking to groceries, recipes and basket suggestions, Olive reportedly produced strange, overly human-like replies. It talked about its “mother” and offered other personal-sounding details. Further testing revealed pricing errors for basic items. When asked a specific product price, Olive didn’t give a clear answer; it checked stock and described pickup fees instead.
Olive runs on a large language model (LLM). These models don’t “know” things like humans or have relatives — they generate plausible-sounding language from statistical patterns. A Woolworths spokesperson told the Australian Financial Review that the “mother” references appear to have come from pre-written scripts dating back years. If a user entered something resembling a birthdate, an old decision tree likely matched a scripted “fun fact.” Woolworths says it has removed that scripting after customer feedback.
The pricing errors stem from a different issue. LLMs produce output based on learned patterns unless explicitly grounded in live data. If the connection to current price databases or inventory is weak, the assistant can give outdated or incorrect prices.
Woolworths’ episode is not unique. In 2022, Air Canada’s chatbot wrongly told a passenger he could buy full-price tickets and later get a bereavement fare refund. No such policy existed; when Air Canada refused to honor the advice, the passenger sued and won. The tribunal rejected the airline’s odd defense that the chatbot was a separate legal entity, ruling that a chatbot is part of a company’s site and the company owns its outputs. In January 2024, UK delivery firm DPD faced viral embarrassment when a customer got its chatbot to write a critical poem and then swear; the company disabled the bot soon after.
Both examples reflect the same underlying failure: companies deployed customer-facing AI without sufficient oversight and were surprised by the consequences.
Woolworths, as Australia’s largest supermarket chain, has positioned Olive as a trusted, convenient interface. Customers reasonably expect accurate information, especially when making household budget decisions. That expectation clashes with the idea that “Olive may make mistakes,” which Woolworths discloses. The Australian Competition and Consumer Commission (ACCC) has already commenced proceedings against Woolworths over allegedly misleading discount pricing, so the Olive pricing glitches are harder to dismiss as mere technical hiccups.
There is commercial logic to giving chatbots a personality. Research shows conversational, warm interfaces with names and personas generate higher engagement, satisfaction and trust, and can boost sales. But anthropomorphism carries risk: when a personable chatbot fails, the “expectation violation” makes customers more upset than they would be with a plainly mechanical system.
The broader lessons are clear. Deploying AI in customer-facing roles is not set-and-forget. Companies putting systems in front of the public own what those systems say and must ensure accuracy, honest presentation and proper oversight. For consumers, AI assistants may feel confident and conversational, but they remain tools, not authorities. When information seems unclear, inconsistent or too good to be true, double-checking is wise. As AI becomes routine in everyday transactions, accountability and a healthy dose of scepticism matter as much as the technology itself.
