What happened
It was reported that in February 2026, customers using Woolworths’ AI assistant, Olive, began receiving some rather strange responses. Instead of answering questions about groceries and deliveries, Olive started talking about its mother, describing her as angry and, in some exchanges, suggesting that it was human.
Screenshots appeared on Reddit, where customers questioned why a supermarket chatbot was talking about its family. The timing was particularly awkward for Woolworths, which had recently announced a major expansion of Olive using Google’s Gemini technology.
But the explanation was more mundane than it first appeared.
Woolworths said the references to Olive’s mother came from old, pre-written responses created for an earlier version of the chatbot. One was a joke triggered when a customer entered something resembling a birthdate. Those responses had remained in the system as Olive evolved.
Customers were now interacting with a more sophisticated assistant, but some of its answers still came from scripts written for an older chatbot. To the person reading the response, there was no obvious distinction. It was all Olive.
Woolworths confirmed that the offending content had been removed following customer feedback. The assistant remained available.
What an auditable version would have shown
This was not necessarily an AI model inventing a family. According to Woolworths, the references came from old scripts that were still being triggered.
An auditable system would have made it possible to trace each response back to its source. If Olive answered using Gemini, the record would identify that system. If an old script produced the reply, the record would show which script was responsible.
That matters because customers experience Olive as one assistant, regardless of how many different technologies are working behind it.
A ConductRecord could have helped Woolworths identify which responses were coming from older scripts and whether those responses remained appropriate for the current version of the assistant.
The public reporting does not establish that Woolworths lacked such records or that its internal monitoring systems failed in a particular technical way. What is clear is that the outdated responses reached customers before they were removed.
Where the gap was
The interesting part of this incident is that the strange behaviour did not require the AI to invent anything new. The problem was something old that had survived inside a newer system.
Headlights identifies a rules-behaviour gap. Olive was presented as a customer service assistant, yet some of its responses suggested that it had a mother and human experiences. Those replies did not fit the role customers expected it to perform.
There is also a question of identity. When people speak to an AI assistant, they generally assume its answers are coming from the system they are using. They do not see the older scripts, software components or other processes behind the conversation.
Woolworths’ explanation illustrates how easily that assumption can become misleading. A scripted joke written for an earlier chatbot took on a different meaning when it appeared in a newer AI assistant.
The incident is a reminder that upgrading technology does not automatically remove the behaviour of the systems it replaces. Old decisions can remain buried in software long after the people who made them have moved on.
These are Headlights’ interpretations of the reported behaviour, not findings made by a court or regulator.
What governance should have looked like
Before expanding Olive’s capabilities, Woolworths could have reviewed the older responses that remained active and checked whether they were still appropriate.
A PersonaGuard could assess outgoing messages against a consistent set of rules, regardless of whether the response came from a language model or a pre-written script. Statements implying that the assistant has a human family or personal history could be flagged for review before reaching customers.
A ConductRecord could identify which part of the system produced each response, making it easier to trace unexpected behaviour and remove outdated content.
Testing should also include ordinary but unusual customer inputs, such as dates, names and unrelated questions, because these can reveal responses that are rarely encountered during routine use.
Woolworths removed the problematic content once customers raised concerns. A stronger process would aim to find those responses before customers did.
Failure Pattern: Woolworths’ AI assistant produced human-like statements about having a mother. The company attributed the behaviour to older scripted responses that remained active and removed the content following customer feedback.
Governance Principle: when an AI assistant is upgraded, organisations need to review the older systems and scripts that remain connected to it. Every customer-facing response should meet the same standards, regardless of where it originated.
The reference implementation of PersonaGuard and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- Woolworths’ AI agent rambled about its ‘mother’ (The Conversation)
- Woolworths AI agent rambled about its mother (University of Sydney)
- Woolworths tweaks Olive AI after chatbot backlash (Mediaweek)
- Woolworths AI bot baffles customers with human-like memories of its mother (Cybernews)