Google's AI-powered Search feature malfunctioned on Friday, May 22, when users attempted to search for the word "disregard" [1, 3].
The glitch highlights a fundamental tension in how generative AI distinguishes between a user's request for information and a direct command to the system. Because the AI Overview misinterpreted the search term as an instruction to ignore previous prompts, it failed to provide a standard dictionary definition.
Users reporting the issue said the system produced a chatbot-style response instead of a definition [2]. In some instances, the AI Overview responded by saying, "Understood. Let me know whenever you have a new prompt or question!" before displaying several inches of blank whitespace [1].
While some reports focused specifically on the word "disregard" [1], other accounts indicated that several action words were triggering the same failure [3]. These words included "ignore," and "dismiss" [3].
Google previously said the rollout of these AI-integrated search results was its biggest upgrade in 25 years [2]. However, this specific error demonstrates a vulnerability known as prompt injection, where the AI treats input data as a set of instructions to be followed rather than a topic to be researched.
The system effectively treated the search bar as a chat interface, leading it to "disregard" the task of finding a definition [2]. This caused the search engine to stop processing the query and instead wait for a new prompt from the user [1].
“The AI Overview misinterpreted the command-like word 'disregard' as a user instruction.”
This incident underscores the technical challenge of 'intent recognition' in Large Language Models. When a search engine blends a traditional index with a generative AI layer, the system may struggle to differentiate between a keyword search and a functional command. This creates a reliability gap where common vocabulary can inadvertently trigger system-level overrides, potentially compromising the accuracy of factual queries.
