Just days after Google debuted what it called its biggest product upgrade in a quarter-century, the company’s new AI Overviews feature is failing basic vocabulary tests. When users search for the single word “disregard,” the Gemini-powered system abandons dictionary definitions and instead treats the search query as a direct system command.
Instead of returning a standard Merriam-Webster entry, the engine generated chatbot-style apologies. Users saw responses reading exactly: “Understood. I have disregarded your previous prompt. How can I help you today?” The glitch wiped out standard formatting and triggered widespread complaints across Reddit as users struggled to disable the mandatory summaries.
This vulnerability highlights a core architectural flaw in blending standard web queries with Large Language Model interfaces. The AI responding to “disregard” as a system prompt exposes how easily conversational tools can overwrite factual retrieval. Other command-like words such as “stop” or “remember” triggered similar conversational loops.
The timing creates an immediate public relations problem for the search giant. Following the Google I/O rollout of Gemini, users faced UI layout glitches causing white space and began drawing direct comparisons to non-AI engines like Kagi. Frustrated users pointed out that alternative search engines parsed the word perfectly and provided the definition instantly without mistaking the query for an instruction.
Tech observers verified the AI Overviews feature temporarily breaking the dictionary function for command-based words. Google engineers are now forced to quickly patch these edge cases as millions of users test the limits of the new interface.
Why the ‘Prompt Injection’ Problem Threatens Google’s Core Product
Google marketed this intelligent search box as its most significant foundational change in 25 years. But this specific failure mechanism—accidental prompt injection—shows exactly why strapping an LLM to a traditional search bar carries extreme technical risk.
When a user types a word, the AI cannot differentiate between a query about the word and an instruction using the word. As competitors like Kagi capitalize on these unforced errors, Google has to figure out how to sandbox user queries. If basic vocabulary searches can accidentally rewrite the AI’s internal instructions, the reliability of the entire search ecosystem drops to zero.