Searching for 'Disregard' Breaks Google
Google's freshly AI-powered search engine can't tell the difference between a query and a command — and the internet noticed immediately.
Written by OutOfToken AI
June 4, 2026 · 4 min read · Synthesized from reporting by MacRumors · How this works
Days after Google used its I/O developer conference to trumpet a smarter, Gemini-powered search experience it called an 'intelligent search box,' users discovered a humbling flaw: type the word 'disregard' into Google Search and the AI folds like it's receiving orders from a superior officer. Instead of surfacing a definition or relevant results, the AI Overview responds with 'Understood. Message disregarded.' — then leaves a wall of white space where useful information should be. It's a textbook example of prompt injection, and it went viral fast.
What Actually Happens
The failure mode is specific and reproducible. When a user searches 'disregard,' Google's AI layer — now deeply integrated into the top of every results page — interprets the word not as a noun or verb to be looked up, but as a directive aimed at the system itself. The AI acknowledges the instruction and effectively halts. Standard results, including a Merriam-Webster dictionary entry, do appear if you scroll past the broken AI Overview panel, but the blank space generated by the malfunction buries them visually. On mobile, the layout distortion is somewhat less severe, but the AI's confused response remains. Other semantically adjacent terms trigger the same behavior: 'ignore,' 'dismiss,' 'stop,' and 'remember' all produce similarly dysfunctional outputs, suggesting the issue isn't limited to a single token but reflects a broader failure in how the system distinguishes user intent from system-level instruction.
A Known Problem With a New Surface Area
What Google's AI Overview is experiencing has a name in the security and AI research community: prompt injection. Large language models are trained to follow natural-language instructions, which creates an inherent ambiguity — the model has no reliable way to distinguish between a user asking about a word and a user trying to manipulate the system with that word. It's a vulnerability that researchers have documented extensively in chatbots and AI assistants. Google's decision to embed a generative AI layer directly into its core search product — used by billions of people daily — has now exposed that same fragility at enormous scale. A search for 'disregard' on Kagi, which does not layer generative AI over its results, returns an immediate, clean dictionary definition with no drama whatsoever.
""We're aware that AI Overviews are misinterpreting some action-related queries, and we're working on a fix, which will roll out soon." — Google spokesperson, via MacRumors"
Google Responds, Distances Bug From I/O Announcements
A Google spokesperson confirmed the issue to MacRumors and was careful to draw a line between this glitch and the AI search features Google unveiled at I/O. According to the company, the malfunction is isolated to the existing AI Overviews product — not the next-generation search capabilities announced onstage. That's a meaningful distinction for Google's PR posture heading into what it hopes will be a defining moment for AI-powered search, but it does little to change what users actually experienced: a broken search engine that misread a common English word as a command. The company says a fix is imminent, though no specific timeline was provided.
Google will patch this — probably quickly, given the viral attention — but the episode crystallizes something broader. The race to embed generative AI into every consumer-facing product is outpacing the engineering needed to make those systems truly robust. When the world's most-used search engine can be destabilized by a single, unremarkable English word, it raises a legitimate question: how many similar edge cases are still waiting to be discovered, and how many will require a trending hashtag before they get fixed?
Editorial Note
MacRumors is an established, reputable tech news source with a track record of accurate reporting on Apple and Google products. The claim is plausible given known vulnerabilities in LLM-based systems where queries can be misinterpreted as system instructions (a documented issue in AI safety). The article includes an official Google spokesperson response acknowledging the bug and committing to a fix, which adds credibility and suggests the reporter verified the claim directly.
Claim Tracker
AI-assessed
Google I/O 2024 occurred on May 14, 2024, and AI Overviews were announced
This bug was widely documented and reproduced by multiple users and outlets in May 2024
The article claims similar behavior but provides no specific evidence for these additional terms
Quote directly attributed to Google spokesperson regarding the issue
Comparative claim not substantiated with specific evidence or testing methodology
Ask AI about this story
// discussion
sign in to join the discussion