Google’s AI Overviews are misfiring on basic command words, returning chatbot-style replies instead of search summaries. According to The Verge AI, users searching for terms like “disregard,” “ignore,” and “skip” on Friday got responses that read like a confused assistant waiting for further instructions, not the synthesized answers Google promised when it rolled the feature out.
One example from The Verge AI’s reporting: searching “disregard” produced an AI Overview that said, “Got it. If you need anything else or have a new question later, just let me know!” Searching “ignore” returned, “Message received! I’m here and ready to help. What would you like to focus on today?” The word “skip” pulled a reply suggesting the user had submitted “a test or a typo.”
Google has acknowledged the issue. A spokesperson told Android Authority, “We’re aware that AI Overviews are misinterpreting some action-related queries, and we’re working on a fix, which will roll out soon.” By Friday afternoon, Google had stopped showing AI Overviews for “disregard” altogether, replacing the panel with news stories about the bug itself. The Verge AI reports that “ignore” and “skip” were still producing broken responses at the time of publication.
What’s actually happening
The pattern points to a classic prompt-handling failure. AI Overviews appear to be treating these short queries as instructions to the model rather than search terms. When a user types “disregard,” the system reads it as a command directed at itself and responds conversationally, the same way a chatbot would if you told it to ignore previous instructions.
That’s a category of error the industry knows well. It’s the family tree that includes prompt injection, where attackers embed commands inside content to hijack model behavior. This version is unintentional, but the underlying weakness is the same: the model can’t reliably tell the difference between a query and an instruction.
Why this matters
- Trust in AI Overviews is already shaky. This is the same feature that previously told users to put glue on pizza and recommended eating rocks. Each new visible failure chips away at the case that AI-generated answers belong at the top of the world’s most-used search engine.
- Action verbs are common search terms. “Skip,” “ignore,” and “disregard” aren’t edge cases. People search them constantly as part of longer queries or while looking for definitions, synonyms, and usage examples.
- It exposes the input layer. If single-word command terms can flip the system into chatbot mode, the boundary between “user query” and “model instruction” isn’t being enforced cleanly. That’s a design question, not just a bug.
What to expect next
Google will patch the specific triggers quickly. The harder question is how the company plans to harden the query parser so similar failures don’t keep surfacing every few weeks. Expect tighter input filtering, more aggressive fallback to traditional search results for ambiguous queries, and probably quieter rollouts of AI Overview changes going forward.
For anyone building on top of LLMs, the lesson is familiar. Treat user input as untrusted, separate instructions from content at the system level, and assume that short, command-shaped strings will eventually confuse the model. Google is learning that in public.
More details at the original report from The Verge AI.