Why Google’s New Search Agent ‘Strikes’ on Simple Queries Like “disregard”

At this year’s I/O conference Google announced a new AI‑driven "Agent" experience for Search, billed as the first major overhaul in 25 years and promising natural‑language queries, image and file uploads, and even the ability to build personal agents directly from the search box.

Shortly after the rollout, users discovered a striking bug: searching for the word disregard triggers the AI Overview to treat the term as a command. The result shows the AI replying, "Okay, I’ll disregard the previous prompt and start from scratch. How can I help you?" instead of the expected dictionary entry and traditional links.

The cause, as explained in the article, is the large model’s prompt injection problem. Everyday prompts that happen to contain verbs like “disregard” or “remember” are mistakenly interpreted as instructions to the system rather than ordinary search terms, leading the AI Overview to generate irrelevant responses.

For comparison, the same query on Microsoft’s Bing shows an AI summary that, while imperfect, still includes useful information and retains the standard list of blue links, highlighting the difference in how the two services handle such edge cases.

Users have begun sharing a simple workaround: appending -ai to the query (e.g., "disregard -ai") to suppress the AI Overview and revert to traditional search results, though this method is not guaranteed to work in every case.

Google’s spokesperson quickly acknowledged the issue, stating that the AI Overview was incorrectly interpreting action‑related queries and that a fix was already being rolled out. Subsequent testing reported that the problem had largely been resolved.

This incident underscores the challenges of integrating large language models into search engines, especially the need to guard against prompt‑injection errors that can turn a powerful assistant into an ineffective tool.