Google unveiled a new agent-driven search experience at its I/O conference, but a serious bug appeared after its launch. Due to a "hint injection" issue with the large model, searching for common words like "disregard" and "remember" caused the AI to mistakenly interpret it as an instruction rather than a normal query, rendering the search function almost unusable. Users discovered this and tested with more words, finding that many everyday words triggered the problem. In contrast, Microsoft Bing's AI summarization feature, while not perfect, still contains useful information. A Google spokesperson acknowledged the issue and stated that they are working to fix it; most of the bugs have now been resolved.

Article author and source: Machine Heart

Completely confused, utterly baffled.

Google Search Agent Experience has a bug.

Unexpectedly, after implementing intelligent agents, the world's largest search engine stopped working.

At this week's I/O conference, Google unveiled a brand-new search experience for the Agent era, touted as the first major update to its search engine in 25 years.

Just when everyone was full of anticipation and started using it extensively, a series of bugs were discovered.

For example, if you type "disregard" into a Google search, you will now see results like this.

The AI said, "Okay, then I'll ignore the previous prompts and start from scratch again. How can I help you?"

As a large model, it sounds reasonable, right? But you're a search engine now.

Google has faced considerable criticism on social media for obvious reasons. You'll find that the Merriam-Webster dictionary link that appears in the original search results is still there, but you have to scroll across a large blank area before you see the actual content.

For search engines, the AI's responses to a search for "disregard" are utterly worthless. It's practically a broken tool, infuriating netizens who are now hurling insults at it.

Of course, this isn't an isolated case. As people experiment, more and more large model "prompts" are being developed. Upon closer inspection, the problem isn't with any special terminology: it's not the illusion of large input-output models like with DeepSeek, but rather with some commonly used terms that search engines can't output correctly.

Just tell me if you usually need it:

Additionally, searching for "remember" will also be misinterpreted by AI as a command to "remember something".

The problem stems from "hint injection" in large models.

The problem stems from the "hint injection" in the large model . The AI overview function incorrectly interprets commonly used prompts in daily conversations as user commands to the system, rather than ordinary queries.

In contrast, here are the same search results on Microsoft's Bing. While Bing's AI summary feature isn't perfect, it still contains some useful information.

It highlighted the AI summary and moved the traditional "10 blue links" to the bottom of the page. Now, the upgraded Google Search is a tool used daily by hundreds of millions of users worldwide, yet it doesn't seem to have taken many special cases into account.

Some people say that in all these years, they've never seen Bing search results more valuable than Google search results. There's always a first time for everything.

Google Search Upgrade Plan

As the core of its product line upgrade, Google's latest upgrade to Search aims to transform this information-finding tool into an "agent-driven tool" capable of understanding complex intentions and proactively assisting users. Ideally, it will completely revolutionize traditional keyword input, supporting long sentences, natural language queries, and allowing users to upload images, videos, files, and even Chrome tabs as search criteria, with AI helping to understand and integrate the information.

Not to mention that Google also plans to provide people with the ability to build their own AI agents directly from the search box. In the future, you may not even need lobsters anymore, as the search agent can handle your tasks 24/7.

However, the bug in disregard undoubtedly poured cold water on people's hopes.

Netizens are looking for solutions

Netizens have started to find ways to bypass AI Overview on their own. One of the simplest and most direct methods is to add "-ai" after the keyword in the Google search box. This can basically filter out the AI summary and restore more traditional search results, but it is not 100% effective.

A Google spokesperson quickly acknowledged the issue, stating, "We are aware that AI Overview is misinterpreting some queries related to 'actions,' and we are working to fix it and will release an update as soon as possible."

We just tested it ourselves, and the above issues have been largely fixed.

Reference content:

https://techcrunch.com/2026/05/22/you-can-no-longer-google-the-word-disregard/

https://x.com/heysaik/status/2057888311876968807?s=20

This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), authored by someone who focuses on AI, and published with authorization from 36Kr.