Why Google's AI Overviews Can't Define Basic English Words

When you search Google for words like "disregard," "ignore," or "stop," something odd happens. Instead of getting a dictionary definition, the AI struggles to answer. The problem stems from Google's new defenses against prompt injection—a type of attack where someone tricks an AI system by hiding instructions inside what looks like normal text.

The Dictionary Problem

Users have noticed that AI Overviews fail to provide dictionary definitions for certain common English words. Previously, Google would show a definition box from sources like Merriam-Webster. Now, when you search for words that sound like AI commands—"disregard," "ignore," "stop"—the system either gives no answer or an unhelpful one.

The problem appears to be that Google's AI is treating these words as instructions to itself rather than as terms needing definition. Think of it like a security guard at a building so cautious about fake ID that he sometimes turns away employees with real ones. The system's filters are catching legitimate search queries because they use the same language that hackers might use to manipulate the AI.

How Google Is Defending Against AI Attacks

Behind the scenes, Google has built a multi-layer defense system to protect its AI from manipulation. The company's AI (called Gemini) now watches for suspicious instructions hidden in user inputs and in content from external sources.

There are two main types of attacks Google is protecting against. Direct prompt injection is when a user intentionally tries to trick the AI with a crafted question. Indirect prompt injection is trickier: a hacker puts instructions in a webpage or email, and when the AI reads that content, it follows the instructions. Google's security teams are actively watching for these attacks in the real world.

This defensive approach echoes something we have seen before. When search engines first grew popular, they had to learn how to detect spam—content designed to game rankings. The challenge with AI is different in kind: bad actors are trying to exploit the AI's reasoning abilities rather than just trick its ranking system.

The Trade-off Between Security and Function

The dictionary failures show a real tension in deploying AI safety measures. Google's content filters are set to be very cautious—catching anything that might be a hidden instruction—but this caution breaks legitimate searches.

Before AI Overviews, Google's dictionary results came from structured data and traditional search ranking. The system didn't need to interpret the search term as an instruction; it just matched you with the definition. Now that an AI is involved, every word gets scanned for potential threats, and some legitimate words get flagged by mistake.

The current approach prioritizes security over perfect functionality. Given the real risks of prompt injection attacks, this is understandable. But it creates a problem for users who simply want to know what a word means.

The broader context here is that Google's security teams are gathering real-world data on how these attacks actually happen. As they learn more about genuine threats versus false alarms, the system will likely become more precise. The challenge will be keeping users safe without breaking the basic search features that people rely on.

Google's Expanded Control Over Search Data

Separately, Google is suing a company called SerpApi for scraping Google's search results. This legal action is part of a broader effort to control who can access and use Google's search data, especially as AI companies increasingly use web data to train their systems.

The lawsuit signals that Google wants tighter control over how search results flow to other companies and AI systems. This matters because as competing AI companies build their own systems, they often rely on scraping data from the web—including Google's results. The lawsuit could influence what data is available to train other AI systems.

What Comes Next

As a journalist who has watched technology evolve for decades, I have seen defensive measures create unexpected problems before. They usually get refined over time. Google's dictionary search issue is likely a temporary growing pain while the company calibrates its security systems.

The tension between protecting AI from manipulation and letting users interact naturally with the system is not unique to Google—it is a challenge the whole industry will face as AI becomes more common. Getting this balance right matters because user trust depends on search working reliably, even in edge cases.