Why Google's AI Broke Dictionary Lookups (And What That Reveals)

Google's AI search results are no longer showing proper dictionary definitions for certain common English words—ones like "disregard," "ignore," and "stop." Instead of pulling up meanings from sources like Merriam-Webster, the system treats these terms as commands directed at the AI itself, rather than words the user wants defined. It's an accidental consequence of Google's recent effort to protect its AI from being tricked by hidden instructions.

What's Happening With Dictionary Searches

Normally, when you search for a word, Google shows you a clean dictionary box at the top with the definition. That's no longer happening for certain terms. The AI system is misinterpreting them.

The problem stems from how Google is defending against "prompt injection"—a type of trick where someone tries to manipulate an AI system by sneaking commands into innocent-looking text. Terms like "disregard," "ignore," and "stop" are common in these attacks because they're used to tell the AI what to do.

Google's safety filters see those words and assume they're part of an attack, not a genuine search request. So instead of showing you a definition, the system blocks the request or changes how it responds. It's like a security guard at a building refusing to let in anyone carrying a clipboard, because some people have used clipboards as disguises before.

How Google Is Protecting Its AI

Google recently published details about its defense strategy against these kinds of tricks. The company's AI system, called Gemini, now scans incoming text to look for hidden commands. If it spots something suspicious, it generates what Google calls a "safe response" instead of proceeding normally.

Google is defending against two types of attacks. Direct attacks happen when you deliberately try to trick the AI through what you type. Indirect attacks are trickier: malicious instructions are hidden in web pages or emails that the AI reads while trying to help you. Google's security team is actively tracking these indirect attacks because they're becoming a real concern.

This defensive approach follows a pattern we have seen before in internet history. When search engines first became popular, people started flooding the web with spam pages designed to rank high in results. Google and other search companies had to build sophisticated spam filters. The challenge with AI is different but similar in spirit: instead of filtering out irrelevant pages, the system needs to spot instructions designed to trick it into behaving badly.

The Trade-off: Safety Versus Usability

The broken dictionary lookups show what happens when security measures are set very strictly. Google's filters appear to be calibrated to catch as many potential attacks as possible, which means some legitimate searches get caught in the net.

This creates a real problem for people just trying to look up what a word means. Before AI was involved in search, dictionary definitions came from structured sources and simple matching—the system just found the right word in a database. There was nothing clever about it, which made it reliable. Now that an AI is involved, the system has to decide whether your search term is a real question or an attack.

What Else Google Is Doing

Alongside its AI defenses, Google is suing SerpApi, a company it accuses of illegally scraping Google search results. This lawsuit is about controlling who can access Google's search data and how they use it—especially important now that AI companies are hungry for training data.

Google's own AI systems need access to web-scale information to work well. By restricting others' access to the same data, the company gains a potential advantage in building competitive AI systems.

Looking Ahead

The decision to prioritize security over perfect search results makes sense given the real risks of AI systems being manipulated. However, the dictionary problem shows how hard it is to block bad behavior without accidentally breaking things that should work.

Google's security team is watching to see which attacks actually happen in the wild versus which threats are theoretical. Over time, the company should be able to fine-tune its filters to catch real attacks while letting legitimate searches through. The challenge is getting that balance right while billions of people rely on Google Search every day.

This kind of trade-off—security versus convenience—is something every tech company will face as AI becomes more common. It's not obvious how to let people interact naturally with AI systems while also keeping them from being tricked by someone with bad intentions.