From Pizza Glue to Broken Basics
Google’s AI Overviews launched with infamous gaffes like advising people to drizzle glue on pizza, a painful but somewhat forgivable early stumble for a new technology. Two years on, the system is supposed to be more accurate and tightly integrated into everyday search, stepping in to summarize pages and answer factual queries directly. Yet users keep catching Google’s AI search making basic mistakes that traditional search rarely, if ever, made. Instead of just surfacing links or structured answers, AI Overviews now intervenes in dictionary lookups, spelling checks, and simple one-word queries—and often gets them wrong. The pattern suggests that Google has replaced stable, rule-based components of Search with a generative model that doesn’t always understand what users are asking. Rather than quietly enhancing results in the background, AI Overviews is now visibly rewriting some of the most mundane tasks people rely on Google Search to perform correctly.
Dictionary Lookups That Turn Into Chat Prompts
One of the clearest examples of Google AI Overviews errors is how it now handles dictionary lookup issues. For years, typing a single word like “disregard” or “ignore” into Google brought up a clean dictionary card at the top of the page, pulled from licensed lexicons. Today, AI Overviews has taken over that slot for many queries—and sometimes misfires completely. When users search for action words such as “disregard,” “ignore,” or “remember,” the system often responds as if it’s a chatbot being instructed, with messages like “Understood! I’ll ignore the previous prompt and start fresh,” instead of showing a definition. Even adding the word “definition” doesn’t reliably fix the problem. Google has acknowledged that AI Overviews is misinterpreting some action-related queries and says a fix is on the way, but it hasn’t detailed which terms are affected or how broadly the issue extends.
When Single-Word Commands Break Search Logic
Under the hood, these AI search failures reveal a deeper conflict between conversational AI and deterministic search features. Action words such as “ignore,” “dismiss,” and “remember” share a grammatical shape: they are short imperatives that look like commands in a chat interface. AI Overviews appears to treat these tokens as instructions by default, even when they appear as lone queries with no prior conversation. That means a request that used to trigger a simple database lookup now passes through a generative model predisposed to follow commands. Instead of a consistent, verifiable definition card, users see a freeform sentence that may vary from person to person. This shift erodes the predictability that dictionary lookups provided. A feature designed for billions of quick vocabulary checks has effectively become an open-ended generation step, highlighting how poorly the system distinguishes between “show me information” and “do what I say.”
Spelling Tests Expose Letter-Level Weaknesses
Spelling mistakes AI systems make in AI Overviews further underline the gap between language generation and precise text handling. The feature has already gone viral for bungling basic spelling questions, such as incorrectly answering how many “r” letters are in “strawberry.” More recently, users tested it with the word “astronomical.” When asked how many “e” letters are in the word, AI Overviews confidently replied that there are two, then provided a mangled spelling: “a-s-t-r-e-n-o-m-i-c-a-e-l.” Despite being powered by large language models trained on massive text corpora, the system struggles when queries require letter-by-letter accuracy instead of fluent sentences. Experts point to tokenization—models processing text in chunks rather than individual characters—as a root cause, but for users, the nuance doesn’t matter. A search tool that can’t reliably count letters in common words raises uncomfortable questions about using AI to answer even the most straightforward queries.
What These Missteps Reveal About AI Search
Taken together, misread dictionary lookups and botched spelling responses point to fundamental design tensions in AI-powered search. Google AI Overviews blurs the line between a traditional search engine and a conversational assistant but has not yet mastered the rules of either. It treats dictionary lookup queries as prompts, action words as directives, and simple spelling questions as generative text tasks rather than deterministic checks. This suggests that core parts of Search now depend on a model whose strengths—flexibility, fluency, context—clash with use cases that demand rigid correctness and consistent outputs. Google says a fix for the action-word bug is coming, yet the recurring failures show this is not just a one-off glitch. Until the system can reliably distinguish between requests for information and instructions to act, users will continue to see AI Overviews stumble on the kind of basic tasks that classic Google Search once handled flawlessly.