Google’s newly integrated generative AI Overviews in Search are having a hard time with something a first-grader can do: spelling and counting letters. As the company pushes mandatory AI Search features to millions of users, the system is consistently failing basic vocabulary tasks.
The AI claims there is one “r” in the word “poop,” thinks there are two “d”s in “journalism” while bizarrely spelling it “j-o-u-r-n-a-d-i-s-m,” and misspells the U.S. President’s last name as “t-r-p-u-m.” It even incorrectly counts two “p”s in the company name “Google.”
AI experts say this is an inherent flaw in how Large Language Models work. Matthew Guzdial, an AI researcher at the University of Alberta, explains that LLMs translate input into numerical encodings called tokens rather than reading traditional letters. The AI knows the encoding for a word, but it does not actually recognize individual alphabet characters.
Sheridan Feucht, a PhD student studying LLM interpretability at Northeastern University, pointed out that finding a perfect token vocabulary is almost impossible due to the fuzziness of how language is chunked. This structural blind spot directly causes the AI counting errors, and expert insights from researchers on token architecture flaws show why it is so difficult to fix.
Google formally acknowledged the issue in an official statement acknowledging the LLM limitations, confirming that counting letters within words is a known challenge. Researchers note that fixing this specific bug is not an urgent priority for developers compared to training the models to handle complex coding or math tasks.
How Basic AI Errors Are Pushing Users to DuckDuckGo
These visible spelling failures arrive shortly after Google’s AI Search feature generated nonsensical and dangerous instructions based on satirical web content. The compounding errors are actively shifting the search market.
Frustration over Google’s accuracy issues has directly fueled competitors. Privacy-focused search engine DuckDuckGo reported a 30% jump in installs as users actively flee the generative AI rollout. The tokenization problem means Google cannot simply patch the spelling errors overnight, leaving the door open for alternative search engines to capture frustrated users.