Google has rolled out an overhaul that puts AI at the front of search, but it has been hit by a reliability dispute as basic spelling errors and information-processing problems have surfaced repeatedly.

On May 27, TechCrunch reported that Google AI Overview showed errors in basic text handling, including miscounting the number of P letters in the spelling of “Google” and incorrectly explaining how the word “poop” is spelled. It also misspelled “journalism” as “j-o-u-r-n-a-d-i-s-m” and misspelled the surname of a U.S. president as “t-r-p-u-m,” while confusing the number of P letters, revealing basic language-processing problems.

Errors also occurred in the process of searching words. Google AI Overview typically provides a dictionary meaning when a word is searched. But when “disregard” was searched, it displayed a chatbot-like sentence instead of a dictionary definition: "Understood. Please let me know if you have a new question."

Google explained that counting letters within words is a known challenge for LLMs and said it is currently working on fixes.

The problems are expanding beyond simple functional glitches into a broader dispute over the reliability of AI search features. Search services are based on the premise of providing accurate information, but as AI summaries intervene, if the consistency and reliability of answers waver, users could become more confused about how far they can trust the results, an observation has emerged.

As confusion appears even in basic information-provision functions, analysis of the cause is also continuing. Experts see the phenomenon as related to structural limitations of large language models (LLMs). LLMs do not read text at the character level like humans, but operate by breaking it into tokens and then processing it numerically. As a result, it is difficult for language models to clearly define word boundaries, and complete character-level processing is not easy no matter how tokens are constructed, analysis suggests.

Sheridan Feucht (셰리던 포이흐트), a doctoral student at Northeastern University, said: "There will not be a perfect tokenizer," pointing out that the model’s tendency to split text into smaller units is structurally inherent.

Some also see these limitations as not immediately leading to problems with overall LLM performance. That is because they still show high utility in other areas such as code generation and solving complex problems. Even so, the case again shows that no matter how much AI advances, its outputs should not be trusted unconditionally.

Google is revamping its entire search engine to this btw pic.twitter.com/PIR4llFhiV