Why can't Google's AI spell its own name correctly?
The AI Overview feature in Google Search is making unexpected errors in counting letters in words. For example, the system claimed the word "Google" has two "p"s, and spelled "journalism" as "j-o-u-r-n-a-d-i-s-m," claiming it had two "d"s. Even when correctly identifying that the US president's surname has one "p," it spelled it "t-r-p-u-m," surprising users. This is reported by Techcrunch.com reports .
This problem is not new. Google has faced difficulties before in integrating generative AI into its 29-year-old flagship product. Previously, AI Overview cited satirical posts advising users to put glue on pizza or eat rocks. In a statement to TechCrunch, company representatives noted that counting letters within words is a specific challenge for LLMs (Large Language Models) and that they are working on it.
In fact, the AI models powering ChatGPT and Google Search were not created to understand spelling rules. Over the years, as new models have been introduced, asking them how many "r"s are in the word "strawberry" has become a kind of joke. These systems, which can solve complex mathematical problems or write code in seconds, remain at a preschool level when it comes to analyzing letters.
The reason lies in how AI perceives text. LLM models do not read text letter-by-letter like humans. They are based on a "transformer" architecture and break text into "tokens" (syllables, words, or groups of characters). AI converts text into digital codes and analyzes their contextual meaning.
As University of Alberta researcher Matthew Guzdial explained, when the system sees the word "the," it sees the code representing its meaning, but it does not know that it consists of the letters "T," "H," and "E." Therefore, even though Google Search is being completely updated, the AI within it still struggles to fix simple spelling errors.
