Google AI Errors: Why the Neural Network Can't Spell 'Google' Correctly

The integration of generative AI into Google’s search system has exposed critical flaws in neural network performance. Despite excelling at complex tasks—from coding to mathematical calculations—LLM-based models display childlike literacy skills. The algorithm incorrectly counted letters in words like 'poop' and 'Trump,' while also mangling the spelling of 'journalism.'

Google representatives acknowledged the issue, admitting that character counting remains a known vulnerability in language models. However, experts doubt a quick fix is possible. The root problem lies in how LLMs process text: they break it into tokens—fragments that may represent words, syllables, or even single characters—before converting them into numerical vectors for response generation.

Researchers from the University of Alberta and Northeastern University emphasize that transformer architectures, the backbone of modern neural networks, lack any inherent understanding of word structure. Even with a flawless token dictionary, models will still fragment text into smaller segments to optimize processing. While these spelling errors don’t hinder AI’s ability to tackle complex tasks, they serve as a reminder for users to verify algorithm-generated results.

This isn’t Google’s first reputational stumble involving AI. Past incidents included search results suggesting users glue pizza or eat rocks, citing satirical sources. These new errors further fuel skepticism toward uncritical trust in automated systems, even when developed by tech giants.