Why Google’s AI models fail at simple word-level tasks

Google’s AI Overviews are facing public scrutiny for failing at simple tasks, such as counting letters in words or spelling common terms correctly. These errors highlight the persistent difficulty that large language models (LLMs)—artificial intelligence systems that power chatbots and text generators—have with character-level accuracy. For example, the AI asserted there are two Ps in the word “Google” and exactly one “r” in the word “poop.” It also claimed there are two d’s in the word “journalism” while spelling it “j-o-u-r-n-a-d-i-s-m.” While the AI correctly identified that there is one P in the last name of the U.S. president, it spelled the name as “t-r-p-u-m.” These errors occur as Google integrates generative AI into its 29-year-old flagship product. The company has previously had to address other system errors; last week, Google patched an issue regarding the word “disregard,” which had returned a default prompt response instead of a definition. In an emailed statement, Google acknowledged the spelling issues, stating that counting within words is a known challenge for LLMs and that they are working to fix this particular issue.

The root of the issue lies in how these models process information. Rather than reading text like a human, LLMs rely on a transformer—the architecture that powers LLMs by breaking text into tokens, which are numerical representations of text used by AI models. Because of this design, the AI does not perceive sentences as units of language made up of individual words and letters. “LLMs are based on this transformer architecture, which notably is not actually reading text. What happens when you input a prompt is that it’s translated into an encoding,” said Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta. Guzdial noted that when the model processes a word like “the,” it recognizes the single encoding for the entire word but does not inherently know about the individual letters “T,” “H,” and “E.”

These errors are difficult to resolve because of the fundamental way text is processed. Sheridan Feucht, a PhD student studying large language model interpretability at Northeastern University, shared that her guess would be that there is no such thing as a perfect tokenizer—the system that converts text into tokens—due to the inherent fuzziness of defining what a word should be for a language model. Feucht indicated that even if human experts agreed on a perfect token vocabulary, AI models would probably still find it useful to chunk things even further. Consequently, precise spelling and counting tasks remain a structural challenge for the current architecture of generative AI.

Why it matters

Google’s AI Overviews are struggling with basic spelling and counting tasks due to the token-based architecture of LLMs, highlighting the limitations of current generative AI models.