Google’s ongoing integration of generative AI into its flagship search engine has hit another series of embarrassing roadblocks. Users have reported that the company’s AI Overviews are struggling with fundamental literacy tasks, such as counting letters within words and basic spelling. Examples include the AI incorrectly identifying the number of letters in common words and misspelling prominent names, such as the U.S. president, despite the technology’s ability to perform complex tasks like coding or mathematical problem-solving.

These errors highlight a persistent technical limitation inherent in Large Language Models (LLMs). Unlike humans, these models do not process text by reading individual letters. Instead, they rely on a transformer architecture that breaks language down into numerical tokens. Because the AI perceives these tokens rather than the actual characters that compose a word, it lacks a native understanding of spelling or character counts. Experts note that this token-based approach is a fundamental design choice, making it difficult to ‘fix’ spelling without fundamentally altering how these models function.

While Google has acknowledged that counting within words is a known challenge and is working on improvements, the recurring nature of these glitches serves as a reminder of the limitations of generative AI. Previous incidents, ranging from incorrect dictionary definitions to bizarre advice, have underscored the risks of relying on AI for factual accuracy. As the industry continues to push AI to the forefront of search, these spelling mishaps serve as a cautionary tale for users to maintain a healthy level of skepticism and verify AI-generated information.