Why Google’s AI can’t spell Google (or anything else)

Google's AI Overview, a new feature for its search engine, is consistently making basic spelling and grammatical errors. Examples include misspelling "journalism" and incorrectly counting the number of letters in words like "Google" and "poop." These glaring mistakes have drawn significant attention and concern.

This isn't the first time Google's AI search features have stumbled. Previous iterations cited satirical articles and provided unhelpful or even dangerous advice. The current errors highlight ongoing challenges as Google integrates generative AI more deeply into its core products.

Google acknowledges that counting within words has been a known challenge for large language models (LLMs), which power these AI features. These models are not built to understand spelling in the human sense. Instead, they break down text into "tokens"—numerical representations of words, syllables, or letters—rather than processing them as coherent linguistic units.

Researchers explain that LLMs use a transformer architecture where input prompts are translated into an encoding. This means the AI recognizes "the" as a single encoded unit but doesn't understand it as a combination of the letters "T," "H," and "E." This token-based approach, while powerful for many AI tasks, inherently limits their ability to handle spelling accurately.

While spelling accuracy isn't the primary utility of LLMs, these persistent errors serve as a critical reminder. They demonstrate that even advanced AI systems are not infallible and should not be blindly trusted. Users must continue to verify information and outputs generated by AI.