Google's AI Overview Struggles with Basic Spelling, Exposing Fundamental LLM Limitations
Key Takeaways
- ▸Google's AI Overview makes basic spelling errors: claiming Google has 2 Ps, misspelling 'journalism,' and mangling the president's name
- ▸LLMs use token-based transformers that convert text to numerical representations rather than reading individual letters like humans do
- ▸This spelling limitation is a known and persistent challenge across all LLMs, not unique to Google
Summary
Google's AI Overview feature in Search has been exposed making embarrassing spelling errors, including claiming there are two Ps in "Google," misspelling words like "journalism" as "j-o-u-r-n-a-d-i-s-m," and spelling the U.S. president's name as "t-r-p-u-m." The company acknowledged the issue to TechCrunch, stating: "Counting within words has been a known challenge for LLMs, and we're working to fix this particular issue."
The spelling failures stem from fundamental architectural limitations in how Large Language Models work. Unlike humans who read words letter-by-letter, LLMs use token-based transformer models that convert text into numerical representations. As AI researcher Matthew Guzdial explained, the model doesn't perceive individual letters—when it sees the word "the," it has one encoding for that entire word, but "does not know about 'T,' 'H,' 'E'." This is not a new problem; it has been a well-known challenge for LLMs across the industry for years.
Researchers remain pessimistic about solving the problem. Sheridan Feucht, a PhD student studying LLM interpretability at Northeastern University, told TechCrunch that there may be no "perfect tokenizer" due to the inherent fuzziness of language representation. While these errors may seem minor compared to other AI applications, they underscore that AI systems—despite their capabilities in coding, complex problem-solving, and text generation—remain fundamentally limited in tasks that humans find trivial.
- Researchers are skeptical these issues can be easily solved due to fundamental architecture constraints, not engineering fixes
