Google's AI Overview struggles with basic spelling tasks, revealing fundamental limits of large language models

Google's AI Overview has produced multiple spelling errors, including stating there are two Ps in 'Google,' two d's in 'journalism' (spelled j-o-u-r-n-a-d-i-s-m), and one r in 'poop.' Google told TechCrunch that 'counting within words has been a known challenge for LLMs, and we're working to fix this particular issue.'

LLMs are built on transformer models that break down text into tokens (linguistic units that can be full words, syllables, or letters) and convert text into numerical representations rather than reading words letter-by-letter like humans do. According to an AI researcher at the University of Alberta, when an LLM sees the word 'the,' it has one encoding of what 'the' means but does not know about the individual letters 'T,' 'H,' 'E.'

Researchers are skeptical the spelling problem can be solved. A PhD student studying LLM interpretability at Northeastern University stated that 'there's no such thing as a perfect tokenizer' and that models would probably still break things down further even with optimized token vocabularies.