🤖Why Google's AI Can't Spell 'Google': Tokens, Explained
TL;DR
Google's AI Overviews has been spotted misspelling its own brand name. The reason is not a bug — it's a clean demo of why LLMs that operate on tokens, not characters, will always have a soft spot on tasks that need letter-level reasoning.
Google's AI Overviews has been spotted misspelling its own brand name. The reason is not a bug — it's a clean demo of why LLMs that operate on tokens, not characters, will always have a soft spot on tasks that need letter-level reasoning.
Key Points
AI Overviews returned spellings like 'Gogle' and 'Goggle' when users asked it to spell the brand
Root cause: LLMs tokenize text into sub-word chunks, then operate on integer IDs, not letters
The model never 'sees' the letters G-O-O-G-L-E individually unless explicitly forced to
Same root cause as the classic 'how many Rs in strawberry' failures from 2024
Workaround for builders: add character-aware tools or spell-out prompts for letter-critical tasks
Why It Matters
Anyone shipping LLMs into spelling-, formatting-, or character-counting-sensitive workflows needs to know this is a structural limit, not a model-quality issue. No amount of scaling fixes it without architectural help.
Quick Facts
Frequently Asked Questions
Why does this matter?
Anyone shipping LLMs into spelling-, formatting-, or character-counting-sensitive workflows needs to know this is a structural limit, not a model-quality issue. No amount of scaling fixes it without architectural help.
What happened?
Google's AI Overviews has been spotted misspelling its own brand name. The reason is not a bug — it's a clean demo of why LLMs that operate on tokens, not characters, will always have a soft spot on tasks that need letter-level reasoning.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.