Google's AI Overview Mangles Spelling and Letter Counts
Key Details
- Google's AI Overview feature in Search has produced incorrect information, including wrong spellings and letter counts.
- Examples include stating there are two 'Ps' in "Google" (correct), one 'r' in "poop" (correct), but spelling "journalism" as "j-o-u-r-n-a-d-i-s-m" and "Trump" as "t-r-p-u-m".
- Google acknowledged the issue, stating that counting letters is a known challenge for LLMs and that they are working on a fix.
"Counting within words has been a known challenge for LLMs, and we're working to fix this particular issue."
— Google spokesperson
Background
Large language models (LLMs), which power AI chatbots and text generators, break text into tokens rather than individual characters, making spelling errors common. Researchers note that LLMs do not "read" text but encode it numerically, leading to difficulties with spelling and letter counting.
Previous similar issues occurred when Google first added AI Overviews, such as recommending eating rocks or putting glue on pizza.
The Root Cause: Tokenization
"LLMs are based on transformer architecture, which notably is not actually reading text."
— Matthew Guzdial, AI researcher at University of Alberta
The fundamental issue lies in how these models process language. Instead of recognizing letters, they work with numerical representations of word chunks called tokens.
"There's no such thing as a perfect tokenizer due to this kind of fuzziness."
— Sheridan Feucht, PhD student at Northeastern University