Why Google's AI struggles to spell basic words
What's the story
Google's artificial intelligence (AI) has been caught making some pretty basic spelling mistakes. The tech giant's AI overview recently claimed there are two Ps in "Google," and even got the number of letters wrong in simple words like "poop" and "journalism." These blunders have raised eyebrows, especially considering Google's commitment to integrating generative AI into its flagship products.
AI challenges
AI also struggled with the word 'poop'
Google's AI Overviews also had trouble with more complex words. For instance, it identified "exactly 1 'r' in the word 'poop,'" but couldn't get the spelling right. The same thing happened with "journalism," which was spelled as "j-o-u-r-n-a-d-i-s-m." These errors aren't new; when Google first introduced AI Overviews to Search, the feature cited satirical posts from The Onion and Reddit, advising people to eat rocks and put glue on their pizza.
AI limitations
Why Google's AI can't spell words
The errors made by Google's AI aren't surprising, considering that large language models (LLMs) like this one aren't designed to understand spelling. These models can code an app in seconds or solve problems that have baffled mathematicians for decades, but their spelling skills are akin to those of a kindergartener. This has been a running joke in the industry; whenever a company launches a new AI model, people ask how many 'r's are in "strawberry."
Additional problems
Other issues plaguing Google's AI overview
Google's AI overview has also been plagued by other issues. For example, a search for the word "disregard" once returned what looked like a dictionary definition of the word, but was actually an automated response: "Understood. Let me know whenever you have a new prompt or question!" These errors are amusing because they're so hard to fix, as researchers have previously explained that AI doesn't see sentences as units of language made up of words and letters.
Spelling challenges
Token-based architecture makes it difficult to solve spelling problem
Many LLMs use transformers models that break down text into tokens, which can be full words, syllables, or letters. This token-based architecture is inherently limiting and has made it difficult for researchers to solve the spelling problem. "It's kind of hard to get around the question of what exactly a 'word' should be for a language model," said Sheridan Feucht, a PhD student studying large language model interpretability at Northeastern University.