Why Google’s AI can’t spell Google (or anything else)

“Counting within words has been a known challenge for LLMs, and we’re working to fix this particular issue,” Google told TechCrunch in an emailed statement.

These basic spelling errors may seem familiar. LLMs, the kind of artificial intelligence that powers chatbots and other text-generators, are not built to understand spelling. It’s been a running joke for years that whenever a company unveils a new AI model, you should ask it how many ‘r’s are in the word strawberry. These AI models — which can code an app in seconds, or solve problems that have stumped mathematicians for decades — are about as good as a kindergartener at spelling.

Google’s AI overview woes reach beyond silly spelling mistakes though. Google already patched an issue from last week in which searching the word “disregard” would yield what looked like a dictionary definition of the word, only the definition was shown as, “Understood. Let me know whenever you have a new prompt or question!” But these spelling errors have remained amusing because they’re so difficult to quash.

As researchers have previously explained when we’ve asked about these spelling conundrums, AI doesn’t perceive sentences as units of language made up of words and letters. Many LLMs are built on transformers models, which break down text into tokens, which can be full words, syllables, or letters, depending on the model. Instead of “reading” like a human would, the AI converts the text into numerical representations of itself, which are then contextualized to help the AI come up with a logical response.

“LLMs are based on this transformer architecture, which notably is not actually reading text. What happens when you input a prompt is that it’s translated into an encoding,” Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta, told TechCrunch. “When it sees the word ‘the,’ it has this one encoding of what ‘the’ means, but it does not know about ‘T,’ ‘H,’ ‘E.’”

The token-based architecture that powers LLMs like Google’s AI overview is inherently limiting, and researchers haven’t been optimistic that they can 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, and even if we got human experts to agree on a perfect token vocabulary, models would probably still find it useful to ‘chunk’ things even further,” Sheridan Feucht, a PhD student studying large language model interpretability at Northeastern University, told TechCrunch. “My guess would be that there’s no such thing as a perfect tokenizer due to this kind of fuzziness.”

This isn’t necessarily an urgent problem on researchers’ minds, since the utility of LLMs doesn’t come in their capacity to spell. But these blatant failures help us remember that AI is not perfect, even if it may sometimes seem like an all-knowing power beyond our comprehension. We cannot blindly trust AI outputs without double-checking their accuracy.