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

How many P’s are there in Google? According to Google there are two.

There’s also “exactly 1 ‘r’ in the word ‘poo,'” says Google’s AI Overview, as well as two ‘d’s in the word journalism, but still spelled: journalism. In any case, Google has determined that there is one P in the US president’s last name, but spelled it trpum.

You didn’t have to be a prophet to predict that Google’s AI-forward Search overhaul would end badly. We’ve done this before. The first time Google added AI summaries to Search, the feature ended up citing satirical posts from The Onion and Reddit advising people to eat rocks and put glue on their pizza.

This time, as Google doubles down on its promise to make generative AI the centerpiece of its 29-year-old flagship product, it’s not surprising that it’s stumbling.

“Counting within words is a well-known challenge for LLMs, and we are working to resolve this specific 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 when a company unveils a new AI model, you should ask how many ‘r’s are in the word strawberry. These AI models – which can code an app in seconds or solve problems that have plagued mathematicians for decades – are about as good at spelling as a toddler.

However, the problems with Google’s AI overview extend beyond silly spelling mistakes. Google already fixed an issue last week where a search for the word “ignore” returned something that looked like a dictionary definition of the word. Only the definition was displayed as: ‘Understood. Let me know if you have a new prompt or question!’ But these spelling mistakes have remained funny because they are so difficult to destroy.

As researchers previously explained when we asked about these spelling problems, AI does not view sentences as linguistic units made up of words and letters. Many LLMs are based on transformer models, which split text into tokens. These can be full words, syllables or letters depending on the model. Instead of ‘reading’ as 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 answer.

“LLMs are based on this transformer architecture, which notably doesn’t actually read text. What happens when you enter a prompt is it translates it into an encode,” 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 doesn’t know anything 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 quite difficult 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 symbolic vocabulary, models would probably still find it useful to ‘chunk’ things even further,” Sheridan Feucht, a PhD student who studies the interpretability of large language models at Northeastern University, told TechCrunch. “My guess would be that there is no such thing as a perfect tokenizer because of this kind of vagueness.”

This is not necessarily a pressing issue in the minds of researchers as the usefulness of LLMs is not in their ability to spell. But these glaring failures help us remember that AI is not perfect, even if it sometimes seems like an omniscient force beyond our understanding. We cannot blindly trust the output of AI without double-checking its accuracy.