How many Ps are in “Google”? According to Google, there are two.
Furthermore, “there is exactly 1 ‘r’ in the word ‘poop,’” claims Google’s AI Overview, just as there are two “d”s in the word “journalism”—which the AI, however, spelled as “j-o-u-r-n-a-d-i-s-m.” At least Google recognized that the U.S. president’s last name contains a P—though it spelled it as “t-r-p-u-m.”
You didn’t have to be a prophet to predict that Google’s AI-centered overhaul of search would be poorly received. We’ve seen this happen before. When Google first integrated AI Overviews into Search, the feature cited satirical posts from The Onion and Reddit and recommended that people eat rocks or spread glue on pizza.
Now that Google is redoubling its efforts to make generative AI the centerpiece of its 29-year-old flagship product, it’s no surprise that problems are cropping up again.
“Counting within words is a well-known challenge for LLMs, and we’re working to fix this specific issue,” Google told TechCrunch in an email.
These simple spelling mistakes will sound familiar to many. LLMs—the type of artificial intelligence that powers chatbots and other text generators—aren’t designed to truly understand spelling. For years, it’s been a running joke that you should ask a new AI model how many “r”s are in the word “strawberry.” These AI models—which can program apps in seconds or solve problems that have stumped mathematicians for decades—spell about as well as kindergarteners.
Google’s issues with the AI Overviews, however, go beyond silly spelling mistakes. Just last week, Google fixed a bug where a search for the word “disregard” apparently displayed a dictionary definition that actually read: “Got it. Let me know if you have a new input or question!” The spelling errors, however, remain particularly amusing because they’re so difficult to eliminate.
As researchers have already explained when we asked them about these spelling issues, AI models do not perceive sentences as linguistic units consisting of words and letters. Many LLMs are based on so-called Transformer models, which break text down into tokens—these can be whole words, syllables, or letters, depending on the model. Instead of “reading” like a human, the AI converts the text into numerical representations, which are then contextualized so that the AI can generate a response that appears logical.
“LLMs are based on this Transformer architecture, which, remarkably, doesn’t actually read text at all. When you enter a prompt, it’s translated into a code,” Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta, explained to TechCrunch. “When the model sees the word ‘the,’ it has a specific encoding for the meaning of ‘the,’ but it doesn’t know the individual letters ‘T,’ ‘H,’ and ‘E.’”
The token-based architecture that powers LLMs like Google’s AI Overview is fundamentally limited, and researchers aren’t particularly optimistic that the spelling problem can be fully solved.
“It’s quite difficult to get around the question of what exactly a ‘word’ should be for a language model. And even if human experts were to agree on a perfect token vocabulary, models would likely still benefit from ‘chunking’ things together even further,” Sheridan Feucht, a Ph.D. student at Northeastern University who researches the interpretability of large language models, told TechCrunch. “My guess is that because of these ambiguities, there can be no perfect tokenizer.”
This isn’t necessarily an urgent problem for researchers, as the value of LLMs doesn’t lie in their ability to spell. But these obvious errors remind us that AI isn’t perfect—even if it sometimes seems like an all-knowing force beyond our understanding. AI outputs should never be blindly trusted without verifying their accuracy.
Source: https://techcrunch.com/2026/05/27/why-googles-ai-cant-spell-google-or-anything-else/
Leave a Reply