How many Ps are in Google? According to Google itself, there are two. There is also “exactly 1 ‘r’ in the word ‘poop’,” the company’s AI Overview states, along with two ‘d’s in the word journalism—yet it spelled it as j-o-u-r-n-a-d-i-s-m. At least the AI correctly identified that there is one P in the last name of the U.S. president, but it spelled that as t-r-p-u-m. You didn’t need to be a prophet to foresee that Google’s AI-heavy search overhaul would be poorly received. We’ve seen this before. The first time Google introduced AI Overviews 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. Now, as Google doubles down on making generative AI the centerpiece of its 29-year-old flagship product, it’s hardly surprising to see it stumble again.

These basic spelling errors may feel familiar. Large language models (LLMs)—the type of artificial intelligence that powers chatbots and other text generators—are not designed to understand spelling. It has 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 roughly as good at spelling as a kindergartener.

But Google’s AI overview troubles go beyond silly spelling mistakes. The company already patched an issue from last week where searching the word “disregard” would yield what looked like a dictionary definition, only the definition was shown as: “Understood. Let me know whenever you have a new prompt or question.” However, these spelling errors have remained amusing because they are so difficult to fix. As researchers have previously explained when we asked about these spelling conundrums, AI does not perceive sentences as units of language made up of words and letters. Many LLMs are built on transformer models, which break down text into tokens—these 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. 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 have not been optimistic that the spelling problem can be solved. “My guess would be that there’s no such thing as a perfect tokenizer due to this kind of fuzziness.”

This is not necessarily an urgent priority for researchers, since the utility of LLMs does not lie in their ability to spell. But these blatant failures serve as a reminder that AI is not perfect, even if it sometimes seems like an all-knowing power beyond our comprehension. We cannot blindly trust AI outputs without double-checking their accuracy.