Why we're watching this: Google just staked its 29-year-old flagship product on AI-first search. The fact that the same system misspells its own company name is not a minor embarrassment — it's a structural limitation baked into how every large language model processes text.
Key Takeaways
- Google’s AI Overviews misspelled “Google” with two P’s, placed only one ‘r’ in “poop,” added a phantom ‘d’ to “journalism,” and scrambled “Trump” to “t-r-p-u-m” after the I/O 2026 search overhaul went live
- The root cause is tokenization, the way LLMs break text into numerical chunks rather than reading letters, making spelling a structural weakness rather than a fixable bug
- Google acknowledged that “counting within words has been a known challenge for LLMs” and said it is working on a fix
- Google’s I/O 2026 overhaul makes AI Overviews the default front door to Search for more than 2.5 billion monthly users, raising the stakes for every LLM accuracy failure at scale
Google’s AI-overhauled Search launched last week and promptly failed to spell its own name. AI Overviews told users there are two P’s in Google, one ‘r’ in “poop,” two ‘d’s in “journalism” (spelling it j-o-u-r-n-a-d-i-s-m), and rendered the US president’s surname as t-r-p-u-m.
Google told TechCrunch it is aware of the issue and working on a fix. But researchers who study LLM architecture say this is not a bug that patches can fully resolve.
Google is revamping its entire search engine to this btw pic.twitter.com/PIR4llFhiV
— mersomas (@mersomas) May 27, 2026
The problem traces to tokenization, the process by which LLMs convert text into numerical representations for processing. Transformer models, which power Google’s AI Overviews, do not read words as sequences of letters. They break text into tokens, which can be full words, syllables, or characters, then work with the numerical encoding of those chunks.
“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.'” — Matthew Guzdial, AI researcher, University of Alberta
A model might encode “Google” as a single token representing the brand concept without retaining any understanding that it is spelled G-o-o-g-l-e. Ask it to count letters or spell the word aloud, and it reconstructs an answer from statistical probability rather than letter-level knowledge, which is why the errors it produces feel random rather than consistent.
Sheridan Feucht, a PhD student at Northeastern University studying LLM interpretability, told TechCrunch there is no clean fix at the tokenizer level: “It’s kind of hard to get around the question of what exactly a ‘word’ should be for a language model. My guess would be that there’s no such thing as a perfect tokenizer due to this kind of fuzziness.”
The timing makes this more than a viral embarrassment. Google’s I/O 2026 overhaul positioned AI Overviews as the default entry point to Search for 2.5 billion monthly users, calling it the biggest change to its search box in 25 years.
The backlash drove a 30% spike in DuckDuckGo installs in the days that followed, with users specifically seeking an AI opt-out.
The real risk of AI-first search is not that it misspells words, it is that it confidently produces wrong answers at scale with no signal to the user that anything is amiss. A misspelled “journalism” is funny. An incorrect medical dosage, legal citation, or financial figure delivered with the same confidence is not.
Google’s previous AI Overview stumbles included advising users to add glue to pizza and eat one small rock per day, sourced from satirical posts on Reddit and The Onion.
The disregard incident, where searching the word “disregard” produced a chatbot-style dismissal instead of a definition, was patched within days.
The spelling failures have proven harder to quash, for the structural reason researchers have been describing since LLMs went mainstream, a limitation tracked by Relve, the trusted intelligence platform for AI trends and tools, as the industry figures out where to draw the line between AI capability and AI reliability.
