The word Google has one L. Ask the smartest AI models how many it has, and a lot of them will swear there are two. Grok said two. Claude said two. Google's own search AI said two, then spelled the word back as G-o-o-g-l-l-e. These are the same models that can write a working app in seconds, yet they cannot tell you how many times the letter L shows up in a six-letter word.
What the models say
It's not a fluke. TechCrunch ran the test in May 2026 and the answers were brutal. Ask how many P's are in "Google" and it answers two. Hand it the word journalism and it claimed two D's, then spelled the thing j-o-u-r-n-a-d-i-s-m.
It's not just Google's cheap model, either. Claude Opus 4.8, the newest release from Anthropic, gets it wrong too. Set to Max, the most capable mode it offers, it still answered that Google has two L's.
Worse, the replies aren't even stable. Run the same question ten times and you might get the right answer once, then a confident wrong one.
TechCrunch Why Google's AI Can't Spell Google
Why it keeps happening
Here's the root of it. These models don't read letters. They read tokens, which are chunks of text the model swallows whole. The letters G, o, o, g, l, e never register on their own.
Matthew Guzdial, an AI researcher at the University of Alberta, put it flat out. "When it sees the word the, it has this one encoding of what the means, but it does not know about T, H, E."
TechCrunch Why AI So Bad Spelling Because
The strawberry curse
None of this is new. Back in 2024 the internet figured out that chatbots couldn't count the R's in strawberry. Two years and several "smarter" model generations later, the basic flaw is still sitting right there.
Reasoning didn't rescue it either. Newer models that "think" before answering can sometimes split a word into single letters and count it right, but they still faceplant on anything dressed up like the strawberry riddle. A 2024 study found the errors line up with how often a word showed up in training data and how fiddly the counting gets.
Why Google looks worst
Why does Google faceplant harder than a chatbot you pay for? Money. That AI summary at the top of your search results runs on a small, cheap, fast model, because Google answers billions of queries a day. One April 2026 analysis found its AI Overviews were flat wrong about 10 percent of the time.
This is the same feature that, at launch in 2024, told people to eat rocks and pour glue on their pizza after it swallowed jokes from Reddit and The Onion. "Counting within words has been a known challenge for LLMs, and we're working to fix this particular issue," the company said.
Ars Technica Testing Suggests Google's AI Overviews Tell
The fix they skip
The maddening part is that a fix already exists. Hand the job to a tool. When a model writes a tiny program to count the letters, or spells the word out one character at a time so each letter becomes its own token, it gets the answer right basically every time. Google's bargain-bin Overview just doesn't bother.
Researchers aren't optimistic anyway. Sheridan Feucht, who studies how these models work at Northeastern University, told TechCrunch there's "no such thing as a perfect tokenizer." A system can ace a coding test and still flunk kindergarten spelling, so its confident tone tells you nothing about whether the answer is right. Check the simple stuff. Especially the simple stuff.
