Why Google’s AI can’t spell Google (or anything else)
Google’s AI Overviews are still making basic spelling and counting mistakes, highlighting how weak LLMs remain at exact text tasks.
Intelligence analysis by GPT-5.4 Mini
TechCrunch says Google’s AI search features are once again stumbling over simple language tasks, including misspellings, counting letters, and odd output in place of definitions. The piece argues these errors are a reminder that LLMs process text as tokens, not as humans read words.
Google’s AI is like a very fast kid who can talk smoothly but still trips over easy spelling questions. It can sound smart without always being right.
The reason is that it does not look at words the way people do. It breaks them into pieces, like building blocks, and sometimes those blocks do not line up well for counting letters.
The article says this is a reminder to double-check AI answers. A machine can be confident and still make simple mistakes.
Analysis
What happened
TechCrunch reports that Google’s AI Overviews are producing embarrassing mistakes in Search, including miscounting letters in words like "Google," misspelling common words, and mishandling a search for the word "disregard." The article frames this as another sign that Google’s AI-first search overhaul is still rough around the edges.
Why the errors happen
The story explains that these failures are not surprising to researchers. Large language models do not "read" text the way people do. Instead, they work on tokenized representations of text, which may split words into chunks, syllables, or letters. That design helps models generate fluent responses, but it makes precise letter counting and spelling much harder than it is for a human.
Google told TechCrunch that counting within words has been a known challenge and that the company is working to fix this specific issue. The article also notes that this is not the first time AI Overviews have created bad answers. Earlier versions of the feature were criticized for citing satire and low-quality sources, and for giving harmful advice such as eating rocks or putting glue on pizza.
Broader takeaway
The article’s main point is not that AI is useless, but that it is still far from trustworthy on basic factual or text-precision tasks. It can code, summarize, and reason in ways that look impressive, yet still fail at something a child could do. That gap is exactly why the piece says AI outputs still need to be checked before people trust them.
Key points
- Google’s AI Overviews are making obvious spelling and counting mistakes in Search.
- The article says this is a known weakness of large language models, not a one-off glitch.
- Google told TechCrunch it is working to fix the letter-counting problem.
- The piece ties the issue to a broader warning: fluent AI output can still be wrong.
- It also recalls earlier AI Overview failures involving bad citations and unsafe advice.
