What happens when you ask an AI to spell the name of the company that built it? For Google's Gemini, the answer is: sometimes correctly, sometimes not. TechCrunch examined this specifically and the explanation isn't a bug - it's a design consequence.
How AI Actually Reads Text
Large language models don't read text the way humans do. They break words into chunks called "tokens" - often whole words, sometimes partial words ("Google" might be a single token, or it might split into "Goo" + "gle"). The model learns patterns between these chunks, but it doesn't process individual letters in sequence.
This means asking an AI to count letters, spell a word character by character, or identify rhymes is asking it to do something it was never built for. It has to infer the answer from statistical patterns in its training data rather than actually reading the characters. Sometimes those patterns give the right answer. Often they don't.
Ask Gemini how many letters are in "strawberry" and it might get it right or it might not - models frequently miss repeated letters like the double "r". Ask it to spell "necessary" character by character and transposed letters are common. These aren't signs the model is deteriorating. They're evidence that character-level reasoning was never what these systems were optimized for.
This Isn't a Google Problem Alone
Every major language model - ChatGPT, Claude, and the rest - has the same structural limitation. The tokenization approach that makes LLMs powerful for reasoning, summarization, and code generation makes them unreliable at tasks a primary school student handles easily. Google just happens to be a particularly pointed example because its own name falls into the category of words that trip up its AI.
For people using AI in their daily work: if you're relying on an AI assistant to proofread character-level details - specific spellings, letter counts for character-limited copy, or precise word constructions - verify those yourself. AI writing tools are genuinely strong at sentence structure, tone, and argument flow. They're far weaker on anything requiring left-to-right character awareness.
The irony of Google's AI struggling to spell "Google" is funny. It's also a precise illustration of why understanding what these tools actually can and can't do matters more than any benchmark score.