Why Google’s AI can’t spell its own name: The fundamental flaw in large language models

BitcoinWorld

Why Google’s AI can’t spell its own name: The fundamental flaw in large language models

Google’s AI Overview, the generative search feature the company has positioned as the future of its flagship product, continues to struggle with a task most humans master by age six: spelling. Users recently discovered that when asked how many ‘P’s are in ‘Google,’ the AI confidently answered ‘two.’ It also stated there is ‘exactly 1 r in the word poop’ and claimed there are two ‘d’s in ‘journalism’ while spelling it as ‘j-o-u-r-n-a-d-i-s-m.’ When asked about the number of ‘P’s in the U.S. president’s last name, it correctly identified one, but then spelled the name as ‘t-r-p-u-m.’

The tokenization problem

These errors are not random glitches. They stem from a fundamental architectural limitation in the large language models (LLMs) that power AI Overview. Unlike humans, who perceive words as sequences of letters, LLMs break text into tokens. A token can be a full word, a syllable, or even a single character, depending on the model. The AI does not ‘read’ letters; it converts text into numerical representations that encode meaning and context. This token-based system is highly effective for generating coherent sentences and answering complex questions, but it has no inherent understanding of individual characters.

Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta, explained to Bitcoin World: ‘LLMs are based on this transformer architecture, which notably is not actually reading text. What happens when you input a prompt is that it’s translated into an encoding. When it sees the word ‘the,’ it has this one encoding of what ‘the’ means, but it does not know about ‘T,’ ‘H,’ ‘E.”

A known but unresolved challenge

Counting letters within words has been a well-documented weakness of LLMs for years. It has become something of a running joke in the AI community that whenever a new model is released, the first test should be asking how many ‘r’s are in ‘strawberry.’ Google acknowledged the issue in a statement to Bitcoin World, saying, ‘Counting within words has been a known challenge for LLMs, and we’re working to fix this particular issue.’

But fixing it may not be straightforward. Sheridan Feucht, a PhD student studying large language model interpretability at Northeastern University, told Bitcoin World: ‘It’s kind of hard to get around the question of what exactly a ‘word’ should be for a language model, and even if we got human experts to agree on a perfect token vocabulary, models would probably still find it useful to ‘chunk’ things even further. My guess would be that there’s no such thing as a perfect tokenizer due to this kind of fuzziness.’

Beyond spelling: Broader reliability concerns

The spelling errors are the latest in a series of high-profile missteps for Google’s AI Overview. Earlier in May, users discovered that searching the word ‘disregard’ would return what appeared to be a dictionary definition, but the definition read: ‘Understood. Let me know whenever you have a new prompt or question!’ — a clear sign that the AI was surfacing a system prompt intended for internal use. Google has since patched that issue.

These failures are not merely amusing. They underscore a deeper challenge for Google as it pushes generative AI to the center of its search experience. The company first introduced AI Overviews in 2024, only to have the feature famously advise users to eat rocks and put glue on their pizza, after citing satirical content from The Onion and Reddit. The current iteration, launched more broadly in early 2026, was supposed to be more refined. Yet basic spelling errors persist.

What this means for users

For the average person using Google Search, these errors are a reminder that AI, for all its impressive capabilities, is not infallible. The same models that can generate code, pass professional exams, and assist with creative writing can also fail at tasks a child can perform. The practical implication is clear: users cannot blindly trust AI-generated outputs without verification. This is especially important as Google integrates AI Overviews more deeply into search results, where users may be inclined to accept the information as authoritative.

Conclusion

Google’s spelling struggles are not a crisis for the company, but they are a useful reality check. They highlight that LLMs, despite their transformative potential, operate under fundamental constraints that researchers have not yet solved. The tokenization architecture that makes these models powerful also makes them blind to the building blocks of written language. As Google continues to double down on AI-driven search, these limitations will remain a source of both humor and caution for users and developers alike.

FAQs

Q1: Why can’t AI models like Google’s AI Overview spell correctly?
LLMs use tokenization, breaking text into tokens (words, syllables, or characters) rather than reading individual letters. This makes them highly effective at understanding context but unable to reliably count or identify specific characters within words.

Q2: Is Google working on fixing these spelling errors?
Yes. Google has acknowledged the issue and stated it is working on a fix. However, researchers note that the problem is inherent to the transformer architecture, and a perfect solution may not exist.

Q3: Should I trust information from Google’s AI Overview?
AI Overviews can be useful for general information, but users should verify critical facts, especially those involving numbers, spelling, or specific details. The AI can produce confident but incorrect answers, as demonstrated by the spelling errors and the ‘disregard’ incident.

This post Why Google’s AI can’t spell its own name: The fundamental flaw in large language models first appeared on BitcoinWorld.