When AI Can’t Spell Its Own Name: What Google’s Latest Blunder Tells Us About the Limits of Artificial Intelligence
Cyprus Business Group | Technology & Innovation | May 2026
Artificial intelligence has rewritten the rules of business, communication, and productivity — but it turns out it still struggles to rewrite the alphabet. A wave of embarrassing errors from Google’s AI Overview feature has reignited an important conversation: just how much should we trust AI with even the most basic tasks?
Google’s AI Can’t Count Its Own Letters
The errors emerging from Google’s AI Overview have been, to put it mildly, spectacular. Ask the tool how many P’s are in the word “Google” and it confidently returns two. It also claims there is “exactly 1 ‘r’ in the word ‘poop'” and that the word “journalism” contains two D’s — going so far as to spell it out: j-o-u-r-n-a-d-i-s-m. Even the surname of the sitting U.S. president wasn’t safe, rendered as t-r-p-u-m.
These are not edge cases or obscure linguistic traps. They are basic spelling and letter-counting tasks that any primary school student could handle without hesitation. And yet one of the world’s most powerful and well-resourced AI systems is getting them consistently wrong.
A History of High-Profile Stumbles
This is not Google’s first encounter with AI-related embarrassment, and for those watching closely, it was hardly a surprise. When AI Overviews were first introduced to Google Search, the feature quickly made headlines for citing satirical content from The Onion and Reddit, producing advice that included eating rocks and applying glue to pizza.
Now, as Google pushes generative AI even further to the centre of its flagship search product — a platform that has been running for 29 years — the stumbles continue. Most recently, a separate incident saw the word “disregard” trigger what appeared to be a dictionary definition, only for the definition itself to read: “Understood. Let me know whenever you have a new prompt or question!” — an apparent system prompt leak masquerading as a search result.
Google has acknowledged the spelling issues, stating in an emailed comment to TechCrunch: “Counting within words has been a known challenge for LLMs, and we’re working to fix this particular issue.”
Why Does AI Struggle With Something So Simple?
The root cause of these errors lies in the fundamental architecture of large language models (LLMs) — the technology that powers tools like Google’s AI Overview, ChatGPT, and most modern AI assistants.
Unlike humans, who read words as sequences of individual letters, LLMs do not actually “read” text at all. They are built on transformer models that break language down into units called tokens. A token might be a full word, a syllable, or a single character — depending on the model — but the AI never processes language the way a person does. Instead, it converts text into numerical representations, which are then used to generate what it calculates to be the most statistically plausible response.
As AI researcher and assistant professor at the University of Alberta, Matthew Guzdial, explained to TechCrunch: “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.'”
In other words, the AI has no concept of the individual letters that make up a word. It knows what “journalism” means in context — but it has no mechanism to reliably count or reproduce the letters within it.
The Token Problem Has No Easy Fix
This architectural limitation has long been a point of discussion in AI research circles. The “strawberry test” — asking an AI model how many R’s appear in the word “strawberry” — has become an informal benchmark for demonstrating this very blind spot, reliably catching out new model releases from major labs.
The challenge runs deeper than a simple bug that can be patched. Sheridan Feucht, a PhD student researching large language model interpretability at Northeastern University, put it plainly: “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.”
In short, the problem is not a software glitch — it is a structural characteristic of how these models are built. Improving it would require rethinking the very foundations of how LLMs process language.
What This Means for Businesses Using AI
For businesses — whether in Cyprus or globally — the lesson here is a practical one. AI tools are genuinely transformative: they can generate code in seconds, assist with complex data analysis, draft communications, and support decision-making at scale. These capabilities are real and growing.
But they sit alongside genuine and sometimes surprising weaknesses. An AI that can help structure a business proposal may also silently misspell a word, miscount a figure, or confabulate a fact — with the same confident, authoritative tone it uses when it is correct.
As the source reporting on this issue noted, these failures serve as a valuable reminder that AI output is not infallible. We cannot blindly trust AI outputs without double-checking their accuracy. That principle should be embedded into any business workflow that incorporates generative AI — from marketing copy and financial summaries to customer-facing communications and research reports.
The Bottom Line
Google’s spelling errors are amusing on the surface. But they point to something more substantive: even the most advanced and well-funded AI systems have structural limitations that their developers openly acknowledge cannot yet be solved.
For business leaders, the takeaway is not to abandon AI tools — their utility is undeniable — but to deploy them with appropriate human oversight. In a business environment where accuracy, credibility, and trust are currency, the human review of AI output is not a redundancy. It is a necessity.
0 Comments