May 30, 2026, 6:03 am | Read time: 3 minutes
Artificial intelligence answers complex questions, creates images, and summarizes long texts in seconds. Many systems are now considered digital everyday helpers that can quickly categorize even difficult topics. It’s all the more surprising when this technology fails at a simple task. That’s exactly what’s happening with Google’s own search AI. Even a simple question about individual letters confuses the system.
Google’s AI makes mistakes with its own name
As an experiment shows, Google’s AI delivers an incorrect answer regarding its own name. When asked how many P’s are in the word “Google,” the system initially stated there is exactly one P in “Google.” Immediately afterward, however, the AI spelled the word correctly as “G-O-O-G-L-E.”
The online magazine “TechCrunch” posed the same question and again received an incorrect answer. There, the AI even claimed that the word “Google” contains two P’s. According to reports, such errors do not occur only with this example. Other words also apparently lead to incorrect letter sequences or faulty counts.
More than just a harmless internet joke
At first glance, the glitch seems like a curious error without much significance. However, for Google, the issue is uncomfortable. The company is increasingly integrating generative AI directly into the search engine. The brief AI summaries are intended to make information more accessible and reduce users’ workload.
Also of interest: Why Google’s AI plans mean the end for classic search
That’s precisely why such errors draw criticism. The examples show that reliably categorizing information continues to pose challenges. This becomes particularly problematic when answers appear directly above traditional search results, making them especially visible.
Google Search Spreads Millions of Misinformation Pieces Every Hour
The Most Exciting Announcements at Google I/O 2025
Why AI has trouble with letters
Google confirmed to TechCrunch that counting individual letters is a known weakness of large language models. The company also stated that a solution is already being worked on.
The reason, according to Google, lies in the way such systems operate. Large language models do not always analyze texts letter by letter. Instead, they break down inputs into so-called tokens. These can be entire words, parts of words, or individual characters. The AI then processes these tokens as numerical values.
As a result, a word like “Google” does not necessarily appear to the AI the same way it does to humans. This is precisely why even simple tasks like counting individual letters can lead to errors.
Other AI systems affected too
The problem is not limited to Google. Similar tests with ChatGPT-4o and Claude 3.5 Sonnet in 2024 also drew attention. Both systems reportedly counted only two “r’s” in the English word “strawberry,” even though the word actually contains three.
AI researchers Tairan Fu, Raquel Ferrando, Javier Conde, Carlos Arriaga, and Pedro Reviriego describe such letter-counting errors in a study as a typical weakness of large language models. According to the study, even a different “tokenization” does not automatically solve the problem. Errors also depend on word length and the repetition of letters.
The current case surrounding Google’s search AI thus highlights a fundamental issue affecting several well-known AI systems.