Google AI Overviews draws criticism after failing to spell basic words

Google's AI Overviews draw criticism again after failing to spell simple words like Google. The company has admitted the issue and said it is working to resolve it

Google’s AI Overviews feature is once again drawing criticism after users discovered that the system struggles with basic spelling and letter-counting tasks. According to a report by TechCrunch, Google’s AI-generated responses in Search results incorrectly claimed there are two “p” letters in the word “Google” and produced several other inaccurate spellings and letter counts in user queries.

 

The issue comes as Google continues expanding generative AI features across Search, despite earlier controversies surrounding incorrect and misleading AI-generated answers.

What is happening

Users recently shared examples of Google AI Overviews producing incorrect responses to simple spelling-related questions. In one case highlighted in the report, the AI reportedly claimed that the word “poop” contains exactly one “r.” In another example, it incorrectly spelled the word journalism as “journadism.”

 

 

The report also noted that the system identified one “p” in the surname of US President Donald Trump but incorrectly rearranged the spelling itself as “Trpum.”

 

Google has acknowledged the issue in a statement to TechCrunch, saying that counting within words has been a known challenge for LLMs and that the company is working on resolving the problem.

 

The errors have gained attention because they involve extremely basic language tasks, despite modern AI systems being capable of handling complex coding, reasoning, and research-related queries.

Previous AI Overview issues

This is not the first time Google’s AI-generated Search responses have faced criticism. Google recently patched an issue where searching for the word “disregard” produced what appeared to be a dictionary entry, but instead displayed the phrase: “Understood. Let me know whenever you have a new prompt or question!”

 

When the company first introduced AI Overviews, users discovered instances where the feature cited satirical sources such as The Onion and Reddit posts as factual information. Some responses also generated bizarre recommendations, including suggesting that users eat rocks or add glue to pizza recipes.

Why AI models struggle with spelling

While citing responses from researchers, TechCrunch explained that the issue largely stems from how large language models (LLMs) process text. Unlike humans, AI systems do not actually read words letter by letter. Instead, many models rely on transformer-based architectures that break text into smaller units called tokens.

 

These tokens may represent entire words, syllables, or partial word fragments rather than individual letters. The system then converts those tokens into numerical representations to predict likely responses based on patterns in training data.

 

Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta, told TechCrunch that AI models understand encoded representations of words instead of the actual characters themselves. Researchers also noted that tokenisation itself is an imperfect process, making precise spelling and letter-counting tasks unexpectedly difficult for modern AI systems despite their broader capabilities.