• News
  • Society
  • Google AI Overview makes mistake in counting letters in word Google and name of its own company
1449

Google AI Overview makes mistake in counting letters in word Google and name of its own company

Main points
  • Google's AI Overview system incorrectly determines the number of letters in words, including in its own company name.
  • Artificial intelligence researcher Matthew Guzdial noted that language models do not process words literally.
  • The recorded errors demonstrate the limitations of language models even in simple text tasks.
Mathematics, artificial intelligence. Collage: Ukrainian News
Mathematics, artificial intelligence. Collage: Ukrainian News

Google's AI Overview system incorrectly determines the number of letters in words, including in its own company name. Among the recorded examples are a statement about two letters "p" in the word Google and errors in the spelling of other words, including "journalism" and the surname of the US President.

This is reported by TechCrunch.

Thus, the system, among other errors, presented the spelling of the surname of the US President as "t-r-p-u-m".

Google explained that the difficulties with counting letters are related to the principle of operation of large language models. The company noted that such systems do not operate with individual letters, but with tokens that can correspond to words, syllables or parts of characters, and is currently working on fixing this.

ADVERTISING

After the launch of AI Overview, similar cases were already recorded in Google search. At that time, the system referred to satirical materials and, among other things, issued recommendations such as using stones or adding glue to pizza.

Matthew Guzdial, an artificial intelligence researcher at the University of Alberta, noted that language models do not process words literally and perceive them as integral units of meaning, and not as a set of individual letters.

Sheridan Feucht, a graduate student at Northeastern University, pointed out that determining the correct text division for such systems remains a difficult task and does not have a unique solution.

The material emphasizes that spelling and basic language operations are not the main function of such models, but the recorded errors demonstrate their limitations even in simple text tasks.

ADVERTISING

Recall that artificial intelligence regularly makes large and small errors in news texts, therefore it cannot be a reliable source of news. This is the conclusion reached by researchers from the European Broadcasting Union (EBU), who analyzed the responses of the most popular AI assistants in different languages.

Who we are: About us, Contacts. How we write news and our principles: Editorial code. We did our best. If you found this valuable – please support us.

To request a correction, please send an email.