Why Google's AI can't spell Google (or anything else)

Google's AI Overview states there are two Ps in the word Google. The feature also says there is exactly one r in the word poop. It further claims there are two ds in the word journalism while spelling the term j o u r n a d i s m. Google identified one P in the last name of the U.S. president yet spelled the name t r p u m. The company told TechCrunch that counting within words has been a known challenge for large language models. Google is working to fix this particular issue. Google patched an issue in which searching the word disregard yielded what looked like a dictionary definition. The definition text instead read Understood. Let me know whenever you have a new prompt or question. Large language models operate on transformer architectures that convert text into numerical encodings of tokens. Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta, told TechCrunch that the model does not know about individual letters such as T H and E when processing the word the. Sheridan Feucht, a PhD student at Northeastern University, told TechCrunch that there is no such thing as a perfect tokenizer. These spelling limitations stem from the way models chunk text rather than process it letter by letter. The utility of the models does not depend on their capacity to spell. Researchers have noted that the models can code applications in seconds or solve complex mathematical problems.