Human text processing differs significantly from AI. Humans naturally understand text at the letter level, while AI models like ChatGPT work with tokenized units.
What does that even mean? Is your transmogrifier conflicting with the transflux capacitor?
"How many (letters) in the word (word)?" is an extremely simple question.
Limits the field of consideration to a hard, finite and small number of options.
If they were truly interested in "teaching" AI, they would allow humans to say "NO. WRONG." and the AI would count the number of people that are disagreeing with it, and at least say "I don't know". This fundamental failure means it's not "Artificial Intelligence", it's "Artificial Stupidity".
"Tokenized Units" translates to "Too stupid to learn from humans", AFAIC.
Another response might be "I don't know. How many do you think there are?" and go with the most consistent responses to that question. If 5 out of 100 people tell the AI that there are 3 "R's" in "Strawberry", and the other 95 answers have less than 5 consistent answers to anything else, at minimum the potentially wrong answer is at least reflective of the consensus of the humans. Google search results run in this way, and one assumes that much-vaunted "AI" is an improvement on that.
FYI I completely disbelieve anything anyone says about "AI", as I think it's more of a Psy-Op than a real tool for anything useful. Persuading low-IQ people that the bots are sentient only serves to fortify the credibility of the AI's "answers", and cause people to believe it, without wondering what's missing from the answer, such as the most-relevant truth that someone doesn't want the masses to know.