ChatGPT is just a statistical engine that outputs text consistent with its previous input. So while it understands that requests to count words often end with s

ChatGPT is just a statistical engine that outputs text consistent with its previous input. So while it understands that requests to count words often end with something like "this passage has <number> words", and therefore puts a number in the sentence to make it statistically probable, it doesn't have any way of understanding that the number is supposed to correlate to the number of tokens in the passage referenced. It doesn't understand the meaning of the words you are putting into it, it just understands how to generate statistically probable text from those words. Furthermore because concepts like "word count" are statistically separate from the semantic content of the sentence, ChatGPT has an especially hard time building any kind of statistical model that outputs the correct number based on the query.