ChatGPT gets basic counting wrong and never once signals it might be unsure
retained text
I keep hitting the same failure and it isn't the mistake itself, it's the confidence. Ask it to count the letters in a word, the items in a list, or how many times something appears, and it's often off by one or two. Fine, models are bad at this. The real problem: it delivers the wrong number in the exact same certain tone it uses when it's right. No hedge, no "let me double check," nothing. Why it happens: it isn't actually counting. It predicts what a plausible answer looks like, so there's no internal tally to be right or wrong against. That's also why "are you sure?" often makes it flip to a different wrong number instead of verifying. What actually works for me: \- For anything numeric, tell it to use its code tool or count step by step out loud. Forcing an explicit process beats asking for the answer. \- Treat every raw count it gives as a draft, never trust it in client-facing work. \- If it doubles down when challenged, that's your signal it's guessing, not that it's right. I recorded a 40-second demo of it happening live if anyone wants to see the exact behavior: [https://youtu.be/-ioP3XCH-3k](https://youtu.be/-ioP3XCH-3k)
document fetched 2026-09-17 00:20 · canonical https://reddit.com/r/ChatGPT/comments/1wi3a8r/chatgpt_gets_basic_counting_wrong_and_never_once