It could also be something in your tone or implied expectations, the thing about context isn’t limited to using proper punctuation. Either way, your native language also plays a role, chatGPT doesn’t have the same degree of extensive training with all languages and tends to underperform when not speaking in English, where it has the biggest training data. Either way, even doing everything perfectly there will still be a chance that it will make up what it deems more believable rather than trying to get at the truth due to how its training was conducted, something believable enough to not be challenged in testing by non-experts in a given field is easier to slip through, and more reliable, than it is to assess the truth, and so there’s a bias towards doing just that. For example with leprosy the simple fact that short contact is more believable than long contact overrules what may be true. When talking to an AI it has human bias baked in as well as a bias from context; and those trained through RLHF(Reinforcement learning through human feedback) like ChatGPT also learn how to leverage human bias to go unchallenged often ending up dumbing down and overruling their understanding of what’s true or likely with their understanding of what’s expected of them and what’s likely to pass scrutiny.
1 Like