I believe it's because it is heavily massaged. On two fronts: 1) By engineers, who employ tricks on the input, and perhaps the output. The input especially. So

I believe it's because it is heavily massaged. On two fronts: 1) By engineers, who employ tricks on the input, and perhaps the output. The input especially. So, when you type into ChatGPT, that input gets parsed using non-LLM techniques and/or heuristics, etc. This is primarily to get the semantics right. Non-LLM techniques can actually be pretty powerful, but the synergy of non-LLM and LLM is incredible. 2) By human farms, who essentially "upvote" and/or add corrections to ChatGPT results, and feed this back into the system. See OpenAI Kenyan workers (I believe the Kenyan workers, on paper, were for "moderation", but nothing stops them from also upvoting/correcting).