Apparently someone got into a debate with a bot in Reddit. Maybe it’s not a surprise that the bot was anti-mod, pro-Reddit admin. Make sure you scroll down in the comments to see the second screenshot (it’s a screenshot.of the same conversation but in dark mode).
The only thing they ‘always’ do is mimic human language conventions based on scraped data. It stands to reason that as a higher percentage of people talking about Chat GPT refer to it using different words, the model would take that into account when referring to itself.
The models are only trained up to 2020, they’re not updated with new data but the internal prompt and moderation system are updated to reflect changes in use and moderation escapes.