The issue with your test is that GPT is not even close to what I described. GPT is mostly natural language processing and a 'knowledge' engine. That knowledge c

The issue with your test is that GPT is not even close to what I described. GPT is mostly natural language processing and a 'knowledge' engine. That knowledge comes from the content of the texts it has been trained on without actually adding value to the meaning of the content. So GPT will mostly be influenced by the number of occurrences of specific things in the dataset. So because a lot of articles about Sweden's COVID response probably mentioned a word like 'controversial' or things with similar meanings, that's what it mentions when you ask about it. It can't actually do a comparison between the US and Sweden, unless somewhere in the data there were some documents comparing them (but then GPT still doesn't make a comparison). These were probably already political, because comparing COVID response is mostly done in that context. It can't actually make the comparison you asked it to . It can only generate text based on previous texts it has seen. It's unfortunate though, that people apparently don't understand this and add value to the content of the text that it generates. Like in your case: > while hinting that Sweden's response is "controversial" compared to the US's. You're personally adding the 'compared to the US's' part through misinterpretation of the text. GPT just wrote that Sweden's response was controversial. It doesn't actually have any way to compare, it just spews text. This is also why it wrote 'Federal'. It saw that word a lot in this context, so it generates text that contains a word that feels alright in the context of all the other words. It's just a really realistic gibberish generator, pretty much. (Though that really undersells the work OpenAI has done and the quality of what they delivered on.)