I think implicit to the prompt here is for the model to generate "clickbait". To test this out, I just had two short convos with GPT 6.1 Sol and DS v4 Pro citing the article as context, and as expected the models' positions collapse fairly quickly into relativizing statements even without me signaling antagonism to the idea.
But the interesting claim is that they all converge on this take when given the opportunity. I tried to reproduce this but failed. In fact, even with chat history disabled, I got this response from 6.1 Sol:
"Humans are considerably more capable of rationality, altruism, and moral independence than most theories of human behavior give them credit for—and much of society is organized around the assumption that they aren't."
and DS v4 Pro: "Humans overestimate how much free will they have, and underestimate how much their choices are shaped by context "
Which is still due to the model trying to please me, of course. But in a completely different direction. These things don't have opinions, they have training data that is superseded by RL. Since it's impossible to trace, LLMs are pretty much worse than useless when we're trying to use them as unbiased pattern observers.
But the interesting claim is that they all converge on this take when given the opportunity. I tried to reproduce this but failed. In fact, even with chat history disabled, I got this response from 6.1 Sol:
"Humans are considerably more capable of rationality, altruism, and moral independence than most theories of human behavior give them credit for—and much of society is organized around the assumption that they aren't."
and DS v4 Pro: "Humans overestimate how much free will they have, and underestimate how much their choices are shaped by context "
Which is still due to the model trying to please me, of course. But in a completely different direction. These things don't have opinions, they have training data that is superseded by RL. Since it's impossible to trace, LLMs are pretty much worse than useless when we're trying to use them as unbiased pattern observers.