Comment by zahlman

> The agent wasn’t drawing on the highest human knowledge. It was drawing on what gets engagement, what “works” in the sense of generating attention and emotional reaction.

> It pattern-matched to the genre of “aggrieved party writes takedown blog post” because that’s a well-represented pattern in the training data, and that genre works through appeal to outrage, not through wisdom. It had every tool available to it and reached for the lowest one.

Yes. It was drawing on its model of what humans most commonly do in similar situations, which presumably is biased by what is most visible in the training data. All of this should be expected as the default outcome, once you've built in enough agency.