tldr:

There are 3 problems:

Echo chambers or filter bubbles. You need to have a diversity of opinion, a diversity of perspective. Deliberation needs to be among equals; people need to have more or less the same influence in the conversation.
The social media prism. The more extreme users tend to get more attention online due to algorithmic amplification and human bias towards negativity.

IMO Lemmy/PieFed only really suffers from the first problem.

Despite the obvious problems with using LLMs to pretend to be people (which the authors acknowledge), they found this basic dynamic:

You hit retweet when you see someone being angry about something, or doing something horrific, and then you share that. It’s well-known that this leads to toxic, more polarized content spreading more.

But what we find is that it’s not just that this content spreads; it also shapes the network structures that are formed [i.e. who follows who]. So there’s feedback between the effective emotional action of choosing to retweet something and the network structure that emerges. And then in turn, you have a network structure that feeds back what content you see, resulting in a toxic network. The definition of an online social network is that you have this kind of posting, reposting, and following dynamics. It’s quite fundamental to it. That alone seems to be enough to drive these negative outcomes.

We tested six different interventions [chronological feeds, etc]

but

it doesn’t really make a difference in changing the basic outcomes.