I don’t really understand how the research design leads to the author’s conclusions. For example they state they created the AI personas based off of the american national election survey but go on to make comments about global internet behavior. With that data it seems like they should only suggest the model as an attempt at approximating the behavior american voters. If someone takes the time to read the actual arxiv paper I’d be curious to know if the methodology is more robust than this article implies. Feels a bit like they designed opposing archetypes based on us political surveys then tried to say interventions in social media design (re-organizing feeds, hiding likes, etc) wouldn’t help generally.

It’s not a perfect solution—there’s all kind of biases and limitations—but it does represent a step forward compared to a list of if/then rules. It does have something more of capturing human behavior in a more plausible way. We give them personas that we get from the American National Election Survey, which has very detailed questions about US voters and their hobbies and preferences. And then we turn that into a textual persona—your name is Bob, you’re from Massachusetts, and you like fishing—just to give them something to talk about and a little bit richer representation.