Could you maybe provide a bit of context about where this idea of using generative AI to support this kind of deliberation come, came from the first place? Because I imagine when people first hear about generative AI and consensus, and it's not the first thing people think of.

Yeah, ab-absolutely.

So the most direct inspiration for this is a piece of work that came out of DeepMind, uh, that was published in Science last year or the year before, that they called it the Habermas Machine, and that's named after a German philosopher called Jürgen Habermas, whose main argument was that legitimate decisions in pluralistic societies have to come from inclusive and rational and well-informed discourse, not just kind of naive majority votes.

And what the DeepMind team showed in their paper was that you can use a large language model as kind of like a skilled facilitator that will take everyone's positions, surface where there are disagreements, and then draft proposals that, uh, try to capture as much of the, the bridging or the agreement amongst the group as possible.

And then participants rate those, and the model iterates and, um, what they showed in the paper is that you get measurably more convergence than you do from people just deliberating unstructured groups.

And that's one of the kind of shining examples in an increasingly rich literature, um, that we refer to as generative social choice theory, and that's just the broad area of language model-generated content helping people to come to agreement.

But the lineage actually goes back much further than that.

Yeah.