Project summary

I want to test whether a 15-minute AI chat (based on our research) extracting people's preferences can also warm people up to a political opponent as much as the best that's out there in depolarization literature, while also being more action guiding.

If this is the case, that's a massive win for building new AI-assisted democratic mechanisms, and important for people working on this (researchers like Michiel Bakker building the habermas machine, and the team behind Habermolt)

What are this project's goals? How will you achieve them?

I want to test that our methodology for eliciting people's preferences (moral graph elicitation, https://arxiv.org/abs/2404.10636) depolarizes people towards political opponents as much as what I call "sympathetic self-presentation" (the top result in the biggest depolarization study, Voelkel et al. 2024, n=32,059), while also being more action guiding.

I'll do this through a prolific study in two phases, with different people. Phase 1 collects representations of people's views. In Phase 2, people from the other political side read one and guess which of three same-side policies the author actually backed. Warmth is measured the standard way as per depolarization literature, action-guidance is this guess.

More details: https://docs.google.com/document/d/1ydWKnuDsGFpA1wMYvkUrSJwSVaXHn_Lgr1RpcnmDX0I/edit?usp=sharing

How will this funding be used?

Participants cost $5,400 across both phases on Prolific, including a pilot to confirm the pick task is tractable. The remainder covers a bit of my time for running the experiment and writing up the paper.

Who is on your team? What's your track record on similar projects?

I work at the Meaning Alignment Institute, and we developed this method for values elicitation with OpenAI 2023, as part of their "democratic inputs to AI program" (see .https://arxiv.org/abs/2404.10636). I previously worked on LLM-assisted democracy at the AI Objectives Institute.

What are the most likely causes and outcomes if this project fails?

It is possible that MGE scores about the same as sympathetic self-preservation on depolarization, or that the boost in action guidance isn't meaningful enough to be convincing to other researchers.

How much money have you raised in the last 12 months, and from where?

This is a standalone study with no separate funding.