Questions:
- What gap are they finding in the literature?
- “Research has shown that digital conversations are prone to stagnation at a low level of cognitive complexity, where participants repeat familiar arguments rather than building on or critically engaging in the ideas of others [Papacharissi, 2004, Nithyanand et al., 2017, Sunstein, 2017, Boukes, 2025].” (Braley et al., 2025, p. 4)
- “One way to frame this issue is to consider that online discussion platforms are often designed to maximize the “exploration” phase of engagement, where participants share ideas but have not yet synthesized them into actionable insights or consensus positions [March, 1991, Raghavan et al., 2018].” (Braley et al., 2025, p. 4)
- “Few systems currently support a transition to the “exploitation” phase, where shared understanding can be translated into decision-making or collective action, because discussion platforms typically lack tools to scaffold reasoning, highlight areas of agreement, or facilitate areas of conflict.” (Braley et al., 2025, p. 4)
- Surprisingly has similar reasoning to luDeliberationWorksDeliberationSystem2025
- What is the LLM helping with?
- Very simple. Prompted LLM sends message to participants with statement that summarizes primary area of agreement for the group, and provide participants with new discussion question for the next round by identifying an area in the conversation that, if discussed with more detail, would bring group to greater consensus.
- Participants vote on the proposed consensus statement, and the subsequent discussion question. Initiator of deliberation defines threshold and diversity of agreement necessary for a consensus statement.
- What is the deliberation structure?
- Rate main issue on scale from 0 to 100 (“Do you support the use of military force against civilians if there are peaceful protests when the next president takes office?“)
- Contribute their own comment, constrained to 280ch
- Evaluate peer-submitted comments using agree/disagree/pass
- Reassess main issue, using 0-100 scale
- Other notes:
- Prohibits direct replies to others’ comments, based on nonviolent communication principles, emphasizing expressing one’s own perspective rather than reacting to others.
- This feels like a convenient justification for not accounting for perspective change.
- Prohibits direct replies to others’ comments, based on nonviolent communication principles, emphasizing expressing one’s own perspective rather than reacting to others.
- What prompts do they use?
- Appendix F is empty.
- Design
- In control, no one proposes a new follow-up discussion, but simply continues with initial prompt
- Measure of success:
- Nuance of conversation
- are there more topic clusters?
- is there more word diversity?
- is there greater cosine similarity between participant’s initial and final comment?
- Favorability of outcomes
- Petition preference: Did participants choose standard petition and the AI-generated petition?
- Consensus levels: comparing standard deviation of votes to see if treatment group reached higher agreement
- Survey questions
- “Made me think differently”
- “Learned something new”
- “Felt included”
- “Legitimate process”
- “Would participate again”
- Nuance of conversation
This feels pretty sloppy to me. Big limitations: participants did not retain agency over whether to adopt or not adopt the LLM generated next discussion question, and stopped at 2 rounds.