- Online deliberation
- ML-supported scaled deliberation
- Polis is agnostic to language, since it depends completely on PCA components based on a matrix of votes on comments. This also means that you can only represent variance along an axis of topic according to the number of PCA components generated. In their pilot of Polis, the only useful component was the first which captured the variance between those in favor and those in opposition to Uber. Interpreting each PCA component requires comparing the “top-loading” comments from each end of the PCA components. Their output report includes highlighted comments with high overall support, and comments that best explain how each opinion group differs form the rest of the conversation. See smallPolisEscalarDeliberacion2021.
- AI-mediated deliberation
- smallOpportunitiesRisksLLMs2023 points out that LLMs can augment human intelligence for tasks like summarization and topic modeling, which is the element that required interpretation in the original version of Polis. There are higher risk tasks like prompting LLMs to come up with discussion questions based on previous conversation towards consensus, and predicting people’s votes.
- braleyModerationHowAI2025 implements the “facilitating conversations” task from above. They show some benefits of AI facilitation, which seem to me like a case for how deliberation would benefit from mediators.
- The biggest “innovation” is prompting LLM to make summary statements on which participants can further vote on, prompting them to come up with more questions towards consensus, and drafting a more favorable petition statement based on the deliberation.
- tesslerAICanHelp2024 built a Habermas Machine that uses Shulze ranked-choice voting method to reduce Strategic manipulation and Winner’s curse. This resulted in more favorable consensus statements over human-generated ones, and demonstrably incorporating minority critiques into revised statements.
- Realizing this definition of “consensus” is not much different from opinion group summaries.
- boehmerGenerativeSocialChoice2025 is a method of summary generation for deliberation methods like Polis, with an algorithmic guarantee of the proportional representation of viewpoints, grounded in Social choice theory. Now that I understand what that means, we might have to be careful about how the guarantee has to do with proportional representation, especially w.r.t. 5 conceptual models of fairness for UAM.
- braleyFaceCrowdGenerating2025 is not entirely relevant.
- LLMs as voting agents
- LLMs as individual reflection mediators
- yeoEnhancingDeliberativenessEvaluating2025 is about using multimodal models to induce reflection pre-deliberation. While I believe that reflection has usefulness in this area, the way they use “storytelling” is 10 GPT-4 personas with a “distinct perspective on the discussion topic”, which completely misses the mark—the point of involving storytelling, at least for me, is to surface the lived experience of actual people and for those stories to be heard by others—not GPT-generated ones.
- fulayEmptyChairUsing2025 is using LLMs to fill in missing points of view in group deliberation settings.
- yeoHelpMeReflect2024