Goldberg, B., Acosta-Navas, D., Bakker, M., Beacock, I., Botvinick, M., Buch, P., DiResta, R., Donthi, N., Fast, N., Iyer, R., Jalan, Z., Konya, A., Danciu, G. K., Landemore, H., Marwick, A., Miller, C., Ovadya, A., Saltz, E., Schirch, L., Shalom, D., Siddarth, D., Sieker, F., Small, C., Stray, J., Tang, A., Tessler, M. H., & Zhang, A. (2026). AI and the future of digital public squares. Collective Intelligence, 5. https://doi.org/10.1177/26339137261459441
Summary
This position paper — the product of a 2024 convening of over 70 experts across academia, industry, and civil society — argues that large language models represent a potential second paradigm shift for digital public squares, following the internet’s first reshaping of public discourse. Situating itself within democratic theory (Habermas’s public sphere, Rawls’s public reason, Landemore’s Open Democracy) and the sociotechnical study of platforms, it contends that LLMs could push online conversation toward more inclusive, deliberative, and participatory forms — but equally risk deepening polarization, surveillance, and distorted information ecosystems. The authors organize their analysis around four families of AI-enabled technology and lay out a concrete research and investment agenda for each, insisting throughout that AI should augment rather than replace human deliberation.
Key Contributions
- A structured framework mapping four AI-enabled technology families — collective dialogue systems, bridging systems, community-driven moderation, and proof-of-humanity systems — to specific opportunities, risks, and mitigations.
- A bridge between democratic theory and applied sociotechnical research, connecting ideals of open democracy and public reason to concrete LLM applications.
- A concrete research and investment agenda per domain: open CDS benchmark datasets, interoperable elicitation-inference toolboxes, meta-platforms for composable deliberation, and methods for measuring bridging outcomes.
- A consolidated cross-sector reference articulating both the promise and peril of AI in democratic discourse.
Methods
The paper synthesizes insights from an April 2024 convening in New York City of the co-authors plus roughly 50 additional experts, integrating applied research, case studies, and reviews of existing platforms (Polis, Remesh, Make.org, All Our Ideas, Community Notes, Perspective API, Policy Synth). Each of the four technology domains is analyzed for current applications, challenges, AI-enabled opportunities, risks and mitigations, and future research directions, drawing on preference aggregation, generative social choice, and matrix factorization alongside democratic theory.
Findings
- Collective dialogue systems sit between surveys and focus groups, offering scalable yet nuanced feedback, but currently demand heavy human labor for setup, facilitation, moderation, and sensemaking. LLMs can ease these bottlenecks via facilitator training (RAG), participant education, intelligent mediation, real-time translation, elicitation inference (vote prediction), and automated summarization.
- Synthetic AI participation in deliberation is promising for representativeness but democratically problematic — risking lost legitimacy, diluted public agency, and convincing misrepresentation of reality. Aggregation of opinions is only one part of genuine deliberation, which requires iterative exchange of reasons.
- Bridging systems can elevate content that builds trust and common ground: Meta’s diverse-engagement ranking reduced polarizing content while raising comment views by 0.69%; X’s Community Notes uses matrix factorization to surface notes rated helpful across the primary axis of division. Bridging can draw on user-diversity signals or content analysis, with a combination likely optimal — but risks elevating insubstantial content, being gamed, or being perceived as illegitimate.
- Community-driven moderation is more robust and legitimate when moderators are empowered with AI tools rather than moderation being centralized. LLMs still struggle with context (ChatGPT 3.5 performed worse than a coin toss on some subreddits); AI can nonetheless support proactive moderation (listening, simulation, guideline co-creation, nudges) and reactive moderation (customizable filters, triaging, appeals tools like AppealMod).
Connections
This paper’s normative framing of platform design and moderation connects to work on the governance and stakes of algorithmic public spaces, such as Gillespie2026-aa and Rieder2026-pp. Its treatment of Community Notes and community-driven, AI-assisted moderation relates to empirical studies of crowd and platform moderation like Bak-Coleman2025-pm and Allen2025-ot. Its concern with LLM-augmented deliberation and synthetic participation intersects with research on AI-generated persuasion and discourse such as Hackenburg2026-ud.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify