return ✕︎


By E. Glen Weyl, Audrey Tang and ⿻ Community


     As we have noted above, one the most common concerns about social media has been its tendency to entrench existing social divisions, creating "echo chambers" that undermine a sense of shared reality. Whatever one's evaluation of the extent to which this is true (relative to what counterfactual), it is natural to ask how these systems might be design with an opposite intention. The largest-scale attempt at this is the Community Notes (formerly Birdwatch) system in the X (formerly Twitter) social media platform.

Twitter Community Notes

     Usually seen as a variant of "fact checking", Community Notes (CN) allows members of the X community to suggest contextualizing additional information on posts that are flagged as potentially misleading as illustrated above. Participants in this process not only submit suggested notes: they also rate the notes suggested by others. These ratings are used to asses the overall perspective of a rater statistically.

     Specifically, raters are placed on one-dimensional spectrum of opinion, discovered by the statistical analysis from the data but in practice corresponding in most applications to the "left-right" divide in the politics of much of the Western hemisphere. Then (or really simultaneously), the support each note receives from any community member is attributed to a combination of its affinity to their position on this spectrum and some underlying, position-agnostic "objective quality". Notes are then shown if this objective quality, rather than their overall support, is sufficiently high. This represents a way of prioritizing content shown based on the principle of collaboration across diversity, consistent with Plurality, to which hundreds of millions of people are currently exposed each week.

     Words, whether written or spoken, are probably the most common and consistent form of communication and collaboration in human history on this planet. Friendships, business, politics, science, culture and much else certainly rely on other modes of interaction, but the exchange of ideas, directions, critics, feelings and more through verbal communication is typically at the center of them. In this chapter, we will explore the power and severe limits of conversation today and hope advances in Plurality might make words an engine for both proliferating and bridging diversity like never before.

Conversation today

     The omnipresence of verbal exchange in human life makes it daunting to classify, especially as briefly as we must, but we will roughly distinguish two patterns (oral and written) and a third that combines the first two (networked) in which written summaries of oral interactions are transmitted across distances and possibly used to stimulate other oral interactions.

     Oldest, typically richest and still most common form of verbal exchange is the in-person meeting. The business meeting and negotiation is how most consequential decisions in such arenas are made. Idealized portraits of democracy typically refer more heavily to the discussions, such as took place among traditional tribes, in the Athenian marketplaces or in New England town halls, than to votes or media. The recent film, Women Talking, brilliantly captures this spirit in its portrait of a traumatized community coming to a plan for common action through discussion. Groups of friends, clubs, students and teachers, all exchange perspectives, learn, grow and form common purpose through in-person talk. In addition to their interactive nature, in-person interactions often carry elements of the richer, non-verbal communication we described above, as participants share a physical context and can perceive many non-verbal cues from others in the conversation.

     The next oldest and most common communicative form is writing. While far less interactive, writing allows words to travel across much greater space and time. Typically conceived as capturing the voice of a single "author", written communications can spread broadly (even globally), especially with the aid of printing and translation, and endure for sometimes thousands of years, allowing for a "broadcast" of messages much farther than amphitheaters or loudspeakers (though recorded audio and visual formats can rival this).

     As this discussion illustrates, there have long been severe trade-offs between the richness of in-person discussion and the reach of the written word. Many structures attempt to harness elements of both through a network where in-person conversations and deliberations are nodes and writing acts as edges. Examples include many constitutional and rule-making processes (where groups deliberate on writing that is then submitted to another deliberation that results in another document that is then sent back to other deliberations), book clubs, editorial boards for publications, focus groups, surveys, and other research processes, etc.

     One of the most fundamental challenges this variety of forms tries to navigate is the trade-off between speed and inclusion. On the one hand, conversations are long, costly and time-consuming. This often means that they have trouble yielding definite and timely outcomes, the "analysis paralysis" often bemoaned in corporate settings and complaint that (sometimes attributed to Oscar Wilde) that "socialism takes too many evenings".

     On the other hand, deliberation often struggles to be inclusive. People affected by a topic are often broadly geographically dispersed, speak different languages, have different conversational norms, etc. Diversity in conversational styles, cultures and language often impede mutual understanding. Furthermore, given that it is impossible for everyone to be heard at length, some notion of representation is necessary for conversation to cross broad social diversity, as we will discuss at length below.

     Perhaps the fundamental limit on all these approaches is that while methods of broadcast (allowing many to hear a single statement) have dramatically improved, broad listening (allowing one person to thoughtfully digest a range of perspectives) remains extremely costly and time consuming. As economics Nobel Laureate and computer science pioneer Herbert Simon observed, "(A) wealth of information creates a poverty of attention." The limits on attention impose potentially sharp trade-offs on richness and inclusion.

     A number of strategies have, historically and more recently, been used to navigate these challenges. Representatives are chosen for conversations by a variety of methods, including

  1. Election: A campaign and voting process is used to select representatives, often based on geographic or political party groups. This is used most commonly in politics, unions and churches, with the advantage of conferring a degree of broad participation, legitimacy and expertise, but often being rigid and expensive.
  2. Sortition: A set of people are chosen randomly, sometimes with checks or constraints to ensure some sort of balance across groups. This is used most commonly in focus groups and surveys, and maintains reasonable legitimacy and flexibility at low cost, but sacrifices (or needs to supplement with) expertise and has limited participation
  3. Administration: A set of people are chosen by a bureaucratic assignment procedure, based on "merit" or managerial decisions to represent different relevant perspectives or constituencies. This is used most commonly in business and professional organizations and tends to have relatively high expertise and flexibility at low cost, but has lower legitimacy and participation.

     Once participants to a deliberation are selected and arrive, facilitating a meaningful interaction is an equally significant challenge and is a science unto itself. Ensuring all participants, whatever their communicative modes and styles, are able to be fully heard requires a range of techniques, including active inclusion, careful management of turn-taking, encouragement of active listening and often translation and accommodation of differing abilities for auditory and visual communication styles. These strategies can help overcome the "tyranny of structurelessness" that often affects attempts and inclusive and democratic governance, where unfair informal norms and dominance hierarchies override intentions for inclusive exchange.

     Some of the challenges to inclusion have been significant ameliorated by technology in recent years. Physical travel distance used to be a severe impediment to deliberation. Phone conferences and even more video conferences have significantly mitigating this challenge, making various formats of distance/virtual meetings increasingly common loci for challenging discussions.

     Similarly expansive of inclusion in written formats has been the advent of internet-mediated writing, including email, message boards/usenets, webpages, blogs, and especially social media. These have offered a variety of alternative approaches for allocating limited attention, including by (effectively) popular vote of friends (via "likes" or "reposts") or website ranking and thus have allowed significant changes in the network structure imposed by editorial processes. While these have in many ways included many who could not speak in the past, the challenges of allocating attention continue to bite, with many of these media lacking context and thoughtful facilitation, leading to many of the problems we have highlighted previously including "misinformation", "disinformation" and flooding by well-resourced actors.

Conversation tomorrow

     A range of recent advances have begun to push back the frontier of these trade-offs, allowing more effective networked sharing of rich in-person deliberation and more thoughtful, balanced and contextualized moderation of more inclusive social media forms.

     As we discussed in the "View from YuShan" chapter above, one of the most successful examples in Taiwan has been the vTaiwan system, which harnesses OSS called "" in English. This platform shares some features with social media services like X, but builds abstractions of some of the principles of inclusive facilitation into its attention allocation and user experience. As in X, users submit short responses to a prompt. But rather than amplifying or responding to one another's comments, they simple vote these up or down. User perspectives are then clustered, to highlight patterns of common attitudes. Representative statements that highlight these differing opinion groups' perspectives are displayed to allow users to understand the key perspectives in the conversation, as are the perspectives that "bridge" the divisions: ones that receive assent across the lines that otherwise divide. Responding to this evolving conversation, users can offer additional perspectives that help to further bridge, articulate an existing position or draw out a new opinion group that may not yet be salient. is a prominent example of what leading plural technologists Aviv Ovadya and Luke Thorburn call "collective response systems" and others call "wikisurveys; other leading examples include All Our Ideas and Remesh, which have various trade-offs in terms of user experience, degree of open source and other features. What these systems share is that they combine the participatory, open and interactive nature of social media with features that encourage thoughtful listening, an understanding of conversational dynamics and the careful emergence of an understanding of shared views and points of rough consensus. Such systems have been used to make increasingly consequential policy and design decisions, around topics such as the regulation of ride-hailing applications and the direction of some of the leading large foundation models (e.g. Anthropic's Claude). They have also been central to inspiring related approaches, such as X Community Notes that we highlighted above, that have even greater reach and influence.

     An approach with similar goals but a bit of an opposite stating point is one that centers in-person conversations but aims to improve the way their insights can be networked and shared. A leading example in this category are the approaches developed by the Massachusetts Institute of Technology's Center for Constructive Communication in collaboration with their civil society collaborators Cortico. These use a mixture of the identity and association protocols we discussed in the Freedom part of the book and natural language processing to allow recorded conversations on challenging topics to remain protected and private while surfacing insights that can travel across these conversations and spark further discussion and be highlighted by community members for broader public, including policy-makers and university administrators. Related tools, of differing degrees of sophistication, are used by organizations like StoryCorps and Braver Angels and have reached millions of people.

     Another more experimental efforts overlap heavily with the approaches we highlighted in "Immersive Shared Reality", aiming to push the richness of remote deliberations towards that possible in person. A recent dramatic illustration was a conversation between Meta CEO Mark Zuckerberg and leading podcast host Lex Fridman, where both were in virtual reality able to perceive minute facial expressions of the other. A less dramatic but perhaps more meaningful example was the Portals Policing Project, where cargo containers appeared in cities affected by police violence and allowed an enriched video-based exchange of experiences with such violence across physical and social distance. Other promising elements include the increasing ubiquity of high-quality, low-cost and increasingly culturally aware machine translation tools and work to harness similar systems to enable people to synthesize values and find common ground building from natural language statements.

Frontiers of augmented deliberation

     Some of these more ambitious experiments begin to point towards a future, especially harnessing large language models (LLMs), where we go much further towards addressing the broad listening problem, empowering deliberation of a quality and scale that has henceforth been hard to imagine.

     One of the most obvious directions that is a subject of active development is how systems like and Community Notes could be extended with modern graph theory and LLMs. The "Talk to the City" project, for example, illustrates how LLMs can be used to replace a list of statements characterizing a group's views with an interactive agent one can talk to and get a sense for the perspective. Soon it should certainly be possible to go further, with LLMs avoiding participants being limited to short statements and up-and-down votes, instead allowing them to fully express themselves in reaction to the conversation, but with the models condensing this conversation and making it legible to others who can then participate. Models could also help look for areas of rough consensus not simply based on common votes but on a natural language understanding of and response to expressed positions.

     Furthermore, while the current discussion around collective response models focuses on identifying areas of rough consensus, another powerful role is to support the regeneration of diversity and productive conflict. On the one hand, they help identify different opinion groups in ways that are not a deterministic function of historical assumptions or identities, potentially allowing these groups to find each other and organize around their perspective. On the other hand, by surfacing as representing consensus positions that have diverse support, they also create diverse opposition that can coalesce into a new conflict that does not reinforce existing divisions, potentially allowing organization around that perspective. In short, collective response systems can play just as important a role in mapping and evolving conflict dynamically as helping to navigate it productively.

     In a similar spirit, one can imagine harnessing and advancing elements of the design of Community Notes to more holistically reshape social media dynamics. While the systems currently lines up all opinion across the platform on a single spectrum, one can imagine mapping out a range of communities within the platform and harnessing its bridging-based approach not just to prioritize notes, but to prioritize content for attention in the first place. Furthermore, bridging can be applied at many different scales and to many intersecting groups, not just to the platform overall. One can imagine a future where different content in a feed is highlighted as bridging and being shared among a range of communities one is a member of (a religious community, a physically local community, a political community), reinforcing context and common knowledge and action in a range of social affiliations.

     Such dynamic representations of social life could also dramatically improve how we approach representation and selection of participants for deeper deliberation, such as in person or in rich immersive shared realities. With a richer accounting of relevant social differences, it may be possible to move beyond geography or simple demographics and skills as groups that need to be represented. Instead, it maybe possible to increasingly use the full intersectional richness of identity as a basis for considering inclusion and representation. Constituencies defined this way could participate in elections or, instead of sortition, protocols could be devised to choose the maximally diverse committees for a deliberation by, for example, choosing a collection of participants that minimizes how marginalized from representation the most marginalized participants are based on known social connections and affiliations. Such an approach could achieve many of the benefits of sortition, administration and election simultaneously, especially if combined with some of the liquid democracy approaches we discuss in the voting chapter below.

     It may be possible to, in some cases, even more radically re-imagine the idea of representation. LLMs can be "fine-tuned" to increasingly accurately mimic the ideas and styles of individuals, as we discussed in the previous chapter. But there is nothing special to individual humans about this approach: it is simply based on a body of text. One can imagine training a model on the text of a community of people and thus, rather than representing one person's perspective, it could operate as a fairly direct collective representative, possibly as an aid, complement or check on the discretion of a person intended to represent that group.

     Deliberation in advanced AI systems is a desireable, since it is relevant for understanding human-aligned AI. OpenAI's Democratic Inputs[1] initiative seeks to explore experimental democratic processes for deliberating and adherence of AI systems. Early experiments also showcased debates between human-AI pairs[2]. Another effort is also explored by Anthropic with Constitutional AI (CAI), whereby a group of ~1000 individuals drafted a constitution for an AI System. Such deliberations where made to unearth how democratic processes can influence democratic AI. Claude also utelizes CAI inspired by constitutions such as the United Nations Universal Declaration of Human Rights. Polis is the main deliberation platform used by Anthropic OpenAI, and AI Alignment Assemblies.

     Most boldly, this idea could in principle extend beyond living human beings. In his classic The Parliament of Things, philosopher Bruno Latour argued that natural features (like rivers and forests) deserve representation in governments. The challenge, of course, is how these can speak. LLMs might offer ways to translate scientific measures of the state of these systems into a kind of "Lorax", Dr. Seuss's mythical creature who speaks for the trees and animals that cannot speak for themselves. Something similar might occur for unborn future generations, as in Kim Stanley Robinson's Ministry for the Future. For better or worse, such LLM-based representatives might be capable of carrying out deliberations faster than most humans can follow and might then convey summaries to human participants, allowing for deliberations that include individual humans and also allow for other styles, speeds and scales of natural language exchange.

Limits of deliberation

     The centrality of natural language to human interaction makes it tempting to forget its severe limitations. Words may be richer symbols than numbers, but they are as dust compared to the richness of human sensory experience, not to mention proprioception. "Words cannot capture" far more than they can. Whatever emotional truth it has, it is simply information theoretically logical that we form far deeper attention in common action and experience than in verbal exchange. Thus, however far deliberation advances, it cannot substitute for the richer forms of collaboration we have already discussed.

     Furthermore, and on the opposite side, talk takes time, even in the sophisticated versions we describe. Many decisions cannot wait for deliberation to fully run its course, especially when great social distance has to be bridged, which will generally slow the process. The other approaches to collaboration we discuss below will typically be needed to address the need for timely decisions in many cases.

     Furthermore, many of the ways in which the slow pace of discussion can be overcome (e.g. using LLMs to conduct partially "in silico" deliberation) illustrate another important limitation of conversation: many other methods are often more easily made transparent and thus broadly legitimate. The way conversations take inputs and produce outputs are hard to fully describe, whether they occur across people or in machines. In fact, one could consider inputting natural language to a machine and producing a machine dictate as just a more sophisticated, non-linear form of voting. But, in contrast to the administrative and voting rules we will discuss below, it might be very hard to achieve common understanding and legitimacy on how this transformation takes place and thus make it the basis for common action in the way that voting and markets often are. Thus checks on the way deliberations occur and are observed arising from those other systems are likely to be important for a long time to come.

     Furthermore, deliberation in the democratic process is also limited by the ability for humans to practically survey more capable LLM (scalable oversight). This poses a unique limitation for model evaluators where supervision is not sufficient, due to overly powerful AI Systems. LLMs have also been demonstrated to adhere to instructions blindly; this can raise issues around LLM Censorship as a factor that can undermine the democratic process within AI systems[3].

     Furthermore, deliberation is sometimes idealized as helping overcome divisions and reach a true "common will". Yet, while reaching points of overlapping and rough consensus is crucial for common action, so is the regeneration of diversity and productive conflict to fuel dynamism and ensure productive inputs to future deliberations. Thus deliberations and their balance with other modes of collaboration must always attend, as we have illustrated above, to this stimulus to productive conflict as much as it does to the resolution towards active and away from explosive conflict.

  1. Tyna, Eloundou (OpenAI) et al. (2023) Democratic inputs to AI ↩︎

  2. Geoffrey Irving, Paul Christiano, Dario Amodei. (2018) AI safety via debate ↩︎

  3. David Glukhov, Ilia Shumailov, Yarin Gal, Nicolas Papernot, Vardan Papyan (2023). LLM Censorship: A Machine Learning Challenge or a Computer Security Problem? ↩︎