PaperlayerView exact revision on arXiv

Source-linked AI summary

Improving social media for democratic discourse

Fan Cheng, Amirhossein Farzmahdi, Pinyuan Feng, Kedar Garzón Gupta, Trenton Jerde, Nikolaus Kriegeskorte, Zi Qi Liow, Akihito Maruya, Savannah Smith, Patrick Stinson, JohnMark Taylor

arXiv:2610.10816v1cs.CYcs.HC

TL;DR

Social media enable political participation but are associated with concerns including polarization and misinformation, making democratic design an important challenge. This white paper develops modular mechanisms as an accessible menu of possibilities for improving democratic discourse.

  • Problem

    Social media are associated with both political participation and concerns such as polarization and misinformation, making their design for democratic discourse an important challenge.

    Source: Introduction · pp. 5, 7

  • Method

    The authors develop proposed social-media mechanisms in a multidisciplinary white paper, describing each in a one-page table to emphasize accessibility and modularity.

    Source: Introduction · p. 7

  • Results

    The paper presents mechanisms that can be considered individually, with many offering complementary contributions to improving democratic discourse.

    Source: Introduction · p. 7

  • Takeaways & Limitations

    The white paper provides an accessible starting point for broader interdisciplinary discussion about improving social media for democratic discourse.

    Source: Introduction · p. 7

    Collective inference faces overconfident individual probability ratings and difficulty modeling dependencies among users’ judgments.

    Source: Promise · p. 26

Abstract

from arXiv · show

Social media have expanded opportunities for communication and political participation, but today's dominant platforms are optimized primarily for engagement and advertising revenue, contributing to concerns about polarization, misinformation, social isolation, and loss of civility. We explore how social media might instead be deliberately designed to support democratic discourse, collective deliberation, and collective intelligence. Drawing on literature across computer science, psychology, political science, and related fields, we present a modular collection of mechanisms that could be implemented individually or in combination. These include user-controlled and open recommender systems, tools for exposure to diverse perspectives, new forms of cognitive and epistemic feedback, collaborative and AI-assisted fact-checking, privacy and visibility controls, mechanisms for improving civility and evidentiary integrity, and reputation systems that reward high-quality participation. The proposals are intended both as a practical menu of design possibilities and as a starting point for broader interdisciplinary discussion about digital public spaces designed around democratic values rather than engagement alone.

Introduction · How a user discovers and actively curates content exposure

The introduction argues that social media’s democratic promise is constrained by platform designs and incentives that prioritize advertising revenue, and presents a modular vision for redesigning digital public spaces. It frames the next section as exploring how users discover and actively curate content exposure.

Introduction

  • Social media can expand participation, collective intelligence, and social movements, but research also associates digital media with polarization, misinformation, isolation, and declining political trust.A review of 496 articles found associations with greater political participation and information consumption in some contexts, alongside declining political trust, rising populism, and increasing polarization, especially in established democracies.Source: Introduction · p. 5
  • These outcomes are not presented as inevitable: platform game design can amplify moral-emotional expression and in-group signaling, so existing architectures warrant re-evaluation.The authors call for more work on platform designs that encourage community and collective problem-solving.Source: Introduction · pp. 5–6
  • Ad-revenue incentives drive platforms to maximize attention and targeting, while democratic discourse requires engagement grounded in meaningful exchanges of facts and opinions.The paper describes platforms’ extraction and monetization of behavioral data as “surveillance capitalism”.Source: Introduction · p. 6
  • Although platforms can optimize for diversity, fairness, well-being, factual accuracy, and long-term benefit, narrow economic incentives constrain such alternatives (Stray et al. 2021).The authors argue that ad-revenue-driven shortcomings do not establish inherent limits on connecting people at global scale.Source: Introduction · p. 6
  • The authors argue that prosocial, democratic design could build long-term trust and satisfaction, support sustainable engagement, and potentially contribute to commercial success (Weyl et al. 2025).They call for participation from industry, people outside its incentive structure, and experts across the humanities, sciences, and engineering.Source: Introduction · pp. 6–7
  • This paper develops an alternative design vision; choosing an alternative funding or business model is outside its scope.The introduction lists public-service, nonprofit, subscription, cooperative, and protocol-and-service models as possibilities.Source: Introduction · pp. 6–7
  • The paper proposes modular mechanisms for improving democratic discourse, intended as an accessible entry point to existing research and a starting point for broader discussion.Mechanisms are presented individually in tables and can be considered in isolation or combined.Source: Introduction · p. 7

How a user discovers and actively curates content exposure

Virtual travel through geographic location filtering … Natural language feed algorithm tuning

The paper proposes design mechanisms that let users explore local perspectives across places, find arguments by stance, curate feeds around community values, and tune algorithms through natural-language instructions. These mechanisms aim to make feed and search choices more responsive to users’ interests and preferences.

Virtual travel through geographic location filtering

  • A clickable map could let users explore live local news and conversations in specific places, sampling cultural and political perspectives beyond their own region.The feature is intended to connect users with people in other cities, countries, continents, and cultures, including regions affected by catastrophes and wars.Source: Description · p. 9, Promise · p. 9
  • Because location tags are sparse and unevenly distributed, platforms could infer locations from text, images, or social connections, while carefully accounting for rural gaps, spoofing, and travel.The passage estimates that 3% of Twitter and 25% of Instagram content is geotagged, while machine-learning approaches can extract location from 50–90% of untagged items.Source: Promise · p. 9

Stance-based search

  • Stance-sensitive search could help users find arguments for and against particular positions and optionally curate positive, negative, or neutral opinions.Current searches generally match specific words rather than the stance being advocated, making position-based searches difficult.Source: Stance-based search · p. 10, Promise · p. 10
  • Implementing stance-based search may require novel algorithms, such as using post embeddings, and users could be allowed to opt out of having their posts surfaced this way.Source: Promise · p. 10

Communal feed curation

  • Communities could curate and publicly share feeds based on shared values, letting others subscribe and making the content and values shaping those feeds more explicit.Unlike groups that require content to be authored by members, these communities could range from niche interests to broader intellectual cohesion.Source: Description · p. 11, Promise · p. 11

Natural language feed algorithm tuning

  • Users could tune their feeds with editable natural-language instructions and adjust each instruction’s effect, prioritizing conscious preferences over habitual engagement.The proposed log includes “up” and “down” controls to correct over- or under-adjustment and could help downregulate addictive, low-quality content.Source: Description · p. 12, Promise · p. 12
  • Reliable LLM-tunable filters may be challenging to design, and existing feedback and reaction mechanisms may not adequately steer algorithms toward community values.Source: Promise · pp. 11–12

Open-source forkable recommenders · Daily algorithm ballot · Content-engagement summary and feed recalibration

These proposals give people and communities more influence over feed ranking through algorithm choice, daily ballots, and summaries that can support user-directed recalibration. They also identify practical risks, including privacy, participation, and the difficulty of measuring ideological diversity.

Open-source forkable recommenders

  • Users can install audited open-source recommenders and compare counterfactual feeds, making algorithm choice, tuning, and ranking differences visible (Yang et al. 2018; Salah et al. 2020; Tan et al. 2021; Hosseinmardi et al. 2024).The counterfactual views show how different objectives change what users read.Source: Description · p. 13

Promise

  • Competing recommenders could spur innovation and independent audits, build literacy about amplification, and reduce dependence on a single platform.These benefits depend on users being able to understand and act on the available choices.Source: Promise · p. 13
  • Forkable recommenders risk confusing users, exposing data to third parties, and attracting few switchers without clear benefits.Suggested mitigations include preset bundles or simple toggles, sandboxed APIs and mandatory audits, and preset options or social proof.Source: Promise · p. 13

Daily algorithm ballot

  • An optional daily ballot could aggregate community feed preferences and adjust ranking settings for the next 24 hours while retaining individualized follows and interests.Users could vote on priorities such as chronological order, ideological diversity, or calmer conversation; reach and effect duration could be adjustable.Source: Description · p. 14

Promise

  • The ballot could make feed curation more transparent, empower users, broaden perspectives, and reduce viral pressure.Source: Promise · p. 14
  • Ballot design must address low participation, vote manipulation, possible engagement decline, and complexity through one-click access, verified-account caps, personalized-content baselines, and plain language.Source: Promise · p. 14

Content-engagement summary and feed recalibration

  • Engagement summaries could show viewed, posted, or engaged-with content, including ideological diversity (Munson et al. 2013), time spent doom-scrolling, and topics in or outside the user’s experience.Users could inspect how the algorithm categorizes them, correct misconceptions, ask questions about their habits, and access summaries on demand or periodically.Source: Description · p. 15, Promise · p. 15
  • Users could set goals such as consuming more politically diverse content and use gaps between goals and actual engagement to recalibrate feeds or support shared accountability.Ideological diversity is difficult to quantify, and distinguishing doom-scrolling from productive scrolling is challenging.Source: Description · p. 15, Promise · p. 15

Stream networks … Challenges

The section proposes user-owned stream networks for decentralized content curation and lists contextualization tools, including an opt-in cross-lingual bridge. It describes potential benefits alongside implementation needs and translation risks.

Stream networks

  • Each user can create and control multiple streams that mix incoming feeds and select posts using priority, recency, content, and rate constraints.Streams may be specified through domain-specific language or visual programming and use machine learning or AI to improve selections.Source: Description · p. 16
  • Stream networks distribute human curation and algorithmic selection across users, forming a decentralized public sphere rather than concentrating attention in a platform ranking system.Streams recursively combine user-owned curation with automated selection, and the network is framed as collective intelligence.Source: Description · p. 16

Promise

  • The stream-network proposal is a general framework for user control over streams, community sharing, and allocating time across interests.Its implementation requires programming environments, more accessible stream-design tools, and interoperability with the traditional social web.Source: Promise · p. 16

Content contextualization & understanding

  • Contextualization proposals aim to help users process information and avoid echo chambers through translation, opinion mapping, discourse graphs, timelines, cross-ideological matching, and opinion clustering.Named ideas include a cross-lingual bridge, live opinion mapping, a conversation graph, timeline generation, MatchAcross, and interactive bubble-chart opinion clustering.Source: Content contextualization &… · p. 17

Cross-lingual bridge layer

  • An opt-in translation layer can automatically translate posts and comments outside users’ fluent languages, with granular rules and inline or split-pane originals.Confidence badges, crowd edits, cultural notes, shared glossaries, and low-latency subtitles for live contexts are proposed to improve readability and translation context.Source: Description · p. 18

Promise

  • The cross-lingual bridge is intended to encourage routine cross-language reading, preserve nuance and accountability, and enrich deliberation across diaspora and global conversations.Community edits are proposed to improve machine translation quality and cultural context over time.Source: Promise · p. 18

Challenges

  • Translation risks include distorted meaning, slang or dialect gaps, interface complexity, mistranslation trolling, and latency in live contexts.Proposed mitigations include crowd fixes and confidence labels, community glossaries, smart defaults, moderation and version history, and progressive refinement or caching.Source: Challenges · p. 18

Live opinion mapping for current events … MatchAcross: connecting users across ideological divides via social proximity

The sections propose tools for navigating current-event opinions, structuring debate, tracking the evolution of public issues, and encouraging cross-ideological conversations through social connections. They describe potential benefits alongside challenges such as real-time classification, mislabeling, and the risk of oversimplifying complex or changing information.

Live opinion mapping for current events

  • A live opinion map centralizes posts about developing news stories so users can explore perspectives using filters such as political leaning, controversy, location, humor, and modality.Classifying potentially millions of posts in real time is technically challenging; ML-based clustering algorithms are suggested as a possible approach (Knittel et al. 2022).Source: Description · p. 19, Promise · p. 19

Conversation graph for meaningful discourse

  • Graph-structured conversations branch replies into supporting points, counterarguments, or clarifying questions, making relationships between ideas and diverse viewpoints easier to follow.AI-suggested labels could reduce the effort of tagging replies, while deliberate mislabeling and sprawling graphs remain challenges; low-activity branches could be collapsed.Source: Description · p. 20, Promise · p. 20

Timeline generation for controversial societal issues

  • Timelines for major public debates organize key events, policy changes, viral posts, and turning points chronologically, with source links and concise summaries of multiple viewpoints to help users catch up.The proposal aims to reduce misinformation and improve civic literacy, but complex events risk oversimplification or factual errors.Source: Description · p. 21, Promise · p. 21
  • Timeline entries should include verifiable citations, and rapidly evolving or disputed entries can be marked under review or updated and reviewed through crowdsourcing.These safeguards address the risk that changing facts or interpretations, or summarizing heated debates, could produce mistakes.Source: Promise · p. 21

MatchAcross: connecting users across ideological divides via social proximity

  • The proposed AI system estimates users’ positions from liked posts, viewed content, and friend networks, then maps them and suggests guided conversations with people holding different views and a social connection.Source: Description · p. 22
  • MatchAcross visualizes users’ positions relative to others and proposes flexible, guided interactions within or just beyond their networks, such as with a friend-of-a-friend.The proposal notes that positions inferred from limited signals may be difficult to estimate, and mislabeling or ranking users could undermine trust.Source: Description · p. 22, Promise · p. 22

Promise

  • The proposed cross-ideological interactions aim to encourage empathy and understanding rather than conflict, while visualizing users’ positions may prompt reflection on their views and biases.Users may nevertheless feel labeled or ranked, and inaccurate matches could cause mistrust.Source: Promise · p. 22

Interactive bubble-chart opinion clustering … Cognicons

The section proposes opt-in opinion maps that let users visualize and discuss differing stances, alongside user-selected cognicons for sharing cognitive responses. It describes potential benefits and practical limits of both mechanisms.

Interactive bubble-chart opinion clustering

  • Opt-in bubble charts let users map their stances, explore opinion clusters, chat with other users, and update their positions after conversations.Users place their own dots on issue dimensions; they can inspect profiles and use one-on-one or group chats to explore shared or differing views.Source: Description · p. 23
  • Creators choose chart axes, or AI can suggest editable dimensions by matching a post’s claims and entities to opposing-value pairs.The AI uses NLP and embedding similarity; charts can use one or multiple dimensions.Source: Description · p. 23

Promise

  • The charts could complement binary reactions, make opinion homogeneity and outliers visible, and foster middle-ground conversations.Source: Promise · p. 23
  • Privacy controls let users choose who sees their dots and whether to accept chats; charts can be delayed until N responses or T time has passed.The delay addresses posts with too little engagement to reveal meaningful clusters.Source: Promise · p. 23

How the user interacts with content while minimizing bias

  • Cognicons let users annotate posts with cognitive responses, such as judging content correct, incorrect, or interesting despite disagreement, rather than relying on algorithmic or moderator labels.The section lists Nikolaus Kriegeskorte and Trenton Jerde alongside Cognicons.Source: How the user interacts with content… · p. 24, Description · p. 25

Engaging & reacting

  • Cognicons offer one-click cognitive feedback that may support thinking together and viewpoint diversity, counterbalancing emoticons’ emotionalizing effect.They can make it easier to flag false information, share quality content without endorsing it, and reward thought-provoking arguments from political opponents.Source: Promise · p. 25, Description · p. 25

Cognicons

  • Too many cognicons could confuse users, so their number should balance nuance with simplicity.Source: Promise · p. 25

Probability ratings … Thread-privacy toggle

The section proposes modular tools for assessing claims, encountering opposing views, reducing public popularity cues, and controlling who can see posts. These mechanisms aim to support more informed and less performative discussion, while each has practical limitations.

Probability ratings

  • Users can add one-click 0–1 probability ratings to posts, with readers seeing full or trusted-associate distributions and aggregates that account for viewpoint representation or user calibration.Possible aggregates include a mean, logit mean, or probabilistic collective inference.Source: Description · p. 26

Promise

  • Continuous ratings offer nuanced, democratically legitimate crowd judgments, but overconfidence, dependency, and miscalibration complicate collective inference; Stinson et al. 2025 describes scalable algorithms.Calibration can be estimated and corrected, while truth labels can help calibrate collective inference.Source: Promise · p. 26

Devil’s advocate button

  • A devil’s advocate button can surface relevant, quality-selected alternative posts or articles, generate an AI response, or summon registered advocates to argue an opposing position.Summoned advocates could be rated for respectful conduct, rate the summoning user, and have the feature limited if they are disrespectful.Source: Description · p. 27

Promise

  • The button offers a one-press escape from echo chambers and an educational aid, though finding and selecting alternative views may be computationally challenging.Source: Promise · p. 27

Number-free feed

  • A number-free feed hides public engagement counts while retaining private reactions as feedback for creators and signals to refine what users see.The proposal aims to shift attention toward substance, reduce crowdthink and popularity barriers for newcomers, and make inflating counts less useful to bots.Source: Description · p. 28, Promise · p. 28

Promise

  • Hiding counts may make good content harder to find, obscure ranking if counts still affect algorithms, reduce creator proof for sponsors, or lower posting motivation.User-applied tags and creator analytics dashboards are suggested as mitigations.Source: Promise · p. 28

Thread-privacy toggle

  • Post- or thread-level audience settings—from any user to conversation participants—can limit pile-ons and let users discuss risky opinions without anonymous accounts.The settings offer finer-grained control than blanket private-account options and may avoid performative dialogue; private posts can still be screenshotted, and visible indicators are suggested.Source: Description · p. 29, Promise · p. 29

Accuracy & civility … Collaborative fact-checking

The section proposes design mechanisms for healthier democratic discourse and describes how platform rules can shape status-seeking behavior. It presents transparent, community-based fact-checking as a way to expand verification and shared responsibility for truth.

Tools for collaborative filtering and maintaining ecosystem health

Game-aware design for democratic discourse

  • Because social status rewards depend on design, ad-hominem flags, lower post visibility, cognicons, and probability ratings could favor thoughtful, accurate contributions.The passage frames probability ratings as a possible basis for status linked to accurate, well-calibrated judgments.Source: Description · p. 31

Promise

  • Democratic design should avoid heavy-handed scoring and consider how free expression, deliberation, collective intelligence, rights, and participation may trade off.Source: Promise · p. 31

Collaborative fact-checking

  • Collaborative fact-checking lets users investigate claims, add context and sources, and vote on accuracy, expanding capacity, trust, media literacy, and shared responsibility for truth.The passage also says the approach may slow misinformation and make verification open to scrutiny.Source: Description · p. 32
  • Collaborative checks need independent review and monitoring; expertise can be made visible through special status, but expert input should not be absolute.Without regular audits, low-quality contributions may be rewarded.Source: Promise · p. 32

AI fact-checking … Resolving popularity and majority bias

The section proposes AI-supported tools for evaluating claims and discourse, combining eyewitness footage, and ranking content by merit rather than popularity or majority appeal. These mechanisms include user choice and safeguards, while raising challenges around credibility judgments, bias, consent, and fairness.

AI fact-checking

  • AI can identify claims in posts and media and link them to academic papers, fact-checking databases, or expert-reviewed articles, with user-selected models exposing varied perspectives.The system is intended to support real-time credibility assessment, but users must ultimately judge source credibility and fact-checking remains time-consuming.Source: Description · p. 33, Promise · p. 33

Ad-hominem detection

  • Ad-hominem flags could be human- or AI-generated and binary or continuous, helping users assess discourse and rank feeds to avoid personal attacks.Ad-hominem judgments can distract from substantive discussion, intimidate participants, and limit their expression.Source: Description · p. 34, Promise · p. 34
  • Ad-hominem detection requires judgment: moderators may lack democratic legitimacy and scale, crowd judgments can be weaponized, and AI needs bias monitoring.Source: Promise · p. 34

Crowd media collage synthesizer

  • A platform could cluster eyewitness clips by time and place, synchronize them into multi-angle reels, preserve source provenance, and let uploaders opt out or reorder their angle.The proposal gives an example clustering window of roughly 200 meters and plus or minus two minutes; newsrooms and fact-checkers could access the synchronized package through an API.Source: Description · p. 35
  • Synchronized reels could replace redundant uploads and limit selective-framing bias, but the system must address sync errors, consent, GPU costs, takedowns, and deepfake risks.Source: Promise · p. 35

Resolving popularity and majority bias

  • A voluntary anonymous evaluation feed could conceal creator identities, give posts equal visibility, then boost visibility based on performance while promoting content valued within distinct communities.The proposal aims to reduce popularity bias and cultural majority bias by judging content on merit and recognizing locally valued posts alongside broadly appealing ones.Source: Description · p. 36, Promise · p. 36
  • Balancing global and local visibility is challenging, and content evaluation is multifaceted.Source: Promise · p. 36

Selective-scope immutable stream ledger … Graded control of thread privacy and post visibility

The sections propose mechanisms to preserve authenticity, limit coordinated manipulation, represent community opinion distributions in feeds, and give users control over post reach and audience. Each mechanism balances its intended benefits against practical or behavioral risks.

Selective-scope immutable stream ledger

  • The Selective-Scope Ledger activates for risky keywords, rapid virality, uploader-selected Evidence mode, or flagged locations and periods, then logs signed video-tile hashes and timestamps.The footage itself moves to cold storage after 30 days; ordinary uploads remain editable without an audit trail. The ledger is intended to support verification while preserving privacy, but benign videos may be caught by its heuristics and attackers could hash inauthentic content.Source: Description · p. 37, Promise · p. 37
  • Ledger tradeoffs include high-volume write and audit demands, possible false locks on benign videos, and interface clutter from displaying full edit histories.Subtree-optimized storage, context classification, and highlighting only main changes are proposed responses.Source: Promise · p. 37

Automated detection of filter algorithm manipulation

  • Fake-account defenses combine credibility scoring, anomaly detection, reduced influence for new or low-trust accounts, and network analysis to limit coordinated influence.The proposal aims to preserve crowd-driven signals and reduce exposure to biased content, while facing false positives, privacy and fairness concerns, scalability, and transparency challenges. Radivojevic et al. (2024) show that AI-driven bots can closely mimic real users and spread biased messages at scale.Source: Description · p. 38, Promise · p. 38

Population-representative feeds

  • Population-representative feeds can rebalance posts either across all feeds or within each user’s feed, with continuous user control over rebalancing and choice of reference population.The design seeks to correct distorted impressions of community opinion, but requires modeling political views and may be disliked; user control and all-feed balancing are proposed to address concerns that exposure to opposing views can polarize (Bail 2021).Source: Description · p. 39, Promise · p. 39

Graded control of thread privacy and post visibility

  • Users can set maximum reach by default or per post, limiting shares, reshares, audience size, visibility lifespan, or access to specified groups.This offers a middle ground between public posts and private messages, and may reduce self-censorship or performative statements; limited-reach posts can still be screenshotted and shared.Source: Description · p. 40, Promise · p. 40
  • Lifespan-limited posts may help bad-faith actors evade accountability, while statistics about such posts could help identify them.Source: Promise · p. 40

Reputation & community rewards … Milestone badges for high-quality participation

The section presents reputation mechanisms designed to reward thoughtful, evidence-based, and prosocial participation. It describes discourse scores and milestone badges, alongside risks including narrow definitions, gaming, and inconsistent standards.

Reputation & community rewards — Long-term incentives that encourage prosocial behavior

Discourse score

  • A discourse score could reward credible sourcing, engagement across ideological differences, and publicly revising views in response to compelling evidence.Both users and automated systems could assign points for these contributions.Source: Description · p. 42

Discourse score — Promise

  • The score aims to shift incentives from popularity toward thoughtful arguments, making well-reasoned posts more visible and reliable voices easier to spot.Source: Promise · p. 42, Description · p. 42
  • A discourse score may narrow good discourse to one measure, invite gaming, or bias visibility toward inoffensive centrist views.Continual adjustment to account for viewpoint diversity and richer notions of good discourse may help address these risks.Source: Promise · p. 42

Milestone badges for high-quality participation — Description

  • Milestone badges recognize contributions such as well-sourced corrections, civil conversations across divides, and high-quality summaries or syntheses.Visible badges can help users identify trustworthy participants and filter feeds, while discouraging fake engagement.Source: Description · p. 43

Milestone badges for high-quality participation — Promise

  • Badges are intended to encourage prosocial participation, highlight credibility, filter low-quality or fake accounts, and increase engagement.Source: Promise · p. 43, Description · p. 43
  • Badge systems may reward quantity over quality, invite loophole-seeking, or lack clear requirements and consistent verification.Source: Promise · p. 43

AI civility toolkit

The AI civility toolkit aims to prevent harm before it occurs while supporting users’ self-regulation. It combines pre-posting assistance and vulnerability protections with adaptive pacing, and its proposed benefits come with design risks and safeguards.

Description

  • The toolkit intervenes before harm through an optional LLM reframer, vulnerability shield, and affect-aware pacing, complementing moderation and supporting users’ self-regulation.It can suggest kinder wording, request evidence, flag potential misunderstandings or manipulation, and rate-limit persuasive content in Guarded Mode; modules offer Light, Balanced, and Strong presets and run on-device when possible.Source: Description · p. 44
  • Description modules include evidence prompts, tone checking, adaptive load balancing, civility statistics, and a vulnerability shield.The stated purposes include nudging users to add sources, highlighting ad-hominem attacks, slurs, or sarcasm, and prompting breaks from emotionally intense or repetitive content.Source: Description · p. 44

Promise

  • The toolkit is proposed to convert heated drafts into constructive dialogue, shield vulnerable users, mitigate doomscrolling, reduce moderator workload, and aid neurodivergent and multilingual users.Source: Promise · p. 44
  • Its promise depends on opt-outs, continuous auditing, on-device inference, diverse training data, and robustness tests to address annoyance, context errors, privacy, cultural bias, and prompt gaming.Source: Promise · p. 44

Glossary

The glossary establishes shared terminology for discussing how social media systems shape information circulation, exposure, participation, and democratic discourse.

Information circulation

  • Messages are discrete information units, sharing circulates them across communities, popularity signals indicate attention or approval, and algorithms can amplify or suppress visibility.Amplification increases reach beyond organic sharing; suppression hides or de-prioritizes content in feeds.Source: Glossary · p. 45

Transparency and exposure

  • Societal and user transparency concern platform impacts and feed explanations, while exposure diversity measures the variety of content users choose relative to what is available.Societal transparency can include public and researcher access to data on information flows and algorithmic effects; user transparency explains why content appears and how the algorithm works.Source: Glossary · p. 45

Nudges

  • Accuracy and civility nudges subtly shape behavior; gatekeeper’s nudges rank for intellectual diversity, while model citizen nudges prioritize content encouraging civic virtues.The glossary describes nudging as small changes in how choices are presented that can influence decision-making.Source: Glossary · p. 45

Diversity and segregation

  • Serendipity recommendations seek surprising but relevant content, whereas filter bubbles omit or de-prioritize viewpoints and cyberbalkanization segregates users into intellectually homogeneous groups.Fu et al. (2023) identify unexpectedness, novelty, diversity, and relevance as criteria for serendipity recommendations.Source: Glossary · pp. 45–46

Ranking and diversity

  • The diversity paradox describes users not feeling more diverse exposure despite diversity-based recommendations; bridging-based and intelligence-based ranking instead prioritize cross-perspective value and likely accuracy.Joris et al. (2024) attribute the paradox to recommendation algorithms’ invisibility. Bridging-based ranking values content approved by intellectually diverse users, while Burton et al. (2024) propose ranking content likely to support accurate beliefs; cognitively diverse opinions may also support collective accuracy.Source: Glossary · p. 46

Polarization

  • Interactional polarization concerns users interacting mainly with like-minded people, positional polarization concerns increasingly distinct political beliefs, and affective polarization concerns hostility toward those with different views.Source: Glossary · p. 46

Online learning

  • Regret measures cumulative deviation from an optimal action in online learning; bandits choose among reward-or-loss options, and contextual bandits make such choices given context.A no-regret policy has regret that grows sublinearly over time and therefore converges to the optimal policy; contextual bandits are used in recommender systems.Source: Glossary · pp. 46–47

Privacy, trust, and reporting

  • Machine unlearning removes data-point influence from trained models; reputation thresholds limit new or low-trust accounts, sock puppets fabricate support, and crowdmapping aggregates geolocated reports.Machine unlearning is used for privacy, safety, or stale knowledge; crowdmapping visualizes group-submitted location data, often for crisis response, citizen journalism, or community reporting.Source: Glossary · p. 47
Loading 2610.10816v1…