Deliberative Democratic Capabilities Gap Map v0.5 (Early Preview)

AI & Democracy Foundation [show contributors]

Introduction

We believe that high-quality deliberative democratic processes for decision-making on AI can help mitigate some of the risks emerging with this new technology—and enable us to maintain democracy in a world increasingly dominated by AI. But there is a lot of work to be done!

AI presents an opportunity to turbocharge improvements to our democratic practices. However, key actors new to the space are unsure what to build, what to fund or what kinds of collaborations would be most helpful.

This map aims to help answer those questions. It highlights what we believe are critical capabilities for democratic processes to address needs for democratic governance of AI, mitigate power concentration, and facilitate cooperation (even internationally). For more context, see our Democracy Levels Website. These capabilities are valuable for decision-making across the many other challenges that we face across every level of society.

What is a capability? A measure of the global community of practice's ability to run processes that can perform key functions required of good enough deliberative processes.

What is the goal? Key actors, making consequential decisions (especially on AI), have access to processes that are representative, informed, substantive, deliberative, robust, and legible. These actors might be governments, multinational partnerships, global regulators or corporations. The deliberative processes will vary depending on their purpose and context. We need the toolbox for each of the possible combinations.

We intend to expand this map to include ecosystem infrastructure that will help accelerate the development and improvement of these capabilities, including for their application in innovative corporate governance models.

Please note: This is a work-in-progress draft. Both the format and the content are incomplete; this is a quiet soft launch to ensure that those who have asked for a shareable link can access and reference this ASAP, but the format will likely change significantly before we do a more public v1 launch. There are many missing directions, products, tools, etc. that we would like to add, and contributions are very welcome. Please leave suggestions via suggested edits here or via this form.

Our roadmap includes creating a much better interface for navigating this, and making this more useful for audiences across the landscape of AI and democracy (e.g., filters for relevance to other democratic and decision-making processes, for ML research, AI alignment, democratic/collective alignment, etc).

What processes are we focusing on?

The primary current focus is on representative deliberative processes (like citizens' assemblies), but most of the underlying capabilities are general across any kind of democratic process (e.g., ensuring representation, forecasting impacts, providing sufficiently substantive outputs, etc.)

We've focused primarily on deliberative democracy and related approaches (for now) because of the unique properties that make such processes useful for decision-making about AI, particularly for decisions around issues that cross jurisdictional divides, where there is an urgent need and the least capability.

Deliberative democratic processes can work with jurisdictions of arbitrary size, infrastructure, and political structure (including globally), and can effectively incorporate the knowledge of diverse participants and subject matter experts. They're useful for making decisions on concrete trade-offs, including relating to AI values and safety. Additionally, many of the capabilities generalize to other kinds of democratic and coordination processes, such as improving epistemics. They have been used at every level of government around the world (OECD Deliberative Democracy Database, 2023) and show immense promise for bringing diverse groups of people together and facilitating their finding of common ground on complex policy decisions. However, as a relatively recent innovation in the modern world, they remain less mature than more established governance approaches.

How to use the map

The gap map is categorized by high-level dimensions that describe key outcomes of deliberative processes, for example, that participants become informed. Within each dimension is a subset of capabilities that contribute to delivering these outcomes, for example, that we can curate the relevant context used to inform participants. Not all of these capabilities will be relevant to every process, but this is meant to be a relatively comprehensive set of the capabilities likely to be necessary.

The goal of the mapping process is effectively: What challenges need to be overcome for us to be able to deliver the best possible processes, regardless of different subjective process design decisions?

We're not asserting any claims here around what deliberative processes should be doing, only identifying and sharing obstacles and open problems.

We've made initial rough assessments of the ecosystem's current capabilities across the following criteria:

Capability Level: ~Current level of performance/quality
Urgency: Comparative estimate of R&D priority. Not an assessment of the importance of the capability.
Opportunity: ~Room for improvement
Consistency: ~How widely accessible is best practice?
Global: How well can this currently work for a global/transnational process?

We then provide context for each capability, including examples of how they're performed now, cutting-edge research, products, and practices, before moving on to concrete gaps.

Gaps are defined as our inability to perform or meet goals that reasonably contribute to the ecosystem's overall capabilities. These are specified as either open research problems or specific product cards that scope the necessary technical or organizational infrastructure.

How to contribute

We intend for this map to be a living document, developed through the contributions of others. We have two channels for contributing:

A Google Document for making specific suggested edits and contributions. For example, this is useful when suggesting missing products, research, and practice examples.
A submission form for making more wide-ranging contributions. For example, this is useful for general feedback or for making large contributions that might impact many different points or the entire map.

Our team will review contributions and update the map periodically. All contributions will be acknowledged in the list above.

Want to stay up to date with significant changes? This work is meant to evolve as a shared resource. We encourage you to:

Stay connected as updates are made by adding your details to this form and by subscribing to general AI & Democracy Foundation updates.
Help strengthen the work by suggesting any relevant research, examples, or case studies that could reflect existing practice, products, and research.

We also invite researchers and teams to share their interest in addressing one of these gaps by reaching out to capabilities@aidemocracyfoundation.org, so that emerging efforts can be made visible and potential collaborations can take shape.

Representativeness

The extent to which key decisions are representative of the constituent population.

To what extent: (1) is there sufficient representation at critical parts of the process, including (a) proposing decisions, and (b) making ultimate decisions? (2) Are there barriers leading to bias in representation?

Capability	Who & How They Do It	Existing Practice, Products and Research	Gaps
Select participants Fair Selection Participant selector Ability to fairly select participants according to some definition of representation. Opportunity: Low Urgency: High Capability: Good Consistency: Good Global: Adequate	Process organizers use existing algorithms and tools to implement the normative selection criteria that they set out. 100+ citizens’ assemblies around the world at every level of governance have used these. They take a pool of willing participants and select a final ‘panel’ that adheres to set criteria.	Research Sortition algorithms have been designed to select participants according to quotas, balancing representativeness, fairness and manipulation resistance (Flanigan, 2021; Baharav, 2024). Research Federated Assemblies research explores algorithms for ensuring representation guarantees across subsidiary panels (Halpern, 2024). Research Sortition tools have been designed to support practitioners with things like selecting the optimal set of alternate participants (Assos, 2025) and tuning stratification quotas (Baharav, forthcoming). Product Panelot. Soon to be redeveloped into: Lottery Lab.	Goal: Deep intersectional representation to more closely match the constituent population (even with high-quality selection pools). Research What are the limits of intersectional representation in sortition algorithms, and where are trade-offs most present? Goal: Stratification approaches that best balance desiderata such as diversity, inclusion and the legitimacy of panels. Research What is the relationship between different selection variables and public trust, deliberative quality, epistemic quality and output quality? Research In which contexts does trust increase with reweighting (e.g. to account for affectedness, power imbalances, history, etc.)? Goal: Managing representation guarantees across iterative and parallel panels. Research Representation across iterations of panels, or many panels in parallel. Goal: Product Tools to better understand the impacts on outcomes and trade-offs implicit in making specific design decisions, such as the depth or kinds of intersections, types of representation, and fairness.
Reach participants Diverse recruitment Informed recruitment strategies Ability to reach potential participants (e.g., to mitigate biases around self-selection, who is reachable, etc.). Opportunity: High Urgency: High Capability: Limited Consistency: Variable Global: Poor	Panel providers typically maintain large pools of pre-recruited participants from which they draw. They may actively recruit to fill any gaps, as these lists are typically unrepresentative of the wider population. Citizens’ assembly process organizers commonly recruit with 2-stage democratic lottery, which involves a round of mailed invitations and then a round of stratified random selection. There is also some use of door-knocking, random phone number dialing, and other recruitment methods. In the global south, where infrastructure that these approaches rely on can be scarce, alternative in-person methods have been used.	Experimental Practice The 2021 Global Assembly designed and trialled a method for global recruitment. Experimental Practice The Sortition Foundation has a network of collaborators to conduct transnational European recruitment (SORT-EU). Partners are responsible for regional or country-based recruitment, and the approaches vary between partners. Research Iswe and Participedia have hosted a workshop on the issue. Research [Ali Cirone’s empirical work on self-selection biases, she presented it here, but not yet a public version].	Goal: There are cheap and efficient ways of recruiting participants that can account for differences in response rates to different types of invitation methods (mail, phone, door-knocking, etc). Goal: Product Automated Recruitment Tool Research What kinds of recruitment methods reach which kinds of people? Practice Given a budget, a location, panel size, and unique quotas, design a recruitment plan that will maximize response rates and the representativeness of the sample. What are the most efficient ways of recruiting participants? Goal: Response rates that enable selection algorithms to accurately select panels that represent the whole population by mitigating the impacts of self-selection biases. Context: Currently, response rates are often low, averaging ~2-5%. If the selection pool is missing groups or seriously unrepresentative, the efficacy of selection algorithms is limited. Research What are the best approaches to recruiting a participant pool that captures the complexity and intersections of society while minimising self-selection biases? Research What strategies can be used to motivate participation in less-democratic contexts? Practice How can we handle the real-world failure modes of recruitment? Goal: Recruitment strategies that accommodate poor databases for identifying and accessing people, locations with no mail access, or poor access to the internet. Research How can we quantify the fairness of different approaches to sampling the population? Practice How best to do global sortition given limited resources or access to population data? Practice Managing recruitment in geographies with incredibly poor access and digital and physical infrastructure.
Support participation Accessibility Participant concierge Ability to provide accessible, welcoming and compelling processes enabling diverse participation. Opportunity: Moderate Urgency: High Capability: Good Consistency: Variable Global: Limited	Process organizers provide basic process information, additional context, and address accessibility needs proactively or when onboarding participants. Citizens’ Assembly organizers have a ‘concierge’ service for participants, which helps address any reservations, context setting, or hurdles that might come up (OECD, 2021), but this is difficult to scale. Online or predominantly digital processes will naturally exclude possible participants who have low levels of digital literacy or low levels of access to the necessary technology. Some processes have trialled providing on-hand support to provide this tech and help operate it, but this is resource-intensive.	Experimental Practice The 2021 Global Assembly trialled a ‘Community Host’ model, where in-country partners worked directly with members of the assembly to provide them a safe place to work, a computer, internet connection, and technical assistance if necessary.	Goal: Participants have all their questions and concerns answered promptly and in an encouraging, welcoming and ongoing manner. Context: People are not often asked to contribute to important public decisions, and they doubt their capability. Research How can we assist or automate onboarding and process troubleshooting to reduce the costs of inclusion? Research How can we reassure and provide participants with safe and productive spaces for deliberation in polarized or post-conflict settings? Goal: Participants in remote regions with limited education, technical capacity, or supportive infrastructure can meaningfully participate. Practice Accessible methods for enabling and actively supporting participation in all contexts. Goal: Processes can span different cultures, languages, and geographies, including many different time zones.
Simulate participation Fidelity Participant simulator, Representative Prediction Ability to simulate the interactions and decisions of actors (e.g., participants, stakeholders, facilitators, experts), subprocesses, or entire processes (e.g., for rapid process iteration). Opportunity: High Urgency: Moderate Capability: Poor Consistency: Weak Global: Poor	Process organizers can infer from recorded preferences for new contexts, but these inferences are usually human estimates and not supported by well-documented algorithms. Collective dialogue organizers can run processes with simulated participants, though more research is needed to resolve their fidelity.	Research Several have explored versions of this idea, under the keywords ‘AI-as-representative’ (Collins, 2023), ‘voting avatars’ (Grandi, 2018), and ‘virtual democracy’ (Kahng et al., 2019), ‘plurals’ (Ashkinaze et al., 2024) and ‘simulated deliberative democracy’ (Leike, 2023). Research General Social Agents (Manning & Horton, 2025) Research Policy Priority Inference (agent-based modelling to evaluate policy impacts and ordering). Research Moral considerations of substituting AI for human involvement in deliberative processes (Revel and Pénigaud, 2025). Research [Luke Thorburn’s forthcoming paper on the democratic use of inferred preferences].	Goal: Deep understanding of when, if at all, simulation can be helpful. Research For what uses, in what contexts and with what level of faithfulness is it helpful or appropriate to use simulations? What are the philosophical, moral, political, etc. implications? Goal: Deliberative processes can be tested and refined before implementation with real participants. Research How can we develop realistic simulation environments that accurately predict how different deliberative formats will perform according to different design choices? Research Can AI generate its own suggested changes and test them to search the latent space for optimal solutions? Research How can lessons from speculative execution and speculative decoding help increase the availability of deliberative processes through reduced costs? Research Solving the technical blockers to effective and truth-worthy multi-agent simulation and modelling. Goal: Simulations are faithful enough to be relied upon when decisions are needed rapidly (e.g., seconds, minutes, hours). Research What are the best methods to measure the faithfulness of simulations? Research What hybrid approaches can combine fast simulation with selective human input to optimize both speed and accuracy for urgent decisions? Goal:Product Simulator/process sandbox
Voiceless perspectives Representation Stakeholders, Representatives Ability to fairly include the perspectives of those that are not represented in the process, including people who are not present (future generations, young people or other representation constraints), and non-human entities (natural phenomena or animals). Opportunity: High Urgency: Moderate Capability: Limited Consistency: Very variable Global: Limited	Process organizers can ask participants to critically reflect on the hypothetical views of those not present. There are some exercises that prompt participants to consider or even role-play views that are not present in the room (Example).	Experimental Practice Animals in the Room and DemocracyNext. Experimental Practice Talk to the River by Dark Matter Labs experimented with using an LLM to represent non-human entities in deliberation. Research Extensive work in political theory and political philosophy exploring the opportunities and challenges of representing future generations and non-human animals in democratic processes (e.g., González-Ricoy and Gosseries, 2017).	Goal: Identify the most legitimate methods for representing non-human animals and future generations in deliberative processes. Context: There are epistemological questions around the accurate or fair representations of the beliefs or interests of non-humans/future generations. Research What are the different methods for representing non-humans/future generations, and how do these methods compare? Research How do we evaluate any moral, philosophical, or efficacy-based justifications for trying to include the voices of non-humans/future generations in deliberations? Goal: Effectively integrate important absent perspectives into the process. Context: It can be hard to identify which views or perspectives might be valuable that are not currently present. Research How do we best track and identify necessary voices that are currently missing? Research How can virtual and augmented reality technologies help participants experience perspectives of animals or future generations?
Aggregate / distill inputs Fair Participation Aggregator, Summarizer Ability to aggregate votes and distill more complex forms of open-ended input into outputs and decisions, in fair and understandable ways, such participants feel their contributions are meaningfully taken into account (and can ideally see how). Opportunity: High Urgency: High Capability: Adequate Consistency: Very variable Global: Limited	Citizens’ Assemblies organizers utilize small group work, which relies on moderated or self-led documenting and integration of inputs. Ultimately, voting is used to ‘end’ the conversation in place of finding consensus. These decisions are sometimes made using Likert voting and supermajority thresholds. In Deliberative Polling, ‘final’ outcomes are measured by poll. Collective dialogue tools such as Pol.is use bridging algorithms to cluster inputs, and Remesh uses elicitation inference to help identify bridging statements across a large number of inputs. Other online systems may not even try to surface consensus or bridging, but instead simply log the inputs of various participants.	Research Generative Social Choice (Fish et al., 2023, Boehmer et al., 2025) Research 'Generative CI' through Collective Response Systems (Ovadya 2023), Elicitation inference (Konya, 2022) Research Smoothed Analysis of Social Choice (Flanigan, 2023) Research Augmenting Polis with LLMs (Small, 2023). Product Collective response tools to digitally elicit, collect, and analyse inputs (e.g. Polis, Remesh). Product Some features of AI-based tooling to accompany face-to-face deliberation and help capture insights and surface analysis (e.g. Dembrane, DeliberAIde, Psi, Cortico) Product Some features of tools like YourPriorities, Decidim, Colectiv, in addition to standard survey tools.	Goal: Collective input of many form factors is quickly and meaningfully synthesized, fairly reflecting participant perspectives. Research How can we assist or automate the aggregation of deliberative input from diverse participants in real time, whilst maintaining nuance around minority perspectives? Practice Translating existing social choice research into practical methodologies with decision aides for matching process to context, such as identifying trade-offs between theoretical guarantees, speed, explainability and legitimacy in the eyes of participants, public and stakeholders. Goal: Distilled and aggregated content is integrated effectively into the process. Research What visualization and presentation methods can best communicate distilled content for engagement by participants/organizers? Research What aggregation approaches are best suited to different stages of a process? Goal: Develop actionable and standardized best practices for aggregation and distillation within deliberative processes. Context: Deliberation facilitators and self-led participant facilitation can use different aggregation methods, resulting in views being represented in vastly different ways. Research What are the best standardized models for ‘in-person’ (in real world or online) facilitators to follow as they collate and record views of participants? Research What standardized methods can we easily teach participants in self-led groups on how best to capture the relevant information? Goal: Participants can track how their contributions have informed the process’s outcomes. Research How can we best embed traceability and transparency into the aggregation/distillation process?

Capability

Who & How They Do It

Existing Practice, Products and Research

Gaps

Select participants

Fair Selection

Participant selector

Ability to fairly select participants according to some definition of representation.

Opportunity: Low

Urgency: High

Capability: Good

Consistency: Good

Global: Adequate

Process organizers use existing algorithms and tools to implement the normative selection criteria that they set out.

100+ citizens’ assemblies around the world at every level of governance have used these. They take a pool of willing participants and select a final ‘panel’ that adheres to set criteria.

Research Sortition algorithms have been designed to select participants according to quotas, balancing representativeness, fairness and manipulation resistance (Flanigan, 2021; Baharav, 2024).

Research Federated Assemblies research explores algorithms for ensuring representation guarantees across subsidiary panels (Halpern, 2024).

Research Sortition tools have been designed to support practitioners with things like selecting the optimal set of alternate participants (Assos, 2025) and tuning stratification quotas (Baharav, forthcoming).

Product Panelot. Soon to be redeveloped into: Lottery Lab.

Goal: Deep intersectional representation to more closely match the constituent population (even with high-quality selection pools).

Research What are the limits of intersectional representation in sortition algorithms, and where are trade-offs most present?

Goal: Stratification approaches that best balance desiderata such as diversity, inclusion and the legitimacy of panels.

Research What is the relationship between different selection variables and public trust, deliberative quality, epistemic quality and output quality?
Research In which contexts does trust increase with reweighting (e.g. to account for affectedness, power imbalances, history, etc.)?

Goal: Managing representation guarantees across iterative and parallel panels.

Research Representation across iterations of panels, or many panels in parallel.

Goal: Product Tools to better understand the impacts on outcomes and trade-offs implicit in making specific design decisions, such as the depth or kinds of intersections, types of representation, and fairness.

Reach participants

Diverse recruitment

Informed recruitment strategies

Ability to reach potential participants (e.g., to mitigate biases around self-selection, who is reachable, etc.).

Opportunity: High

Urgency: High

Capability: Limited

Consistency: Variable

Global: Poor

Panel providers typically maintain large pools of pre-recruited participants from which they draw. They may actively recruit to fill any gaps, as these lists are typically unrepresentative of the wider population.

Citizens’ assembly process organizers commonly recruit with 2-stage democratic lottery, which involves a round of mailed invitations and then a round of stratified random selection. There is also some use of door-knocking, random phone number dialing, and other recruitment methods. In the global south, where infrastructure that these approaches rely on can be scarce, alternative in-person methods have been used.

Experimental Practice The 2021 Global Assembly designed and trialled a method for global recruitment.

Experimental Practice The Sortition Foundation has a network of collaborators to conduct transnational European recruitment (SORT-EU). Partners are responsible for regional or country-based recruitment, and the approaches vary between partners.

Research Iswe and Participedia have hosted a workshop on the issue.

Research [Ali Cirone’s empirical work on self-selection biases, she presented it here, but not yet a public version].

Goal: There are cheap and efficient ways of recruiting participants that can account for differences in response rates to different types of invitation methods (mail, phone, door-knocking, etc).

Goal: Product Automated Recruitment Tool

Research What kinds of recruitment methods reach which kinds of people?
Practice Given a budget, a location, panel size, and unique quotas, design a recruitment plan that will maximize response rates and the representativeness of the sample.
What are the most efficient ways of recruiting participants?

Goal: Response rates that enable selection algorithms to accurately select panels that represent the whole population by mitigating the impacts of self-selection biases.

Context: Currently, response rates are often low, averaging ~2-5%. If the selection pool is missing groups or seriously unrepresentative, the efficacy of selection algorithms is limited.
Research What are the best approaches to recruiting a participant pool that captures the complexity and intersections of society while minimising self-selection biases?
Research What strategies can be used to motivate participation in less-democratic contexts?
Practice How can we handle the real-world failure modes of recruitment?

Goal: Recruitment strategies that accommodate poor databases for identifying and accessing people, locations with no mail access, or poor access to the internet.

Research How can we quantify the fairness of different approaches to sampling the population?
Practice How best to do global sortition given limited resources or access to population data?
Practice Managing recruitment in geographies with incredibly poor access and digital and physical infrastructure.

Support participation

Accessibility

Participant concierge

Ability to provide accessible, welcoming and compelling processes enabling diverse participation.

Opportunity: Moderate

Urgency: High

Capability: Good

Consistency: Variable

Global: Limited

Process organizers provide basic process information, additional context, and address accessibility needs proactively or when onboarding participants.

Citizens’ Assembly organizers have a ‘concierge’ service for participants, which helps address any reservations, context setting, or hurdles that might come up (OECD, 2021), but this is difficult to scale.

Online or predominantly digital processes will naturally exclude possible participants who have low levels of digital literacy or low levels of access to the necessary technology. Some processes have trialled providing on-hand support to provide this tech and help operate it, but this is resource-intensive.

Experimental Practice The 2021 Global Assembly trialled a ‘Community Host’ model, where in-country partners worked directly with members of the assembly to provide them a safe place to work, a computer, internet connection, and technical assistance if necessary.

Goal: Participants have all their questions and concerns answered promptly and in an encouraging, welcoming and ongoing manner.

Context: People are not often asked to contribute to important public decisions, and they doubt their capability.
Research How can we assist or automate onboarding and process troubleshooting to reduce the costs of inclusion?
Research How can we reassure and provide participants with safe and productive spaces for deliberation in polarized or post-conflict settings?

Goal: Participants in remote regions with limited education, technical capacity, or supportive infrastructure can meaningfully participate.

Practice Accessible methods for enabling and actively supporting participation in all contexts.

Goal: Processes can span different cultures, languages, and geographies, including many different time zones.

Simulate participation

Fidelity

Participant simulator, Representative Prediction

Ability to simulate the interactions and decisions of actors (e.g., participants, stakeholders, facilitators, experts), subprocesses, or entire processes (e.g., for rapid process iteration).

Opportunity: High

Urgency: Moderate

Capability: Poor

Consistency: Weak

Global: Poor

Process organizers can infer from recorded preferences for new contexts, but these inferences are usually human estimates and not supported by well-documented algorithms.

Collective dialogue organizers can run processes with simulated participants, though more research is needed to resolve their fidelity.

Research Several have explored versions of this idea, under the keywords ‘AI-as-representative’ (Collins, 2023), ‘voting avatars’ (Grandi, 2018), and ‘virtual democracy’ (Kahng et al., 2019), ‘plurals’ (Ashkinaze et al., 2024) and ‘simulated deliberative democracy’ (Leike, 2023).

Research General Social Agents (Manning & Horton, 2025)

Research Policy Priority Inference (agent-based modelling to evaluate policy impacts and ordering).

Research Moral considerations of substituting AI for human involvement in deliberative processes (Revel and Pénigaud, 2025).

Research [Luke Thorburn’s forthcoming paper on the democratic use of inferred preferences].

Goal: Deep understanding of when, if at all, simulation can be helpful.

Research For what uses, in what contexts and with what level of faithfulness is it helpful or appropriate to use simulations? What are the philosophical, moral, political, etc. implications?

Goal: Deliberative processes can be tested and refined before implementation with real participants.

Research How can we develop realistic simulation environments that accurately predict how different deliberative formats will perform according to different design choices?
Research Can AI generate its own suggested changes and test them to search the latent space for optimal solutions?
Research How can lessons from speculative execution and speculative decoding help increase the availability of deliberative processes through reduced costs?
Research Solving the technical blockers to effective and truth-worthy multi-agent simulation and modelling.

Goal: Simulations are faithful enough to be relied upon when decisions are needed rapidly (e.g., seconds, minutes, hours).

Research What are the best methods to measure the faithfulness of simulations?
Research What hybrid approaches can combine fast simulation with selective human input to optimize both speed and accuracy for urgent decisions?

Goal:Product Simulator/process sandbox

Voiceless perspectives

Representation

Stakeholders, Representatives

Ability to fairly include the perspectives of those that are not represented in the process, including people who are not present (future generations, young people or other representation constraints), and non-human entities (natural phenomena or animals).

Opportunity: High

Urgency: Moderate

Capability: Limited

Consistency: Very variable

Global: Limited

Process organizers can ask participants to critically reflect on the hypothetical views of those not present. There are some exercises that prompt participants to consider or even role-play views that are not present in the room (Example).

Experimental Practice Animals in the Room and DemocracyNext.

Experimental Practice Talk to the River by Dark Matter Labs experimented with using an LLM to represent non-human entities in deliberation.

Research Extensive work in political theory and political philosophy exploring the opportunities and challenges of representing future generations and non-human animals in democratic processes (e.g., González-Ricoy and Gosseries, 2017).

Goal: Identify the most legitimate methods for representing non-human animals and future generations in deliberative processes.

Context: There are epistemological questions around the accurate or fair representations of the beliefs or interests of non-humans/future generations.
Research What are the different methods for representing non-humans/future generations, and how do these methods compare?
Research How do we evaluate any moral, philosophical, or efficacy-based justifications for trying to include the voices of non-humans/future generations in deliberations?

Goal: Effectively integrate important absent perspectives into the process.

Context: It can be hard to identify which views or perspectives might be valuable that are not currently present.
Research How do we best track and identify necessary voices that are currently missing?
Research How can virtual and augmented reality technologies help participants experience perspectives of animals or future generations?

Aggregate / distill inputs

Fair Participation

Aggregator, Summarizer

Ability to aggregate votes and distill more complex forms of open-ended input into outputs and decisions, in fair and understandable ways, such participants feel their contributions are meaningfully taken into account (and can ideally see how).

Opportunity: High

Urgency: High

Capability: Adequate

Consistency: Very variable

Global: Limited

Citizens’ Assemblies organizers utilize small group work, which relies on moderated or self-led documenting and integration of inputs. Ultimately, voting is used to ‘end’ the conversation in place of finding consensus. These decisions are sometimes made using Likert voting and supermajority thresholds.

In Deliberative Polling, ‘final’ outcomes are measured by poll.

Collective dialogue tools such as Pol.is use bridging algorithms to cluster inputs, and Remesh uses elicitation inference to help identify bridging statements across a large number of inputs.

Other online systems may not even try to surface consensus or bridging, but instead simply log the inputs of various participants.

Research Generative Social Choice (Fish et al., 2023, Boehmer et al., 2025)

Research 'Generative CI' through Collective Response Systems (Ovadya 2023),

Elicitation inference (Konya, 2022)

Research Smoothed Analysis of Social Choice (Flanigan, 2023)

Research Augmenting Polis with LLMs (Small, 2023).

Product Collective response tools to digitally elicit, collect, and analyse inputs (e.g. Polis, Remesh).

Product Some features of AI-based tooling to accompany face-to-face deliberation and help capture insights and surface analysis (e.g. Dembrane, DeliberAIde, Psi, Cortico)

Product Some features of tools like YourPriorities, Decidim, Colectiv, in addition to standard survey tools.

Goal: Collective input of many form factors is quickly and meaningfully synthesized, fairly reflecting participant perspectives.

Research How can we assist or automate the aggregation of deliberative input from diverse participants in real time, whilst maintaining nuance around minority perspectives?
Practice Translating existing social choice research into practical methodologies with decision aides for matching process to context, such as identifying trade-offs between theoretical guarantees, speed, explainability and legitimacy in the eyes of participants, public and stakeholders.

Goal: Distilled and aggregated content is integrated effectively into the process.

Research What visualization and presentation methods can best communicate distilled content for engagement by participants/organizers?
Research What aggregation approaches are best suited to different stages of a process?

Goal: Develop actionable and standardized best practices for aggregation and distillation within deliberative processes.

Context: Deliberation facilitators and self-led participant facilitation can use different aggregation methods, resulting in views being represented in vastly different ways.
Research What are the best standardized models for ‘in-person’ (in real world or online) facilitators to follow as they collate and record views of participants?
Research What standardized methods can we easily teach participants in self-led groups on how best to capture the relevant information?

Goal: Participants can track how their contributions have informed the process’s outcomes.

Research How can we best embed traceability and transparency into the aggregation/distillation process?

Informedness

The extent to which those making decisions understand the information critical to making that decision.

To what extent: (1) do participants gain critical context about tradeoffs and consequences of different decisions? (2) is this sourced from (a) experts, (b) the existing authorities, who may have extensive context, (c) a broad diversity of constituents, (d) the most impacted stakeholders, and (e) the powerful stakeholders, whose incentives are critical to having the decision "stick"?

Capability	Who & How They Do It	Existing Practice, Products and Research	Gaps
Curate context Informative Research Assistant, Consultant Ability to provide complete context to participants, including things like background information, subject matter fundamentals, relevant considerations, trade-offs, and possible options. Opportunity: High Urgency: Critical Capability: Limited Consistency: Very variable Global: Limited	Process designers work with commissioning authorities to source internal materials, sometimes use independent experts for research or request stakeholder input. The responsibility for translating materials into relevant forms varies. Citizens’ assembly organizers typically work with commissioning authorities to compile background information kits that contain basic process information, background subject matter information and other relevant information. The material is usually presented in plain language with diagrams and visual explainers where possible. Collective dialogues can have a short introductory module with subject matter fundamentals to quickly ground participants in the details. These short learning dialogues are constructed in collaboration with commissioning authorities.	Experimental Practice Harmonica and Colectiv are experimenting with AI interviewers deployed over chat apps. Product Formless and Voicepanel explore semi-structured questionnaires that can ask pre-set questions and relevant follow-ups (used commercially but could be adopted for deliberation). Experimental Practice Many AI systems are being developed for universal search engines inside organisations, having access to all unrestricted information and surfacing useful information depending on a search. Research [Zaria Jalan’s work on a method for quantifying changes in context] Experimental Practice Retrieval Augmented Generation (RAG) is a popular machine learning technique used to make LLMs provide answers only from validated data sources. This is deployed with learning materials and other pre-validated data sources. Product NotebookLM Experimental Practice Contextualization Engines (Ovadya, 2021).	Goal: Efficiently generate background information materials that are sufficiently holistic and informative. Context: Sourcing and analysing sufficient context is a resource-intensive and skill-intensive process. Research How to suitably treat information hierarchies and data privacy while accumulating and mapping the information space? Research How to enable AI-provided context that is appropriately comprehensive and sufficiently unbiased? Research How to fairly identify and fill perspective or empirical gaps in the background information? Goal: Processes are quickly responsive to participant needs when further context and information are required during the process. Research To what extent can AI be used to provide reliable live-time fact-checking within deliberations? Research How can we design responsive information systems that provide accurate context in real-time? Goal:Product Context Mapper
Activate learning Informative Tutor Ability for diverse participants to efficiently and effectively learn relevant information, such that they can actively apply their learnings in the process. Opportunity: High Urgency: Critical Capability: Adequate Consistency: Very variable Global: Limited	Processorganizers provide information and learning activities to bring participants up to speed on the topic. Collective dialogues can have a short introductory module with subject matter fundamentals to quickly ground participants in the details. These short dialogues are interactive and allow participants to learn from one another as well as read background information. Citizens’ assemblies organizers rely on building understanding through a ‘learning journey’ (multiple rounds of informing-questioning-answering) where participants engage with pre-written materials, speakers, Q&A responses, and experiences to build group understanding. Some of this is done individually (async), some is done in group environments.	Experimental Practice Games that build intuition, such as Leveraged Play Research AI agents for learning (Mollick et al, 2024) Research Limits of expert-led brief packs (albeit focused on deliberative polling) (Gleason, 2012) Research Chatbots to improve higher education (Triberti et al. 2024) Research LLMs for self-reflection on policy issues (Yuxin Ji et al. 2023)	Goal: Participants learn as much as possible in the time available. Research What are the best methods for efficiently educating people? Research How can we design adaptive learning systems that provide personalized learning programs? Research How can individual learning be mediated through group learning to lift all boats? Research How can individual learning agents identify and pair learning partners for defined objectives (idea crosspollination, depolarization, information gaps, etc) Research How can AI systems translate, generate and integrate learning materials into diverse formats (text, audio, visual, etc)? Goal:Product Multi-modal tailored learning support. Goal: Track the quality of individual and group learning within the deliberative process, unobtrusively. Research How to unobtrusively measure individual and group understanding?
Enumerate scenarios Considered Outputs Output testing Ability to generate lists of likely scenarios, including less common edge cases, in which decisions will be applied, to help participants better understand the issue space. Opportunity: High Urgency: High Capability: Limited Consistency: Very variable Global: Limited	Processorganizers rely on experts or organizers themselves to develop scenarios as a way of informing participants. Citizens’ assembly organizers will use scenarios as a way to help participants understand the logic of the issue and sometimes to help test draft proposals.	Research PolicyCraft involves using AI to generate or brainstorm cases to test policies in a participatory way. Research Generative Social Choice has to produce representative slates of statements that proportionally represent the full spectrum of user opinions from unstructured text.	Goal: Scenarios can be reliably enumerated to inform deliberations. Context: Understanding the relationships between policy considerations improves overall understanding and deliberation quality. Research How can we enumerate a spanning set of scenarios or cases that a policy needs to address? Research How can we identify the likelihood that a set is missing key scenarios? Research How can we track and mitigate biases within scenario mapping? Goal: Enumerated scenarios can be effectively and fairly integrated into deliberations. Research How can we develop criteria and methods for prioritizing scenarios based on likelihood, impact, and relevance to deliberative decisions? Research How should we best treat low probability but high impact edge cases? Research How can we effectively account for uncertainty in scenario consequences? Practice Represent scenarios in an interactive and educational process (not predictive modelling).
Forecast impacts Trade-off consideration World modeler, Expert Ability to effectively and easily model complex systems, to help participants understand the impacts of potential decisions. Opportunity: High Urgency: Critical Capability: Poor Consistency: Weak Global: Poor	Process organizers may draw on experts or use rudimentary tools to facilitate ‘if this, then that’ exercises. Citizens’ Assembly organizers may be given draft recommendations after the penultimate day to produce analysis to help assembly members understand the possible barriers to implementation and impacts of decisions.	Experimental Practice RAND Participatory Modelling of Climate Impacts on Public Health in Long Beach, California. Experimental Practice Rudimentary modelling and impact simulating tools such as Delib simulator, Forio’s Public Policy simulator, and Beamm.Brussels. Experimental Practice Early experiments using LLMs as a forecaster to help understand the likelihood of certain outcomes (e.g. FiveThirtyNine, PolicySynth). Research Future scenario generation with Generative AI (Ferrer i Pico et al. 2025) Research City-level scenario development (Hao et al. 2024) Research Gen (open-source stack for generative modeling and probabilistic inference) Research Policy Priority Inference (agent-based modelling to evaluate policy impacts and ordering) [System simulations / World models / forecasting counterfactuals [+evals]]	Goal: Participants are able to accurately understand the potential impacts of their decisions. Research How can the impacts of interventions on complex systems be simulated quickly and accurately? Research What kinds of systems are appropriate for simulation? Research What is the Pareto front of speed, accuracy and easy-to-use interactability?
Routing and synthesizing Informative Participation Facilitator Ability to route and synthesize data, revealing critical information, e.g. identifying common ground, high-potential ideas, thoughtful perspectives, insightful experiences, cruxes, forecasts, while helping to minimize the time required to do tasks. Opportunity: Moderate Urgency: Critical Capability: Adequate Consistency: Very variable Global: Limited	Process organizers generally follow structured plans that stitch together data generation and gathering activities with synthesis and understanding activities. Citizens’ assembly organizers rely on custom templates, impartial note takers, manual clustering with post-its, and group sensemaking. Participants might set key outcomes and motivation values to which experts generate proposals in response. Collective dialogue tools such as Pol.is and Remesh are designed to synthesize across many points of view with bridging algorithms and elicitation inference.	Research Elicitation inference (Konya, 2022) Research Collective Constitutional AI (Anthropic, 2023) Research Bridging Systems (Ovadya, Thorburn, 2023) Research Habermas Machine (Tessler et al. 2024) Product Dembrane, DeliberAIde Product Polis, Remesh Product Talk to the City, GoVocal, PSi	Goal: Effectively support participants in finding common ground. Context: Processes can be path-dependent due to time constraints and vast amounts of data. Research What are the best methods for generating and integrating effective, actionable and value-aligned proposals for bridging divides? Research How to ensure that promising bridging proposals that are identified are sufficiently concrete? Goal: Processes can route and synthesize content in a dynamic fashion according to particular needs and goals. Research When should we route divergent info to help participants understand the full range of perspectives, and when should we route convergent info to help find consensus?
Integrate less-involved Influence Mass Contributor Ability to provide those not in the room deliberating with opportunities to constructively and fairly contribute input into the process. Opportunity: Moderate Urgency: High Capability: Limited Consistency: Very variable Global: Limited	Processorganizers use mass engagement methods like surveys and submissions to gather opt-in input to informing processes. The outcomes of these engagement programs are usually then shared with those inside the process. Citizens’ assembly organizers will make clear distinctions between those inside the process and those outside. They develop constructive ways for the maxi public to contribute without constraining or contradicting the internal processes of high-context assembly members, such as by asking for contributions on questions, concerns, hopes and information sources rather than rushing to judgement on proposals or outputs.	Experimental Practice Recent experiments (such as in Bowling Green, Kentucky) combine tools like Polis and Google Jigsaw’s Sensemaker to quickly understand a community's priorities. Experimental Practice Integrating Polis with Citizens Assemblies Research Goñi suggests that big data can be combined with ‘little data’ to ensure core threads of wider public opinion are elicited without losing the personal nuance and stories that underpin them. Practice Experiments in using multiple engagement methods to access different groups and piece outputs together (e.g. Le Grand Débat in France)	Goal: Maxi-public engagements attract and support participation from diverse social groups. Research What methods can best incentivize and enable participation in maxi-public engagements beyond the ‘usual suspects’? Research How can we fairly balance self-selection biases with open opportunities for contribution in wider public engagement? Goal: Maxi-public engagements are effectively and fairly integrated into mini-public deliberations. Context: Finding processes and workflows that can bridge the ‘mini’ (high context) and ‘maxi’ (low context) public effectively is challenging. Research How can we design synthesis and filtering systems that distill massive public input into actionable insights? Research How much weight, if any, should be given to discussions or engagements with a wider public outside the central deliberation? Research How can maxi-public contributions be integrated in a way that maintains nuance and connections to personal stories/contexts? Research What feedback mechanisms/traceability measures can help participants understand how their contributions influenced outcomes in the process? Research Which types of wider public contributions are productive and which are prone to contradicting or creating hierarchical confusion between engagements?

Capability

Who & How They Do It

Existing Practice, Products and Research

Gaps

Curate context

Informative

Research Assistant, Consultant

Ability to provide complete context to participants, including things like background information, subject matter fundamentals, relevant considerations, trade-offs, and possible options.

Opportunity: High

Urgency: Critical

Capability: Limited

Consistency: Very variable

Global: Limited

Process designers work with commissioning authorities to source internal materials, sometimes use independent experts for research or request stakeholder input. The responsibility for translating materials into relevant forms varies.

Citizens’ assembly organizers typically work with commissioning authorities to compile background information kits that contain basic process information, background subject matter information and other relevant information. The material is usually presented in plain language with diagrams and visual explainers where possible.

Collective dialogues can have a short introductory module with subject matter fundamentals to quickly ground participants in the details. These short learning dialogues are constructed in collaboration with commissioning authorities.

Experimental Practice Harmonica and Colectiv are experimenting with AI interviewers deployed over chat apps.

Product Formless and Voicepanel explore semi-structured questionnaires that can ask pre-set questions and relevant follow-ups (used commercially but could be adopted for deliberation).

Experimental Practice Many AI systems are being developed for universal search engines inside organisations, having access to all unrestricted information and surfacing useful information depending on a search.

Research [Zaria Jalan’s work on a method for quantifying changes in context]

Experimental Practice Retrieval Augmented Generation (RAG) is a popular machine learning technique used to make LLMs provide answers only from validated data sources. This is deployed with learning materials and other pre-validated data sources.

Product NotebookLM

Experimental Practice Contextualization Engines (Ovadya, 2021).

Goal: Efficiently generate background information materials that are sufficiently holistic and informative.

Context: Sourcing and analysing sufficient context is a resource-intensive and skill-intensive process.
Research How to suitably treat information hierarchies and data privacy while accumulating and mapping the information space?
Research How to enable AI-provided context that is appropriately comprehensive and sufficiently unbiased?
Research How to fairly identify and fill perspective or empirical gaps in the background information?

Goal: Processes are quickly responsive to participant needs when further context and information are required during the process.

Research To what extent can AI be used to provide reliable live-time fact-checking within deliberations?
Research How can we design responsive information systems that provide accurate context in real-time?

Goal:Product Context Mapper

Activate learning

Informative

Tutor

Ability for diverse participants to efficiently and effectively learn relevant information, such that they can actively apply their learnings in the process.

Opportunity: High

Urgency: Critical

Capability: Adequate

Consistency: Very variable

Global: Limited

Processorganizers provide information and learning activities to bring participants up to speed on the topic.

Collective dialogues can have a short introductory module with subject matter fundamentals to quickly ground participants in the details. These short dialogues are interactive and allow participants to learn from one another as well as read background information.

Citizens’ assemblies organizers rely on building understanding through a ‘learning journey’ (multiple rounds of informing-questioning-answering) where participants engage with pre-written materials, speakers, Q&A responses, and experiences to build group understanding. Some of this is done individually (async), some is done in group environments.

Experimental Practice Games that build intuition, such as Leveraged Play

Research AI agents for learning (Mollick et al, 2024)

Research Limits of expert-led brief packs (albeit focused on deliberative polling) (Gleason, 2012)

Research Chatbots to improve higher education (Triberti et al. 2024)

Research LLMs for self-reflection on policy issues (Yuxin Ji et al. 2023)

Goal: Participants learn as much as possible in the time available.

Research What are the best methods for efficiently educating people?
Research How can we design adaptive learning systems that provide personalized learning programs?
Research How can individual learning be mediated through group learning to lift all boats?
Research How can individual learning agents identify and pair learning partners for defined objectives (idea crosspollination, depolarization, information gaps, etc)
Research How can AI systems translate, generate and integrate learning materials into diverse formats (text, audio, visual, etc)?

Goal:Product Multi-modal tailored learning support.

Goal: Track the quality of individual and group learning within the deliberative process, unobtrusively.

Research How to unobtrusively measure individual and group understanding?

Enumerate scenarios

Considered Outputs

Output testing

Ability to generate lists of likely scenarios, including less common edge cases, in which decisions will be applied, to help participants better understand the issue space.

Opportunity: High

Urgency: High

Capability: Limited

Consistency: Very variable

Global: Limited

Processorganizers rely on experts or organizers themselves to develop scenarios as a way of informing participants.

Citizens’ assembly organizers will use scenarios as a way to help participants understand the logic of the issue and sometimes to help test draft proposals.

Research PolicyCraft involves using AI to generate or brainstorm cases to test policies in a participatory way.

Research Generative Social Choice has to produce representative slates of statements that proportionally represent the full spectrum of user opinions from unstructured text.

Goal: Scenarios can be reliably enumerated to inform deliberations.

Context: Understanding the relationships between policy considerations improves overall understanding and deliberation quality.
Research How can we enumerate a spanning set of scenarios or cases that a policy needs to address?
Research How can we identify the likelihood that a set is missing key scenarios?
Research How can we track and mitigate biases within scenario mapping?

Goal: Enumerated scenarios can be effectively and fairly integrated into deliberations.

Research How can we develop criteria and methods for prioritizing scenarios based on likelihood, impact, and relevance to deliberative decisions?
Research How should we best treat low probability but high impact edge cases?
Research How can we effectively account for uncertainty in scenario consequences?
Practice Represent scenarios in an interactive and educational process (not predictive modelling).

Forecast impacts

Trade-off consideration

World modeler, Expert

Ability to effectively and easily model complex systems, to help participants understand the impacts of potential decisions.

Opportunity: High

Urgency: Critical

Capability: Poor

Consistency: Weak

Global: Poor

Process organizers may draw on experts or use rudimentary tools to facilitate ‘if this, then that’ exercises.

Citizens’ Assembly organizers may be given draft recommendations after the penultimate day to produce analysis to help assembly members understand the possible barriers to implementation and impacts of decisions.

Experimental Practice RAND Participatory Modelling of Climate Impacts on Public Health in Long Beach, California.

Experimental Practice Rudimentary modelling and impact simulating tools such as Delib simulator, Forio’s Public Policy simulator, and Beamm.Brussels.

Experimental Practice Early experiments using LLMs as a forecaster to help understand the likelihood of certain outcomes (e.g. FiveThirtyNine, PolicySynth).

Research Future scenario generation with Generative AI (Ferrer i Pico et al. 2025)

Research City-level scenario development (Hao et al. 2024)

Research Gen (open-source stack for generative modeling and probabilistic inference)

Research Policy Priority Inference (agent-based modelling to evaluate policy impacts and ordering)

[System simulations / World models / forecasting counterfactuals [+evals]]

Goal: Participants are able to accurately understand the potential impacts of their decisions.

Research How can the impacts of interventions on complex systems be simulated quickly and accurately?
Research What kinds of systems are appropriate for simulation?
Research What is the Pareto front of speed, accuracy and easy-to-use interactability?

Routing and synthesizing

Informative Participation

Facilitator

Ability to route and synthesize data, revealing critical information, e.g. identifying common ground, high-potential ideas, thoughtful perspectives, insightful experiences, cruxes, forecasts, while helping to minimize the time required to do tasks.

Opportunity: Moderate

Urgency: Critical

Capability: Adequate

Consistency: Very variable

Global: Limited

Process organizers generally follow structured plans that stitch together data generation and gathering activities with synthesis and understanding activities.

Citizens’ assembly organizers rely on custom templates, impartial note takers, manual clustering with post-its, and group sensemaking. Participants might set key outcomes and motivation values to which experts generate proposals in response.

Collective dialogue tools such as Pol.is and Remesh are designed to synthesize across many points of view with bridging algorithms and elicitation inference.

Research Elicitation inference (Konya, 2022)

Research Collective Constitutional AI (Anthropic, 2023)

Research Bridging Systems (Ovadya, Thorburn, 2023)

Research Habermas Machine (Tessler et al. 2024)

Product Dembrane, DeliberAIde

Product Polis, Remesh

Product Talk to the City, GoVocal, PSi

Goal: Effectively support participants in finding common ground.

Context: Processes can be path-dependent due to time constraints and vast amounts of data.
Research What are the best methods for generating and integrating effective, actionable and value-aligned proposals for bridging divides?
Research How to ensure that promising bridging proposals that are identified are sufficiently concrete?

Goal: Processes can route and synthesize content in a dynamic fashion according to particular needs and goals.

Research When should we route divergent info to help participants understand the full range of perspectives, and when should we route convergent info to help find consensus?

Integrate less-involved

Influence

Mass Contributor

Ability to provide those not in the room deliberating with opportunities to constructively and fairly contribute input into the process.

Opportunity: Moderate

Urgency: High

Capability: Limited

Consistency: Very variable

Global: Limited

Processorganizers use mass engagement methods like surveys and submissions to gather opt-in input to informing processes. The outcomes of these engagement programs are usually then shared with those inside the process.

Citizens’ assembly organizers will make clear distinctions between those inside the process and those outside. They develop constructive ways for the maxi public to contribute without constraining or contradicting the internal processes of high-context assembly members, such as by asking for contributions on questions, concerns, hopes and information sources rather than rushing to judgement on proposals or outputs.

Experimental Practice Recent experiments (such as in Bowling Green, Kentucky) combine tools like Polis and Google Jigsaw’s Sensemaker to quickly understand a community's priorities.

Experimental Practice Integrating Polis with Citizens Assemblies

Research Goñi suggests that big data can be combined with ‘little data’ to ensure core threads of wider public opinion are elicited without losing the personal nuance and stories that underpin them.

Practice Experiments in using multiple engagement methods to access different groups and piece outputs together (e.g. Le Grand Débat in France)

Goal: Maxi-public engagements attract and support participation from diverse social groups.

Research What methods can best incentivize and enable participation in maxi-public engagements beyond the ‘usual suspects’?
Research How can we fairly balance self-selection biases with open opportunities for contribution in wider public engagement?

Goal: Maxi-public engagements are effectively and fairly integrated into mini-public deliberations.

Context: Finding processes and workflows that can bridge the ‘mini’ (high context) and ‘maxi’ (low context) public effectively is challenging.
Research How can we design synthesis and filtering systems that distill massive public input into actionable insights?
Research How much weight, if any, should be given to discussions or engagements with a wider public outside the central deliberation?
Research How can maxi-public contributions be integrated in a way that maintains nuance and connections to personal stories/contexts?
Research What feedback mechanisms/traceability measures can help participants understand how their contributions influenced outcomes in the process?
Research Which types of wider public contributions are productive and which are prone to contradicting or creating hierarchical confusion between engagements?

Substantiveness

The extent to which decisions are substantive (e.g., actionable, consequential) rather than nonsubstantive (e.g., vague, simplistic, inconsequential).

To what extent: (1) is the decision directly actionable and implementable? (2) does the decision meaningfully address the issues? (3) does the decision grapple with the necessary levels of complexity? (4) is uncertainty appropriately managed and accounted for? (5) are risks to implementability accounted for?

Capability	Who & How They Do It	Existing Practice, Products and Research	Gaps
Represent complexity Richness Stress Tester Ability for final outputs to be nuanced, concrete, decisive, and comprehensive. Opportunity: High Urgency: Critical Capability: Limited Consistency: Weak Global: Limited	Process organizers create workflows that optimize for desired outcomes, properties such as concreteness and group agreement can be in tension. Citizens’ assembly organizers create workflows that develop proposals in templated formats which help structure final outputs. Ultimately, the process is constrained by time, the requirement to find common ground, and access to the relevant expertise.	Experimental Practice Moral Graphs (Edelman et al., 2023) Research [Abstract Syntax Trees for policy documents?] Research [Sally Dong et al’s forthcoming paper on concreteness]	Goal: Recommendations are effectively stress tested to be as robust as possible. Research What are the best methods for providing impartial robustness checking and critical friend support for output refinement? Research How can we ensure that outputs go beyond abstract, high-level principles to specific, actionable proposals? Goal: Agreement is found on concrete proposals and not empty platitudes. Research How can we measure the concreteness of statements and recommendations? Goal: Outputs address the complexity and nuance of the issue. Research How to balance finding common ground within a limited time, while minimally sacrificing depth of final outputs?
Adaptable outputs Adaptability Policymaker Ability for final outputs to be adaptable to changing contexts while retaining clear intended outcomes and specificity. Opportunity: High Urgency: High Capability: Limited Consistency: Weak Global: Limited	Process organizers can create workflows that intentionally ask participants to consider adaptability in the development of their outputs, or structure outputs such that they’re inherently adaptable. Citizens’ assembly organizers advise assembly members to design their recommendations in ways that may not constrain the organizing authority in the future in a way that would be inconsistent with the intent of the policy recommendation. This relies on the intent being clear to those implementing and them honoring that intent with a new policy if conditions change.	Practice The Permanent Secretary within the Ostbelgian Parliament supports dialogue between the Citizens’ Council and Parliamentarians to help convey the intent and negotiate the implementation of recommendations (Niessen and Reuchamps, 2022).	Goal: Deliberative outputs remain relevant, adaptable and useful as contexts shift. Research How can deliberative outputs be developed to accommodate revisions over time whilst preserving their intended motivations? Research What are the best methods for enabling iterative and ongoing citizen engagement so recommendations can be updated as contexts shift? Research How can deliberative outputs be formatted as functions such that they can automatically adapt?
Implementable outputs Actionability Facilitator, Templater Ability to produce outputs in immediately actionable forms (e.g. policies, budgets, AI constitutions, town plans etc.) Opportunity: High Urgency: Critical Capability: Adequate Consistency: Very variable Global: Poor	Process organizers create workflows that structure the development of outputs such that they’re implementable. Citizens’ assembly organizers create opportunities for experts and commissioning authorities to review draft outputs and comment on implementability and scope. Workflows are designed to support the templated development of outputs in required forms. The commissioning authority might provide requirements or guidelines for outputs to maximize their uptake (e.g., town plans, rates, functions)		Goal: Outputs from processes respect the relevant legal and institutional confines in which they are operating. Research What are the most effective methods of testing the compatibility of outputs with legal/constitutional/jurisdictional or other fundamental constraints on recommendation uptake? Goal: Outputs closely engage with key stakeholder concerns, perspectives and realities. Context: Outputs that fail to preempt significant stakeholders' concerns are unlikely to translate to impact Research What are the best ways of anticipating key objections core power holders may raise against recommendations? Goal: Outputs are presented in clear and precise language to enable translation to action. Research What are the most effective methods and formats for presenting process outputs to decision makers, and what tools can support this process? Goal:Product Smart Templates (Card): System for helping ensure that the outputs of a task are in the right form, e.g., by evaluating content’s fit to a specified template and instruction, and giving feedback as comments, suggested edits, or through chat. Goal:Product Deliberative tool that allows large groups (1000+) to collaboratively author a cohesive document, with guarantees on the relative influence of each contributor.

Capability

Who & How They Do It

Existing Practice, Products and Research

Gaps

Represent complexity

Richness

Stress Tester

Ability for final outputs to be nuanced, concrete, decisive, and comprehensive.

Opportunity: High

Urgency: Critical

Capability: Limited

Consistency: Weak

Global: Limited

Process organizers create workflows that optimize for desired outcomes, properties such as concreteness and group agreement can be in tension.

Citizens’ assembly organizers create workflows that develop proposals in templated formats which help structure final outputs. Ultimately, the process is constrained by time, the requirement to find common ground, and access to the relevant expertise.

Experimental Practice Moral Graphs (Edelman et al., 2023)

Research [Abstract Syntax Trees for policy documents?]

Research [Sally Dong et al’s forthcoming paper on concreteness]

Goal: Recommendations are effectively stress tested to be as robust as possible.

Research What are the best methods for providing impartial robustness checking and critical friend support for output refinement?
Research How can we ensure that outputs go beyond abstract, high-level principles to specific, actionable proposals?

Goal: Agreement is found on concrete proposals and not empty platitudes.

Research How can we measure the concreteness of statements and recommendations?

Goal: Outputs address the complexity and nuance of the issue.

Research How to balance finding common ground within a limited time, while minimally sacrificing depth of final outputs?

Adaptable outputs

Adaptability

Policymaker

Ability for final outputs to be adaptable to changing contexts while retaining clear intended outcomes and specificity.

Opportunity: High

Urgency: High

Capability: Limited

Consistency: Weak

Global: Limited

Process organizers can create workflows that intentionally ask participants to consider adaptability in the development of their outputs, or structure outputs such that they’re inherently adaptable.

Citizens’ assembly organizers advise assembly members to design their recommendations in ways that may not constrain the organizing authority in the future in a way that would be inconsistent with the intent of the policy recommendation. This relies on the intent being clear to those implementing and them honoring that intent with a new policy if conditions change.

Practice The Permanent Secretary within the Ostbelgian Parliament supports dialogue between the Citizens’ Council and Parliamentarians to help convey the intent and negotiate the implementation of recommendations (Niessen and Reuchamps, 2022).

Goal: Deliberative outputs remain relevant, adaptable and useful as contexts shift.

Research How can deliberative outputs be developed to accommodate revisions over time whilst preserving their intended motivations?
Research What are the best methods for enabling iterative and ongoing citizen engagement so recommendations can be updated as contexts shift?
Research How can deliberative outputs be formatted as functions such that they can automatically adapt?

Implementable outputs

Actionability

Facilitator, Templater

Ability to produce outputs in immediately actionable forms (e.g. policies, budgets, AI constitutions, town plans etc.)

Opportunity: High

Urgency: Critical

Capability: Adequate

Consistency: Very variable

Global: Poor

Process organizers create workflows that structure the development of outputs such that they’re implementable.

Citizens’ assembly organizers create opportunities for experts and commissioning authorities to review draft outputs and comment on implementability and scope. Workflows are designed to support the templated development of outputs in required forms. The commissioning authority might provide requirements or guidelines for outputs to maximize their uptake (e.g., town plans, rates, functions)

Goal: Outputs from processes respect the relevant legal and institutional confines in which they are operating.

Research What are the most effective methods of testing the compatibility of outputs with legal/constitutional/jurisdictional or other fundamental constraints on recommendation uptake?

Goal: Outputs closely engage with key stakeholder concerns, perspectives and realities.

Context: Outputs that fail to preempt significant stakeholders' concerns are unlikely to translate to impact
Research What are the best ways of anticipating key objections core power holders may raise against recommendations?

Goal: Outputs are presented in clear and precise language to enable translation to action.

Research What are the most effective methods and formats for presenting process outputs to decision makers, and what tools can support this process?

Goal:Product Smart Templates (Card): System for helping ensure that the outputs of a task are in the right form, e.g., by evaluating content’s fit to a specified template and instruction, and giving feedback as comments, suggested edits, or through chat.

Goal:Product Deliberative tool that allows large groups (1000+) to collaboratively author a cohesive document, with guarantees on the relative influence of each contributor.

Deliberation

The extent to which decisions are considered and deliberative (rather than superficial and reactive).

To what extent are those involved: (1) able to (and supported to) move from shallower to deeper goals and values? (2) able to (and supported to) collaborate where necessary? (3) able to address issues within the available time?

Capability	Who & How They Do It	Existing Practice, Products and Research	Gaps
Facilitate deliberation Process Completion Facilitator Ability to develop appropriate workflows and navigate mixed groups to successful outcomes. Opportunity: High Urgency: Critical Capability: Good Consistency: Variable Global: Adequate	Process organizers design process workflows and lead participants through them with varying degrees of involvement (light and heavy touch). Citizens’ assembly organizers have facilitation teams that lead the assembly members through a mostly predetermined workflow, focused on ensuring the overall task and outcomes are met within the allotted time. They do their best to be impartial while helping to maximize the processes’ efficacy.	Research Impact of AI facilitation on public perceptions (Jungherr and Rauchfleisch, 2025). Practice Facilitating Deliberation: A Practical Guide by MosaicLab Product Online tools to hold conversations between people (e.g. Stanford Online Deliberation Platform, Frankly, PSi). Product Online tools to help discussion in small groups, with active live cross-pollination between groups throughout (e.g. ThinkScape).	Goal: Use digital tools to enhance facilitation quality and capacity. Research How can digital tools assist human facilitators to more effectively facilitate deliberations? Research How to develop an AI facilitator that is attentive to power imbalances, adaptive to group dynamics and effective in guiding groups towards successful outcomes? Research What are the effects of AI facilitation on public perceptions, group dynamics and deliberative quality?
Support collaboration Synergy Facilitator Ability to collaboratively work together to develop policies and other complex artefacts. Opportunity: High Urgency: Critical Capability: Adequate Consistency: Variable Global: Limited	Process organizers design process workflows that intentionally avoid any data conflicts or versioning issues with participant-produced information, while also creating space for the reconciliation of differing views. Citizens’ assembly organizers routinely combine small group deliberations with plenary reporting to divide work, refine and provide feedback as a group, and then act on that feedback in small groups. For example, one very common specific challenge is the reconciliation of multiple versions of a document, created by different subgroups.	Research PolicyPad: Editor supporting experts in collaboratively drafting policies (Feng et al., 2025)	Goal: Participants can efficiently write proposals in ways that support large group collaboration while identifying differences for reconciliation through deliberation. Research How can we convey similarities and differences between the outputs? (Including satellite deliberations (across cultures and languages) with the goal of finding common ground?)
Enable reason-giving Reflexivity Discussion Partner Ability to facilitate mutual understanding and reason-giving, including by supporting the development of individual critical thinking skills and preferences. Opportunity: High Urgency: Critical Capability: Adequate Consistency: Very variable Global: Limited	Process organizers design workflows that allow for structured individual reflection between moments of discussion or learning. Citizens’ assembly organizers design workflows that build social bonds between participants, before they’ve exchanged perspectives on the issue at hand, to help tackle cognitive biases or assumptions. They design individual reflection sessions and sometimes recommend asynchronous journaling to develop an individual's views further. They often facilitate epistemic skills-building exercises focused on identifying unconscious biases or training critical thinking, and reinforcing these exercises with reminders throughout the process. They also have a range of facilitation methods that create different settings or modes for expression, allowing participants to find a mode that suits them.	Experimental Practice Deliberation and participation monitoring and their visual representations from DemocracyNext and MIT’s Centre for Constructive Communications. Practice Epistemics Resources (MosaicLab Critical Thinking Overview, newDemocracy Critical Thinking Video, newDemocracy Unconscious Biases Video) Practice Legislative Theatre: Using Performance to Make Politics Practice The Graphic Facilitators Guide Practice Game Storming Practice Convivial Toolbox (3D Model Building where participants create physical representations of systems, processes, or future scenarios)	Goal: Participants are well equipped to engage critically with evidence and arguments. Research How can we develop participants' reasoning and critical thinking skills within a process? Research How can we help participants reason about long-term consequences and intergenerational impacts that are difficult to visualize or experience directly? Goal: Participants can critically reflect on their preferences in response to new information. Research How can we support people to critically self-reflect on their preferences? Goal: Participants from all backgrounds are supported to understand the arguments of others. Research What methods can support participants to understand better the perspectives of others (e.g., automated language simplification, visual summary)? Goal: Participants from all backgrounds are supported to effectively express their views, and these are given fair consideration. Research What techniques can help citizens effectively surface, reflect on and convey their perspective to others? Research How can we track the perspectives offered and ensure that they all receive appropriate engagement? Goal: Reasons are effectively transmitted across individuals/groups. Research How to rotate groups/route comments to provide an optimal exposure and testing of different reasons?
Localize participation Intelligibility Translator Ability to run processes in multiple languages and cultural contexts in real time and account for linguistic differences in the precise intent of outputs. Opportunity: High Urgency: Critical Capability: Limited Consistency: Weak Global: Limited	Process organizers use human interpretation when working with a small number of languages. Citizens’ assembly organizers will use human interpretation when financially feasible, otherwise rely on Google Translate for quick translation of documents. The European Commission’s Citizens’ Panels involve live human interpretation across 26 languages.	Research There has been incredible progress on automated real-time translation in the past several years (Stahlberg, 2020) Product Google Meet now does real-time audio translation between English and Spanish. Product Decidim has integrated DeepL Pro, so every comment appears instantly in each participant’s chosen language. Experimental Practice Google Translate can do a passable job of translating across an entire collective dialogue process (Konya, 2025)	Goal: Participants can fully engage in multilingual deliberation regardless of their native language or geographical location. Research How do we best ensure that speakers of low-resource languages are not disadvantaged within deliberative forums? Research How can translation best be provided for those in very remote and hard-to-access regions? Goal: Translation is attentive to nuances and subtleties in communication. Context: Translation often focuses on literal meaning while losing tone, emotion, cultural cues, and non-verbal communication that are crucial for building trust and understanding in deliberative settings. Research How can we ensure that nuances in communication are effectively captured during live translation of deliberations? Goal: Translation occurs in real time and integrates smoothly into the process. Research What different translation tools and techniques are particularly suited for the different stages of a deliberative process? Practice How can these different techniques be combined and integrated most seamlessly? Goal: Product Babelfish
Navigate contexts Context Sensitivity Mediator Ability to facilitate tolerance, discussions and collaboration across divides (historical and ongoing). Opportunity: High Urgency: Critical Capability: Adequate Consistency: Good Global: Limited	Process organizers make adaptations to normal methods to account for added complexities. Citizens’ assembly organizers will make adaptations, such as not listing names on name badges, to disrupt ethnic profiling before social bonds can be built. They might also rotate the location of the venue to balance time spent in different regions or neighbourhoods. Facilitation teams might be intentionally representative of all sides of an issue or completely distinct. The process overall would spend more time building connections and prioritize hearing from everyone before moving through the workflow.	Experimental Practice Collective dialogues in difficult situations (Konya, 2025). Research Deliberative democracy in divided societies (e.g., Dryzek, 2005; O’Flynn and Caluwaerts, 2018).	Goal: Processes enable meaningful dialogue and deliberation across deep divisions. Research How to determine the point at which polarization will impact deliberation beyond usual practice? Research What processes are particularly suited to fragile and polarized contexts, and what adaptations in process designs are necessary? Research How can AI support depolarization, and what new problems might it create in low-trust environments?

Capability

Who & How They Do It

Existing Practice, Products and Research

Gaps

Facilitate deliberation

Process Completion

Facilitator

Ability to develop appropriate workflows and navigate mixed groups to successful outcomes.

Opportunity: High

Urgency: Critical

Capability: Good

Consistency: Variable

Global: Adequate

Process organizers design process workflows and lead participants through them with varying degrees of involvement (light and heavy touch).

Citizens’ assembly organizers have facilitation teams that lead the assembly members through a mostly predetermined workflow, focused on ensuring the overall task and outcomes are met within the allotted time. They do their best to be impartial while helping to maximize the processes’ efficacy.

Research Impact of AI facilitation on public perceptions (Jungherr and Rauchfleisch, 2025).

Practice Facilitating Deliberation: A Practical Guide by MosaicLab

Product Online tools to hold conversations between people (e.g. Stanford Online Deliberation Platform, Frankly, PSi).

Product Online tools to help discussion in small groups, with active live cross-pollination between groups throughout (e.g. ThinkScape).

Goal: Use digital tools to enhance facilitation quality and capacity.

Research How can digital tools assist human facilitators to more effectively facilitate deliberations?
Research How to develop an AI facilitator that is attentive to power imbalances, adaptive to group dynamics and effective in guiding groups towards successful outcomes?
Research What are the effects of AI facilitation on public perceptions, group dynamics and deliberative quality?

Support collaboration

Synergy

Facilitator

Ability to collaboratively work together to develop policies and other complex artefacts.

Opportunity: High

Urgency: Critical

Capability: Adequate

Consistency: Variable

Global: Limited

Process organizers design process workflows that intentionally avoid any data conflicts or versioning issues with participant-produced information, while also creating space for the reconciliation of differing views.

Citizens’ assembly organizers routinely combine small group deliberations with plenary reporting to divide work, refine and provide feedback as a group, and then act on that feedback in small groups.

For example, one very common specific challenge is the reconciliation of multiple versions of a document, created by different subgroups.

Research PolicyPad: Editor supporting experts in collaboratively drafting policies (Feng et al., 2025)

Goal: Participants can efficiently write proposals in ways that support large group collaboration while identifying differences for reconciliation through deliberation.

Research How can we convey similarities and differences between the outputs? (Including satellite deliberations (across cultures and languages) with the goal of finding common ground?)

Enable reason-giving

Reflexivity

Discussion Partner

Ability to facilitate mutual understanding and reason-giving, including by supporting the development of individual critical thinking skills and preferences.

Opportunity: High

Urgency: Critical

Capability: Adequate

Consistency: Very variable

Global: Limited

Process organizers design workflows that allow for structured individual reflection between moments of discussion or learning.

Citizens’ assembly organizers design workflows that build social bonds between participants, before they’ve exchanged perspectives on the issue at hand, to help tackle cognitive biases or assumptions. They design individual reflection sessions and sometimes recommend asynchronous journaling to develop an individual's views further. They often facilitate epistemic skills-building exercises focused on identifying unconscious biases or training critical thinking, and reinforcing these exercises with reminders throughout the process. They also have a range of facilitation methods that create different settings or modes for expression, allowing participants to find a mode that suits them.

Experimental Practice Deliberation and participation monitoring and their visual representations from DemocracyNext and MIT’s Centre for Constructive Communications.

Practice Epistemics Resources (MosaicLab Critical Thinking Overview, newDemocracy Critical Thinking Video, newDemocracy Unconscious Biases Video)

Practice Legislative Theatre: Using Performance to Make Politics

Practice The Graphic Facilitators Guide

Practice Game Storming

Practice Convivial Toolbox (3D Model Building where participants create physical representations of systems, processes, or future scenarios)

Goal: Participants are well equipped to engage critically with evidence and arguments.

Research How can we develop participants' reasoning and critical thinking skills within a process?
Research How can we help participants reason about long-term consequences and intergenerational impacts that are difficult to visualize or experience directly?

Goal: Participants can critically reflect on their preferences in response to new information.

Research How can we support people to critically self-reflect on their preferences?

Goal: Participants from all backgrounds are supported to understand the arguments of others.

Research What methods can support participants to understand better the perspectives of others (e.g., automated language simplification, visual summary)?

Goal: Participants from all backgrounds are supported to effectively express their views, and these are given fair consideration.

Research What techniques can help citizens effectively surface, reflect on and convey their perspective to others?
Research How can we track the perspectives offered and ensure that they all receive appropriate engagement?

Goal: Reasons are effectively transmitted across individuals/groups.

Research How to rotate groups/route comments to provide an optimal exposure and testing of different reasons?

Localize participation

Intelligibility

Translator

Ability to run processes in multiple languages and cultural contexts in real time and account for linguistic differences in the precise intent of outputs.

Opportunity: High

Urgency: Critical

Capability: Limited

Consistency: Weak

Global: Limited

Process organizers use human interpretation when working with a small number of languages.

Citizens’ assembly organizers will use human interpretation when financially feasible, otherwise rely on Google Translate for quick translation of documents.

The European Commission’s Citizens’ Panels involve live human interpretation across 26 languages.

Research There has been incredible progress on automated real-time translation in the past several years (Stahlberg, 2020)

Product Google Meet now does real-time audio translation between English and Spanish.

Product Decidim has integrated DeepL Pro, so every comment appears instantly in each participant’s chosen language.

Experimental Practice Google Translate can do a passable job of translating across an entire collective dialogue process (Konya, 2025)

Goal: Participants can fully engage in multilingual deliberation regardless of their native language or geographical location.

Research How do we best ensure that speakers of low-resource languages are not disadvantaged within deliberative forums?
Research How can translation best be provided for those in very remote and hard-to-access regions?

Goal: Translation is attentive to nuances and subtleties in communication.

Context: Translation often focuses on literal meaning while losing tone, emotion, cultural cues, and non-verbal communication that are crucial for building trust and understanding in deliberative settings.
Research How can we ensure that nuances in communication are effectively captured during live translation of deliberations?

Goal: Translation occurs in real time and integrates smoothly into the process.

Research What different translation tools and techniques are particularly suited for the different stages of a deliberative process?
Practice How can these different techniques be combined and integrated most seamlessly?

Goal: Product Babelfish

Navigate contexts

Context Sensitivity

Mediator

Ability to facilitate tolerance, discussions and collaboration across divides (historical and ongoing).

Opportunity: High

Urgency: Critical

Capability: Adequate

Consistency: Good

Global: Limited

Process organizers make adaptations to normal methods to account for added complexities.

Citizens’ assembly organizers will make adaptations, such as not listing names on name badges, to disrupt ethnic profiling before social bonds can be built. They might also rotate the location of the venue to balance time spent in different regions or neighbourhoods. Facilitation teams might be intentionally representative of all sides of an issue or completely distinct. The process overall would spend more time building connections and prioritize hearing from everyone before moving through the workflow.

Experimental Practice Collective dialogues in difficult situations (Konya, 2025).

Research Deliberative democracy in divided societies (e.g., Dryzek, 2005; O’Flynn and Caluwaerts, 2018).

Goal: Processes enable meaningful dialogue and deliberation across deep divisions.

Research How to determine the point at which polarization will impact deliberation beyond usual practice?
Research What processes are particularly suited to fragile and polarized contexts, and what adaptations in process designs are necessary?
Research How can AI support depolarization, and what new problems might it create in low-trust environments?

Robustness

The extent to which the process is robust to suboptimal conditions or adversarial or strategic behavior.

To what extent is the process or system vulnerable to: (1) suboptimal conditions or broken assumptions? (e.g., low turnout, larger power asymmetries) (2) strategic behavior and manipulation? (3) false claims? (e.g., of manipulation)

Capability	Who & How They Do It	Existing Practice, Products and Research	Gaps
Resist manipulation Integrity Guarantor, Supervisor Ability to resist manipulation that would decrease trustworthiness, legitimacy or unfairly influence the outcome. Opportunity: High Urgency: High Capability: Adequate Consistency: Good Global: Limited	Process organizers anticipate vulnerabilities in processes and do their best to mitigate risk with countermeasures. Citizens’ assembly organizers design processes with an understanding of where manipulation is possible and more likely, and develop mitigating strategies, such as by reinforcing the epistemic capabilities of participants before interacting with new information, developing selection algorithms with manipulation resistance, and establishing governance protocols for impartiality of key actors.	Research Strategyproofness guarantees in voting mechanisms ensure participants cannot benefit from misrepresenting preferences (e.g., theoretical applications) (Satterthwaite, 1975) AI Research Adversarial testing for Generative AI develops methods to detect and prevent manipulation of AI-assisted deliberation tools. Research Strategic classification literature examines how actors modify behavior when being evaluated by algorithms (Hardt, 2015; Milli, 2019; Miller, 2020). Research Sortition algorithms with manipulation resistance (Flanigan, 2021; Baharav, 2024)	Goal: Manipulation attempts can be reliably detected and prevented across different stages of the assembly process. Research How can we develop real-time detection systems for coordinated manipulation attempts during participant recruitment and selection? Research What behavioral indicators reliably signal attempts to game deliberative processes? Research How can we distinguish between legitimate persuasion and manipulative influence in deliberative settings? Goal: Assembly designs are robust to both internal and external manipulation attempts. Research How can we quantify and test the manipulation resistance of different assembly design choices? Research What are the tradeoffs between openness/transparency and manipulation resistance? Research How can we design information presentation formats that minimize susceptibility to framing effects? Goal: Post-hoc verification can establish that outcomes were not unduly influenced. Research How can we develop manipulation impact metrics that distinguish between minor and outcome-altering influences? Practice Create standardized integrity assessment frameworks for evaluating completed assemblies.
Maximize neutrality Neutral Verification Ability to increase, demonstrate, or measure the neutrality of key aspects of a process. Opportunity: High Urgency: Critical Capability: Adequate Consistency: Very variable Global: Poor	Process organizers seek to demonstrate their impartiality by creating governance procedures with strong incentives and review mechanisms. Citizens’ assembly organizers will hire external facilitation teams and commit to their independence through formal governance arrangements. They will allow external auditors or evaluators to review the process and monitor for bias.	Research Algorithmic fairness literature provides mathematical definitions and tests for different types of neutrality (e.g., demographic parity, equalized odds). (Kleinberg, 2016) Research Natural language processing tools to detect biases. (Recasens, 2013; Pryzant, 2020) Practice The French Convention Citoyenne’s uses multi-stakeholder oversight committees to verify neutrality of facilitation. Practice The French Convention Citoyenne pour le Climat published all expert presentations and methodological choices for public scrutiny.	Goal: Neutrality can be defined and measured across different assembly components. Research How can we translate mathematical bias guarantees from algorithmic settings to real-world human facilitation? Research What are the appropriate metrics for measuring neutrality in information presentation, question framing, and synthesis? Research How do we define and measure neutrality when legitimate value disagreements exist about what constitutes "neutral"? Practice Develop standardized neutrality assessment tools that can be applied across different cultural contexts. Goal: Real-time monitoring and correction of neutrality violations is feasible. Research How can we design unobtrusive monitoring systems that don't themselves bias the deliberative process? Research What are acceptable thresholds for intervention when neutrality violations are detected? Research How can facilitators maintain neutrality while also ensuring productive deliberation? Goal: Product "Neutrality dashboards" that provide real-time feedback to facilitators and organizers.
Navigate conflict De-escalation Mediator Ability to address, resolve and navigate conflict that emerges as a result of the process. Opportunity: High Urgency: Moderate Capability: Adequate Consistency: Very variable Global: Limited	Process organizers will isolate and address conflicting participants to mediate or resolve the issue. Citizens’ assembly organizers establish social connections early on and have participants self-write their code of conduct to build norms. Facilitators will monitor for the potential escalation of disagreements and proactively separate and mediate.	Research Deliberative democracy in divided societies (e.g., Dryzek, 2005; O’Flynn and Caluwaerts, 2018; Curato 2025)	Goal: Early warning systems can reliably detect emerging conflicts before escalation. Research How can we identify verbal and non-verbal cues that predict conflict escalation in deliberative settings? Research What role can sentiment analysis and emotion recognition play in real-time conflict monitoring? Research How do we distinguish between productive tension that enhances deliberation and destructive conflict? Goal: Conflict navigation preserves both participant dignity and deliberative quality. Research How can we measure whether conflict resolution preserved or suppressed minority viewpoints? Research What are culturally-sensitive approaches to conflict that work across different contexts? Goal: Post-conflict repair mechanisms restore trust and collaborative capacity. Context: Even well-managed conflicts can leave residual effects on group dynamics. Research What restorative practices are most effective in deliberative settings? Research How can we measure and address the "conflict hangover" effect on subsequent deliberations?
Handle unexpected Resilient Stress tester, Crisis navigator Ability to withstand changing contexts and less-than-ideal conditions. Opportunity: High Urgency: Moderate Capability: Adequate Consistency: Very variable Global: Limited	Process organizers will understand the conditions under which their process breaks, adapting up until that point where possible. Citizens’ assembly organizers will have thresholds for estimating when it will not be possible for a process to meet required outcomes, or when conditions change and adaptations are required to retain process integrity, such as when a significant number of participants do not show up, skewing the representativeness and therefore legitimacy of the process.		Goal: Critical failure thresholds can be identified and monitored in real-time. Context: Knowing when a process risks losing legitimacy allows for timely intervention or principled termination. Research How can we define and measure "minimum viable" conditions for different assembly objectives? Research What are the tipping points where adaptation compromises core democratic values? Practice Develop real-time dashboards that track process health across multiple dimensions? Goal: Adaptive protocols maintain legitimacy while responding to unexpected conditions. Research What pre-commitments and transparency measures best preserve legitimacy during adaptations? Research How do we communicate changes to stakeholders without undermining confidence in outcomes? Goal: Resilience can be built into assembly design from inception. Research How can we systematically stress-test assembly designs before implementation? Research What redundancies and buffers are most cost-effective for different types of disruptions? Research How do we balance efficiency with resilience in resource-constrained environments? Goal: Hybrid human-AI systems can provide legitimate backup mechanisms. Research Under what conditions can AI-simulated participants maintain democratic legitimacy? AI Research How can we ensure simulated participants accurately represent missing demographics? Research What transparency and consent mechanisms are required for hybrid assemblies? AI Research How do we prevent gaming or manipulation of AI backup systems?

Capability

Who & How They Do It

Existing Practice, Products and Research

Gaps

Resist manipulation

Integrity

Guarantor, Supervisor

Ability to resist manipulation that would decrease trustworthiness, legitimacy or unfairly influence the outcome.

Opportunity: High

Urgency: High

Capability: Adequate

Consistency: Good

Global: Limited

Process organizers anticipate vulnerabilities in processes and do their best to mitigate risk with countermeasures.

Citizens’ assembly organizers design processes with an understanding of where manipulation is possible and more likely, and develop mitigating strategies, such as by reinforcing the epistemic capabilities of participants before interacting with new information, developing selection algorithms with manipulation resistance, and establishing governance protocols for impartiality of key actors.

Research Strategyproofness guarantees in voting mechanisms ensure participants cannot benefit from misrepresenting preferences (e.g., theoretical applications) (Satterthwaite, 1975)

AI Research Adversarial testing for Generative AI develops methods to detect and prevent manipulation of AI-assisted deliberation tools.

Research Strategic classification literature examines how actors modify behavior when being evaluated by algorithms (Hardt, 2015; Milli, 2019; Miller, 2020).

Research Sortition algorithms with manipulation resistance (Flanigan, 2021; Baharav, 2024)

Goal: Manipulation attempts can be reliably detected and prevented across different stages of the assembly process.

Research How can we develop real-time detection systems for coordinated manipulation attempts during participant recruitment and selection?
Research What behavioral indicators reliably signal attempts to game deliberative processes?
Research How can we distinguish between legitimate persuasion and manipulative influence in deliberative settings?

Goal: Assembly designs are robust to both internal and external manipulation attempts.

Research How can we quantify and test the manipulation resistance of different assembly design choices?
Research What are the tradeoffs between openness/transparency and manipulation resistance?
Research How can we design information presentation formats that minimize susceptibility to framing effects?

Goal: Post-hoc verification can establish that outcomes were not unduly influenced.

Research How can we develop manipulation impact metrics that distinguish between minor and outcome-altering influences?
Practice Create standardized integrity assessment frameworks for evaluating completed assemblies.

Maximize neutrality

Neutral

Verification

Ability to increase, demonstrate, or measure the neutrality of key aspects of a process.

Opportunity: High

Urgency: Critical

Capability: Adequate

Consistency: Very variable

Global: Poor

Process organizers seek to demonstrate their impartiality by creating governance procedures with strong incentives and review mechanisms.

Citizens’ assembly organizers will hire external facilitation teams and commit to their independence through formal governance arrangements. They will allow external auditors or evaluators to review the process and monitor for bias.

Research Algorithmic fairness literature provides mathematical definitions and tests for different types of neutrality (e.g., demographic parity, equalized odds). (Kleinberg, 2016)

Research Natural language processing tools to detect biases. (Recasens, 2013; Pryzant, 2020)

Practice The French Convention Citoyenne’s uses multi-stakeholder oversight committees to verify neutrality of facilitation.

Practice The French Convention Citoyenne pour le Climat published all expert presentations and methodological choices for public scrutiny.

Goal: Neutrality can be defined and measured across different assembly components.

Research How can we translate mathematical bias guarantees from algorithmic settings to real-world human facilitation?
Research What are the appropriate metrics for measuring neutrality in information presentation, question framing, and synthesis?
Research How do we define and measure neutrality when legitimate value disagreements exist about what constitutes "neutral"?
Practice Develop standardized neutrality assessment tools that can be applied across different cultural contexts.

Goal: Real-time monitoring and correction of neutrality violations is feasible.

Research How can we design unobtrusive monitoring systems that don't themselves bias the deliberative process?
Research What are acceptable thresholds for intervention when neutrality violations are detected?
Research How can facilitators maintain neutrality while also ensuring productive deliberation?

Goal: Product "Neutrality dashboards" that provide real-time feedback to facilitators and organizers.

Navigate conflict

De-escalation

Mediator

Ability to address, resolve and navigate conflict that emerges as a result of the process.

Opportunity: High

Urgency: Moderate

Capability: Adequate

Consistency: Very variable

Global: Limited

Process organizers will isolate and address conflicting participants to mediate or resolve the issue.

Citizens’ assembly organizers establish social connections early on and have participants self-write their code of conduct to build norms. Facilitators will monitor for the potential escalation of disagreements and proactively separate and mediate.

Research Deliberative democracy in divided societies (e.g., Dryzek, 2005; O’Flynn and Caluwaerts, 2018; Curato 2025)

Goal: Early warning systems can reliably detect emerging conflicts before escalation.

Research How can we identify verbal and non-verbal cues that predict conflict escalation in deliberative settings?
Research What role can sentiment analysis and emotion recognition play in real-time conflict monitoring?
Research How do we distinguish between productive tension that enhances deliberation and destructive conflict?

Goal: Conflict navigation preserves both participant dignity and deliberative quality.

Research How can we measure whether conflict resolution preserved or suppressed minority viewpoints?
Research What are culturally-sensitive approaches to conflict that work across different contexts?

Goal: Post-conflict repair mechanisms restore trust and collaborative capacity.

Context: Even well-managed conflicts can leave residual effects on group dynamics.
Research What restorative practices are most effective in deliberative settings?
Research How can we measure and address the "conflict hangover" effect on subsequent deliberations?

Handle unexpected

Resilient

Stress tester, Crisis navigator

Ability to withstand changing contexts and less-than-ideal conditions.

Opportunity: High

Urgency: Moderate

Capability: Adequate

Consistency: Very variable

Global: Limited

Process organizers will understand the conditions under which their process breaks, adapting up until that point where possible.

Citizens’ assembly organizers will have thresholds for estimating when it will not be possible for a process to meet required outcomes, or when conditions change and adaptations are required to retain process integrity, such as when a significant number of participants do not show up, skewing the representativeness and therefore legitimacy of the process.

Goal: Critical failure thresholds can be identified and monitored in real-time.

Context: Knowing when a process risks losing legitimacy allows for timely intervention or principled termination.
Research How can we define and measure "minimum viable" conditions for different assembly objectives?
Research What are the tipping points where adaptation compromises core democratic values?
Practice Develop real-time dashboards that track process health across multiple dimensions?

Goal: Adaptive protocols maintain legitimacy while responding to unexpected conditions.

Research What pre-commitments and transparency measures best preserve legitimacy during adaptations?
Research How do we communicate changes to stakeholders without undermining confidence in outcomes?

Goal: Resilience can be built into assembly design from inception.

Research How can we systematically stress-test assembly designs before implementation?
Research What redundancies and buffers are most cost-effective for different types of disruptions?
Research How do we balance efficiency with resilience in resource-constrained environments?

Goal: Hybrid human-AI systems can provide legitimate backup mechanisms.

Research Under what conditions can AI-simulated participants maintain democratic legitimacy?
AI Research How can we ensure simulated participants accurately represent missing demographics?
Research What transparency and consent mechanisms are required for hybrid assemblies?
AI Research How do we prevent gaming or manipulation of AI backup systems?

Legibility

The extent to which the processes and decisions are accessible, understandable, and verifiable.

To what extent is information (a) accessible, (b) understandable, (c) verifiable about the: (1) processes/ systems used to make decisions? (2) the execution of these processes? (3) decisions being made (4) reasons and inputs feeding into decisions?

Capability	Who & How They Do It	Existing Practice, Products and Research	Gaps
Inform less-involved Parascalability Experience communicator Ability to communicate the “deliberative journey” of a smaller group process to the broader population (especially critical when providing ways for a mass public to participate back with their feedback, perspectives, or direct power via referendums). Opportunity: High Urgency: High Capability: Adequate Consistency: Very variable Global: Limited	Process organizers communicate to the wider public through various media with a focus on conveying qualities that build trust in the process. Citizens’ assembly organizers provide insights into the experience through interviews with participants, asking them to reflect on the process without anticipating outcomes. The goal is to build buy-in to the legitimacy of the process without outcome affiliation biasing reactions.	Product Make.org Panoramic AI Product Cortico Conversation Library Practice CESE Communications Team (YouTube, Instagram, Website) Research French Citizens Convention on Climate, in which the general public were quite sceptical of the process, albeit this was because some thought that the government would just cherry pick what they wanted (Giraudet et al., 2022)	Goal: Non-participating people understand and buy into the process. Research How can AI support the creation of compelling media experiences that support parascaling? Research What are the most compelling features of processes for building trust? Research What methods can help the wider public understand how and why a process is set-up? Research How to create a visual and engaging public archive of deliberations? Goal: Non-participating people can monitor the process. Research How to communicate outputs effectively to different audiences (policy-makers, media, general public) without losing essential nuance and decisiveness? Research How can we help citizens trace a process’ impact on policy-making? Research What are the opportunities and pitfalls of live streaming/recording processes for wider public following?
Ensure transparency Transparency Observer Ability for the process to be open to the public (where possible given privacy considerations). Opportunity: Moderate Urgency: Moderate Capability: Adequate Consistency: Variable Global: Limited	Process organizers determine their openness to observation and the extent to which they will make process documentation public. Citizens’ assembly organizers allow public observers under reasonable conditions of distance and non-involvement. Items like background documents, speaker lists, agendas, final outputs and commissioning authority response documents are usually public. Governance processes may involve independent guarantors witnessing key ‘closed room’ decisions.	Practice Five lessons from the College of Guarantors of the French Citizens’ Convention on the End of Life	Goal: The use of digital tools is as transparent as possible. Research What design choices help promote the transparency of deliberative technologies, and what trade-offs does this raise? Research Does the integration of deliberative technologies raise fundamentally new transparency challenges to processes and if so, what are they? Research To what extent can we clearly communicate the inner workings of AI-augmented deliberative tools?
Make verifiable Proof Verifier Ability for integrity of the process to be verified and audited. Opportunity: Moderate Urgency: Moderate Capability: Limited Consistency: Limited Global: Limited	Process organizers key details are made directly observable by auditors. Citizens’ assembly organizers demonstrate selection processes, allow independent observation for key process details and in some cases record or document discussions for note taking or publication later on.		Goal: Key process details are able to be independently verified. Research How can automatic logging of key events improve access for verifiers?

Capability

Who & How They Do It

Existing Practice, Products and Research

Gaps

Inform less-involved

Parascalability

Experience communicator

Ability to communicate the “deliberative journey” of a smaller group process to the broader population (especially critical when providing ways for a mass public to participate back with their feedback, perspectives, or direct power via referendums).

Opportunity: High

Urgency: High

Capability: Adequate

Consistency: Very variable

Global: Limited

Process organizers communicate to the wider public through various media with a focus on conveying qualities that build trust in the process.

Citizens’ assembly organizers provide insights into the experience through interviews with participants, asking them to reflect on the process without anticipating outcomes. The goal is to build buy-in to the legitimacy of the process without outcome affiliation biasing reactions.

Product Make.org Panoramic AI

Product Cortico Conversation Library

Practice CESE Communications Team (YouTube, Instagram, Website)

Research French Citizens Convention on Climate, in which the general public were quite sceptical of the process, albeit this was because some thought that the government would just cherry pick what they wanted (Giraudet et al., 2022)

Goal: Non-participating people understand and buy into the process.

Research How can AI support the creation of compelling media experiences that support parascaling?
Research What are the most compelling features of processes for building trust?
Research What methods can help the wider public understand how and why a process is set-up?
Research How to create a visual and engaging public archive of deliberations?

Goal: Non-participating people can monitor the process.

Research How to communicate outputs effectively to different audiences (policy-makers, media, general public) without losing essential nuance and decisiveness?
Research How can we help citizens trace a process’ impact on policy-making?
Research What are the opportunities and pitfalls of live streaming/recording processes for wider public following?

Ensure transparency

Transparency

Observer

Ability for the process to be open to the public (where possible given privacy considerations).

Opportunity: Moderate

Urgency: Moderate

Capability: Adequate

Consistency: Variable

Global: Limited

Process organizers determine their openness to observation and the extent to which they will make process documentation public.

Citizens’ assembly organizers allow public observers under reasonable conditions of distance and non-involvement. Items like background documents, speaker lists, agendas, final outputs and commissioning authority response documents are usually public. Governance processes may involve independent guarantors witnessing key ‘closed room’ decisions.

Practice Five lessons from the College of Guarantors of the French Citizens’ Convention on the End of Life

Goal: The use of digital tools is as transparent as possible.

Research What design choices help promote the transparency of deliberative technologies, and what trade-offs does this raise?
Research Does the integration of deliberative technologies raise fundamentally new transparency challenges to processes and if so, what are they?
Research To what extent can we clearly communicate the inner workings of AI-augmented deliberative tools?

Make verifiable

Proof

Verifier

Ability for integrity of the process to be verified and audited.

Opportunity: Moderate

Urgency: Moderate

Capability: Limited

Consistency: Limited

Global: Limited

Process organizers key details are made directly observable by auditors.

Citizens’ assembly organizers demonstrate selection processes, allow independent observation for key process details and in some cases record or document discussions for note taking or publication later on.

Goal: Key process details are able to be independently verified.

Research How can automatic logging of key events improve access for verifiers?