Source-linked AI summary
Envisioning Communities: A Participatory Approach Towards AI for Social Good
Elizabeth Bondi, Lily Xu, Diana Acosta-Navas, Jackson A. Killian
TL;DR
AI for social good lacks an agreed definition, and utility-centered evaluations may overlook distribution and fundamental liberties. The paper proposes PACT, combining the capabilities approach with participation by affected communities. It concludes with guiding questions for research that elicits and respects communities’ definitions of social good.
Problem
Method
PACT combines the capabilities approach with participation by affected communities throughout AI research and gives them influence over desired capability sets.
Source: §1 · p. 2, §7 · p. 10Results
The paper argues that a community-centered, participatory approach grounded in capabilities is a way to define AI for social good and assess projects.
Source: §7 · p. 10Takeaways & Limitations
The authors offer guiding questions to help researchers continually reassess their work and respect a community’s own definition of social good.
Source: §6 · p. 6, §7 · p. 10Participatory community interactions can become exploitative without intention, compensation, or long-term engagement.
Source: §6 · p. 7
Abstract
from arXiv · showhide
Research in artificial intelligence (AI) for social good presupposes some definition of social good, but potential definitions have been seldom suggested and never agreed upon. The normative question of what AI for social good research should be "for" is not thoughtfully elaborated, or is frequently addressed with a utilitarian outlook that prioritizes the needs of the majority over those who have been historically marginalized, brushing aside realities of injustice and inequity. We argue that AI for social good ought to be assessed by the communities that the AI system will impact, using as a guide the capabilities approach, a framework to measure the ability of different policies to improve human welfare equity. Furthermore, we lay out how AI research has the potential to catalyze social progress by expanding and equalizing capabilities. We show how the capabilities approach aligns with a participatory approach for the design and implementation of AI for social good research in a framework we introduce called PACT, in which community members affected should be brought in as partners and their input prioritized throughout the project. We conclude by providing an incomplete set of guiding questions for carrying out such participatory AI research in a way that elicits and respects a community's own definition of social good.
1 INTRODUCTION
Because AI for social good lacks a clear definition, the paper argues that affected communities should determine whether a project serves social good in its context. It proposes PACT, grounding participatory assessment in the capabilities approach rather than aggregate benefit alone.
1 INTRODUCTION
- AI for social good has a nebulous definition, leaving it unclear which projects qualify, and context prevents a single tractable rule from deciding what counts as social good.The paper asks whether, for example, a contact-tracing app belongs in the category despite privacy risks.Source: §1 · p. 1
- Affected communities should determine whether a project may bring about social good in the context where they live.The paper connects this position to calls to elevate traditionally marginalized populations.Source: §1 · p. 1
- PACT integrates capabilities-based ethical guidelines with assessment across stakeholder groups and projects, rejecting aggregate net benefit when it disregards distribution and injustice.The authors warn that additive accounting can favor majorities and exclude underrepresented community members from design.Source: §1 · p. 2
- The capabilities approach evaluates substantive liberties and opportunities to pursue valued ways of being and doing, requiring concrete steps that empower all affected community members.It is intended to attend to inequalities that aggregate utility measures can overlook.Source: §1 · p. 2
- PACT uses participation so communities can define the substantive liberties and functionings that constitute social good in their context.The paper describes this as mutual learning among participants.Source: §1 · p. 2
2 GROWING CRITICISMS OF AI FOR SOCIAL GOOD
Critiques of AI for social good question its ethical direction, local grounding, and vague definition. The paper situates its approach among calls to shift power toward affected communities and efforts to define the field.
2 GROWING CRITICISMS OF AI FOR SOCIAL GOOD
- Critics argue that AI for social good lacks a moral compass and that industry-led technosolutionism can overlook local expertise and context.Related calls urge researchers to examine how their choices normalize particular kinds of research and to remember the people behind the numbers.Source: §2 · p. 2
- AI for social good’s vague definition can encompass projects with conflicting aims, such as police reform and predictive policing, and may steer attention toward incremental fixes over broader reform.The authors report critiques that this vagueness impedes the field’s progress and clarity.Source: §2 · p. 2
- Decolonial AI calls for systems that challenge colonial oppression through critical technical practice, reciprocal co-design, and solidarity with affected communities [55].Kalluri similarly urges scrutiny of power structures and projects that empower intended beneficiaries, including through auditing or recourse.Source: §2 · pp. 2–3
- Existing AI for social good guidance includes ethical principles and recommendations, but its broad definition leaves room for projects with widely diverging outcomes.Other work offers rough guidelines in lieu of a definition.Source: §2 · p. 3
3 THE CAPABILITIES APPROACH
The capabilities approach offers an alternative to utility aggregation by assessing substantive liberties, their distribution, and the opportunities people have reason to value. The paper presents this approach as a way to assess social good while attending to power and community-defined priorities.
3 THE CAPABILITIES APPROACH
- Utilitarian aggregation can overlook minority effects, while both maximizing utility and threshold-based net improvements may disregard distribution, liberties, and preferences shaped by circumstance.The contact-tracing example illustrates tensions between public-health benefit and privacy rights, with losses potentially differing across groups.Source: §3 · p. 3
- The authors propose assessing social good through opportunities to live valued lives, considering how resource distribution and environmental factors create or eliminate those opportunities.They treat capabilities as intermediate landmarks toward shifting power to disadvantaged people, who can decide what that shift should mean.Source: §3 · p. 3
- The capabilities approach centers substantive liberties and the conditions that let people pursue valued ways of being and doing, rather than treating aggregate utility as sufficient.The UN Human Development Index was inspired by this approach, which shifted assessment away from GDP toward people-centered policies.Source: §3 · pp. 3–4, Figure 2 · p. 4
- The approach addresses power through attention to social environments, institutions, group relations, and practices, and explicitly emphasizes enhancing the power of society’s most vulnerable members.Capabilities are presented as intermediate steps toward shifting power and broadening liberties for historically deprived groups.Source: §3 · p. 4
- The paper builds on prior AI-capabilities work by arguing that capabilities should be identified participatorily, distinguishing its approach from work using Nussbaum’s proposed capability set rather than eliciting patients’ desired capabilities.It cites a trafficking-identification app and an empirical study of women’s health in India as examples of capabilities-oriented AI4SG work.Source: §3 · p. 4
4 AI AND CAPABILITIES
The paper frames AI’s potential contribution to social good in terms of expanding and equalizing capabilities. It calls for assessing how projects alter people’s options and the conditions that shape them, with community members judging relevant trade-offs.
4 AI AND CAPABILITIES
- AI projects can be assessed along capability expansion and distribution, and may contribute to equalizing and enhancing capabilities when they interact productively with other social factors.The paper gives a visual question-answering system for visually impaired users as an example of expanding access to environmental information.Source: §4 · p. 4
- Capabilities depend on personal characteristics, accessible external resources, and the wider material and social environment, not merely freedom from obstruction.Figure 3 depicts AI projects as potentially altering these elements and the resulting capability sets.Source: §4 · p. 4, Figure 3 · p. 5
- AI can expand choice sets by providing tools or resources and by fostering skills, abilities, and traits, while also altering the environments in which people act.Examples include navigation, communication, health diagnoses, and climate forecasts.Source: §4 · p. 4
- Predictive policing may increase security while also reinforcing disparate law-enforcement impacts and reducing vulnerable community members’ substantive liberties.Such changes can alter both individual capability sets and how liberties are distributed.Source: §4 · p. 5
- Assessing trade-offs requires alignment with opportunities community members value; localized utility gains do not justify reducing other members’ ability to choose valued lives.The paper states that affected community members must make this assessment.Source: §4 · p. 5
5 COMMUNITY PARTICIPATION
PACT places affected communities in a continuing partnership with AI researchers, giving stakeholders a say in tradeoffs and embedding participation throughout project stages. The approach begins by defining capability sets and evaluates success by whether projects realize them.
5 COMMUNITY PARTICIPATION
- Community members should help navigate capability tradeoffs because AI-based interventions may affect groups differently and can open or close alternatives.PACT calls for consultation with stakeholders from all groups on their preferred ways to navigate those tradeoffs.Source: §5 · p. 5
- Stakeholder-informed projects may support fertile functionings, where securing one functioning is likely to secure others.Examples include work and social affiliations.Source: §5 · p. 5
- PACT aligns with Design Justice by enabling people to decide which capability sets they value, potentially supporting other capabilities and shifting power.The paper says this could plausibly act against disadvantage when participatory space is offered to the most vulnerable.Source: §5 · p. 5
- PACT applies participation to choosing capability sets at project outset and evaluating impacts on groups’ substantive liberties.Source: §5 · p. 5
- The approach embeds community participation throughout project stages, beginning with capability sets rather than immediately applying established tools to social issues.It calls for iterative evaluation of whether the desired capability sets were realized.Source: §5 · pp. 5–6, Figure 4 · p. 6
6 GUIDING PRINCIPLES FOR A PARTICIPATORY APPROACH
The proposed guiding principles keep community-defined capabilities central, with continued participation in project decisions and evaluation. They address partner representation, capability priorities, tradeoffs, and the roles of researchers.
6 GUIDING PRINCIPLES FOR A PARTICIPATORY APPROACH
- The guiding questions are ongoing navigation prompts, not yes-or-no checks, and call for centering desired capabilities without negatively affecting others.The authors expect that nearly all questions will remain unresolved and should prompt regular reassessment.Source: §6 · p. 6, Figure 5 · p. 7
- Identify impacted groups with partners, include representatives from each group where possible, and maintain consultation under agreed partnership norms.The authors also recommend that community representatives and experiential experts have the power to end the partnership.Source: §6 · p. 6
- Compensate stakeholders and recognize substantial contributors; without intention, compensation, or long-term interaction, participatory work can become exploitative.Source: §6 · p. 7
- Deliberative dialogue may seek overlapping consensus, while simple majority voting can neglect marginalized groups; preference aggregation also has theoretical limits.The cited discussion notes Arrow’s impossibility theorem’s limitations with as few as three voters, while saying these need not inhibit practical system design.Source: §6 · pp. 7–8
6.1 Determining Capability Sets
- Choose capability sets through community consensus and consider fertile functionings that may secure additional capabilities.The paper gives reading as an example that may support work, social association, and control over one’s environment.Source: §6.1 · pp. 8–9
- Prioritize vulnerable groups and stratify data: women were 47% more likely to be seriously injured in crashes, while women lost 156,000 jobs and men gained 16,000 in December 2020.Source: §6.1 · p. 8
6.2 Evaluating AI for Social Good
- Communities should judge whether a system succeeds against its intended capabilities, although measuring capabilities is difficult and the authors endorse no measurement method.Source: §6.2 · p. 9
- Evaluate capability effects throughout and beyond the project lifecycle, using data or feedback while checking for harms to other valued capabilities.The authors describe data analysis and feedback through local organizations as possible approaches, while noting examples may not apply to every system.Source: §6.2 · p. 9
7 CONCLUSION: THOUGHTS ON AI FOR SOCIAL GOOD AS A FIELD
PACT centers marginalized community members in defining and shaping capability-focused AI for social good. The authors argue that supporting this work requires changes to research incentives and evaluation, alongside broader inclusion in AI research.
7 CONCLUSION: THOUGHTS ON AI FOR SOCIAL GOOD AS A FIELD
- PACT prioritizes marginalized community members throughout AI design, development, and deployment, letting them choose desired capability sets and influence their realization.The authors frame this as a community-centered approach to defining AI for social good.Source: §7 · p. 10
- Participatory community work requires substantially more time, energy, and resources than typical AI research, while field deployment can take months or years versus a week for an experimental section.The authors note that this timeline weighs especially heavily on PhD students and untenured faculty evaluated over shorter timescales.Source: §7 · p. 10
- Institutions can support community engagement by requiring it in public-facing AI research and incentivizing meaningful engagement with impacted communities.The authors argue this can accompany the production of more impactful research.Source: §7 · p. 10
- Conference evaluation should value potential social impact, nontechnical contributions, and situated programs that may not focus on generalizability.The authors also encourage recognition for socially beneficial work through awards.Source: §7 · p. 10
- Researchers can use interdisciplinary and domain-specific venues to share community-focused work; the authors argue broader reforms could make participatory approaches the norm in AI4SG.Suggested venues include COMPASS, IAAI, MD4SG/EAAMO, LIMITS, and ecology or public health conferences.Source: §7 · p. 10
- AI research must also welcome all communities as researchers to address inequities in access to resources, collaborators, and opportunities.The paper cites systemic anti-Blackness and reports that around 20% of tenure-track computer science faculty were women as of 2018.Source: §7 · p. 10