AI for Social Good (AI4SG)

AI for Social Good (AI4SG) is sponsored by National Science Foundation (NSF). The AI For Social Good (AI4SG) project offers curriculum that engages students from all disciplines to propose or develop AI-powered solutions that address pressing issues in their communities, aligning with United Nations Sustainable Development Goals.

Funding AI for Good: A Call for Meaningful Engagement Funding AI for Good: A Call for Meaningful Engagement Computer Science Harvard University Cambridge Massachusetts United States hongjin˙lin@g. harvard.

edu Computer Science Carnegie Mellon University Pittsburgh Pennsylvania United States Urban Studies and Planning MIT Cambridge Massachusetts United States Computer Science Carnegie Mellon University Pittsburgh Pennsylvania United States Computer Science Harvard University Cambridge Massachusetts United States Artificial Intelligence for Social Good (AI4SG) is a growing area that explores AI’s potential to address social issues, such as public health.

Yet prior work has shown limited evidence of its tangible benefits for intended communities, and projects frequently face inadequate community engagement and sustainability challenges. While existing HCI literature on AI4SG initiatives primarily focuses on the mechanisms of funded projects and their outcomes, much less attention has been given to the funding agenda and rhetoric that influences downstream approaches.

Through a thematic analysis of 35 funding documents—representing about $410 million USD in total investments, we reveal dissonances between AI4SG’s stated intentions for positive social impact and the techno-centric approaches that some funding agendas promoted, while also identifying funding documents that scaffolded community-collaborative approaches for applicants.

Drawing on our findings, we offer recommendations for funders to embed approaches that balance both contextual understanding and technical capacities in future funding call designs. We further discuss how the HCI community can positively shape AI4SG funding design processes. Figure 1.

AI4SG aims to bring about positive, long-lasting social impact. Prior literature emphasizes that impactful innovations are the result of a balance of technology capacities and contextual understanding (right panel). A bias towards technology capacities leads to innovations without a real-world positive impact (left panel).

We use this spectrum as a lens to investigate how current AI4SG funding documents orient towards more techno-centric or balanced approaches. Artificial Intelligence (AI) for social good (AI4SG) initiatives aim to create long-lasting positive impact in social domains such as public health, by leveraging AI technologies like machine learning (Shi et al. , 2020 ; Li et al.

, 2021 ; Perrault et al. , 2020 ; Rolnick et al. , 2022 ; Tomašev et al.

, 2020 ) . Yet prior work finds limited evidence of its tangible benefits for intended communities, and projects frequently face deployment and sustainability challenges in real-world contexts (Vinuesa et al. , 2020 ; Eubanks, 2018 ; SALLY HO and GARANCE BURKE, 2023 ; Tate Ryan-Mosley, 2023 ) .

HCI scholarship provides insights into the socio-technical mechanisms underpinning these challenges, and recent work points to the significant influence of funding agendas—the priorities, goals, and approaches set by funders—in shaping the technology initiatives aimed at social impact (Lin et al. , 2024 ; Saha et al. , 2022a ; Erete et al.

, 2023 ) . In AI4SG specifically, Lin et al. ( 2024 ) show from community organizations’ perspectives that funding agendas heavily influence how partnerships are formed, which problem spaces are prioritized, and what kinds of solutions are pursued during the early stages of projects.

Prior work also notes that funding often incentivizes mission-driven organizations to adopt data-driven innovations, sometimes at the cost of their original social missions (Bopp et al. , 2017 ; Lin et al. , 2024 ) .

For AI developers and practitioners, short funding timelines can constrain considerations of ethics and unintended consequences (Do et al. , 2023 ; Moitra et al. , 2022 ; Okolo and Lin, 2024 ) .

However, existing HCI literature on AI4SG initiatives primarily focuses on funded projects ’ approaches and outcomes through the perspectives of project teams (typically consisting of AI technologists and mission-driven organizations); much less attention is given to the design of funding agendas themselves, before projects even get selected.

In this work, we begin addressing this gap by investigating AI4SG funding agendas, following up on Laura Nader’s call to “study up” (Nader, 1972 ) systems and actors that exert normative influence. In doing so, we aim to shift attention from assigning blame to AI4SG project teams, and towards recognizing the broader power structures shaping their approaches.

We turn to public funding documents—textual instruments that communicate funding agendas, including calls for proposals and grant announcements. Prior literature has established that funding documents are not neutral instruments: they construct priorities (Palmero et al. , 2025 ) , shape applicant behaviors (Peng et al.

, 2024 ) , and structure downstream research activities (Widder et al. , 2024 ; Smith et al. , 2023 ) .

In AI4SG, by defining who is eligible, what constitutes a “problem,” how these problems should be addressed, and what “success” means, funding documents construct discourses about AI and social good, signaling what approaches are fundable or imaginable.

Although there is complexity in how written funding documents influence what actually gets funded, examining them reveals how priorities, values, and power relations are embedded even before any project begins. Funding documents can reinforce dominant AI imaginaries —“collective visions, beliefs, symbols, and expectations that individuals and communities hold about AI and its potential outcomes” (Zhong et al.

, 2025 ) —as seen in public texts portraying AI as “inevitable and massively disrupting” (Sophie Bennani-Taylor, 2024 ; Bareis and Katzenbach, 2022 ) . Yet they can also guide applicants toward community engagement strategies previously unfamiliar to them.

In analyzing funding documents, our goal is not to evaluate how they eventually lead to funded project outcomes, but to provide thematic insights into the rhetoric and conceptual framing of AI4SG in funding documents that influence downstream approaches .

Specifically, we ask the following research questions: RQ1 : How do funding documents frame social good, and what is AI’s envisioned role in advancing social good, as articulated in funding documents? RQ2 : How do funding documents describe the expected outcomes of AI4SG projects? RQ3 : How do funding documents characterize how these goals might be achieved?

We collected and conducted a reflexive thematic analysis of 35 funding documents—19 funding calls and 16 grant announcements—representing about $410 million USD in total investments (see Appendix A for a complete list). These documents represent a range of efforts, broadly framed under the banners of AI or data-driven technologies for “social good,” “development,” and other social domains.

Our dataset includes funding programs from 23 distinct funders, including private technology companies, philanthropic foundations, nonprofit organizations, and government agencies based in Western societies such as the United States.

Since AI4SG builds on earlier “tech for good” efforts (Aula and Bowles, 2023 ) , we ground our analysis on lessons learned from Computer-Supported Cooperative Work (CSCW), Information and Communication Technologies and Development (ICTD), and Human-Computer Interaction for Development (HCI4D).

This prior literature has extensively examined different approaches in socially driven technology projects, emphasizing that impactful innovations require both a thorough understanding of the problem and appropriate technological capabilities (Figure 1 ).

From our synthesis of this body of work, a spectrum from techno-centric to balanced approaches emerges, indicating a strong association between where projects fall on this spectrum and project outcomes.

A techno-centric approach assumes novel technology alone can solve complex social issues, often producing suboptimal results and strained relationships between researchers and communities (e.g., (Ames, 2019 ; Karusala and Kumar, 2017 ; Walsham, 2017 ; Dada, 2006 ; Brown and Mickelson, 2019 ) ).

In contrast, a more balanced approach, grounded in a solid understanding of the community’s challenges and contexts through community-collaborative approaches (Cooper and Zafiroglu, 2024 ) , tend to result in more community buy-in and sustainability (e.g., (Toyama, 2015 ; Gandhi et al. , 2007 ; Burrell and Toyama, 2009 ) ).

Using this spectrum, our study analyzes how current AI4SG funding rhetoric is oriented towards a more techno-centric or a more balanced approach. Funding documents in our corpus overwhelmingly mobilized AI against large-scale social problems that cannot be fully addressed with any single intervention (notably public health and development).

Yet several funding documents framed AI’s potential as deterministically beneficial and positioned AI interventions as solutions to these social problems while skirting contextual and stakeholder specificity. Some other documents scoped interventions to more well-defined sub-problems and explicitly acknowledged both AI’s benefits and risks.

Expected outcomes of funding documents in our dataset frequently centered on technical deliverables, while a subset adopted broader success metrics attentive to benefits to the communities involved. Furthermore, although interdisciplinary collaboration was widely encouraged, funding documents varied in their emphasis and scaffolding of community engagement and who were considered “experts.

” With a few exceptions, expectations for community engagement stopped at a consultation mode towards later stages of projects (e.g., implementation stage), where community stakeholders were expected to provide inputs but not able to shape the broader project scopes.

To the best of our knowledge, this study presents the first qualitative analysis of how AI4SG funding documents frame social good, envision AI’s role in advancing it, and anticipate desired outcomes to be achieved.

While prior work has largely focused on funded projects’ tendency towards more techno-centric approaches, we uncover how both techno-centric and balanced approaches are instantiated in different components of funding documents—before project selection even occurs.

In doing so, this work makes three contributions to existing HCI literature on “tech for good” initiatives and document analysis of AI imaginaries: First, we provide a novel focus on AI4SG funding documents as a discourse instrument rather than neutral texts.

Through a reflexive thematic analysis (Braun and Clarke, 2006 , 2019 ) , we highlight the disconnects between the positive social impact AI4SG aims for and the techno-centric approaches promoted in some funding documents (Section 5. 1 ).

Second, building on the more balanced approaches and gaps identified in our dataset, we propose opportunities for designing future funding calls for funders (including supporting relationship-building, maintenance post-project deployment, and community engagement training for project teams), and outline future research directions for HCI researchers (Section 5. 2 ).

Finally, we call for meaningful community engagement in the funding design process for funders, and specify the roles that HCI practitioners and researchers can play in shaping AI4SG funding agendas and design processes (Section 5. 3 ). Prior HCI research on AI4SG has highlighted both the objective of existing funded projects to produce positive social impact and the challenges they face in doing so (Section 2.

1 ). Scholars have called on AI4SG to draw on lessons from previous “tech for good” projects, which emphasize that impactful innovations require a balance of appropriate technical capacity and deep contextual understanding (Section 2. 2 ).

Given the central role of funding agendas in shaping the AI4SG ecosystem, we draw on emerging scholarship examining grant-making and private tech companies’ investment in AI4SG (Section 2. 3 ).

Furthermore, prior research in HCI and innovation studies has enacted document analysis to examine how AI is imagined in public texts and established the role of funding documents in shaping research priorities and collaboration structures (Section 2. 4 ). However, HCI has yet to systematically analyze funding documents that define and construct AI4SG initiatives.

This study begins to address this gap by analyzing AI4SG funding documents through the lens of existing literature on “tech for good. ” While prior work points out funded projects’ bias towards emphasizing technical capacity, our work contributes an analysis of how this bias is instantiated across some funding documents, while also identifying more balanced approaches that emphasize contextual understanding. 2.

1. The Promises and Challenges of AI for Social Good Artificial Intelligence for Social Good (AI4SG) has emerged as a growing field of use-inspired research and real-world implementations focused on applying AI to complex social issues (Shi et al. , 2020 ; Perrault et al.

, 2020 ; Tomašev et al. , 2020 ; Rolnick et al. , 2022 ) .

Its stated goal is to deliver transformative, positive, and lasting impact across many social aspects (Perera, 2024 ; Cowls, 2021 ; Tudor et al. , 2024 ) . Some notable examples include projects in agriculture that predict crop yields using remote sensing data (You et al.

, 2017 ) and provide price forecasting for farmers in India (Ma et al. , 2019 ) . While enthusiasm for AI4SG remains high, growing evidence shows that many projects fall short of delivering lasting positive social impact and often end at the publication stage without deployment (Lin et al.

, 2024 ; Aula and Bowles, 2023 ; Ben Green, 2019 ) . In high-stakes contexts such as public service screening or criminal justice, AI may be inappropriate altogether (Eubanks, 2018 ; SALLY HO and GARANCE BURKE, 2023 ; Pruss, 2023 ) .

Additionally, AI4SG is often shaped by promotional hype (Joyce and Cruz, 2024 ) , magical thinking (Lupetti and Murray-Rust, 2024 ) , and “enchanted determinism” (Campolo and Crawford, 2020 ) , where the fascination with the mysterious capabilities of new technologies captures public attention toward techno-solutionism—where complex, structural challenges are reduced to problems amenable to AI interventions (Morozov, 2013 ) .

This could be exacerbated by AI’s transparency issues (especially in the Global South context that AI4SG projects are intended for (Okolo and Lin, 2024 ) ). For instance, studies in India show a strong belief in “AI authority” over public institutions and lowered expectations for accountability (Kapania et al. , 2022 ; Ramesh et al.

, 2022 ) . HCI scholars have started to examine the mechanisms behind AI4SG projects. We highlight two key insights below.

First, the aspiration to employ AI for social good is not a recent phenomenon; AI4SG should draw lessons learned from prior “tech for good” movements (Aula and Bowles, 2023 ; Lin et al. , 2024 ) . ICTD and HCI4D have long explored the development and use of emerging technologies in marginalized communities, especially in the Global South (Walsham, 2017 ; Dell and Kumar, 2016 ; Avgerou, 2010 ) .

Second, while past work often focuses on investigating perspectives of project teams —typically AI developers and nonprofit organizations, emerging research highlights funding agendas’ role in shaping interdisciplinary partnerships (Lin et al. , 2024 ; Saha et al. , 2022a ; Erete et al.

, 2023 ) . Through interviews with staff of community organizations participating in AI4SG, Lin et al. ( 2024 ) found that funding agenda influences program goals, proposed solutions, and definitions of success.

Focusing on nonprofit organizations, other studies show how funding agendas shape their incentives to adopt emerging tech, and collaboration choices (Bopp et al. , 2017 ; Erete et al. , 2016 ; Goldenfein and Mann, 2023 ) .

AI developers and researchers’ approaches to AI ethics and considerations of unintended consequences are also shown to be impacted by funding timeline (Do et al. , 2023 ; Widder et al. , 2023 ; Moitra et al.

, 2022 ) . Emphasizing the need to examine how power structures impact technology development in social contexts, scholars have taken up Laura Nader’s (1972) call to “study up” by focusing on systems and actors that shape social systems and exert normative influence (Nader, 1972 ) . Emerging “study up” scholarship in algorithmic fairness (Barabas et al.

, 2020 ) , public sector AI (Kawakami et al. , 2024a ) , and machine learning data (Miceli et al. , 2022 ) demonstrates that examining those in power can reveal critical leverage points for more accountable AI development.

While prior work has highlighted funding’s significant influence on project approaches and outcomes, we still lack a direct understanding of how funding frames AI’s role in addressing social good challenges; there is a need to “study up” AI4SG funding. 2. 2.

Lessons from Prior “Tech for Good” Projects: Techno-Centric vs Balanced Approaches Prior work in CSCW, ICTD, and HCI4D has established a substantial body of work on technology projects aimed at social impact. Despite existing areas of debate within this body of work, there is a broad consensus that impactful innovations require not only engineering expertise but also situated knowledge and community engagement.

Review studies on ICTD projects specifically point out that a bias towards technical problems while overlooking community contexts often leads to limited adoption and community impact (Brown and Mickelson, 2019 ; Heeks, 2003 ; Dodson et al. , 2012 ) , whereas sustained community engagement throughout the project life-cycle is a key factor for success (Brown and Mickelson, 2019 ) .

In this section, we synthesize this rich body of work and categorize characteristics of projects into a more techno-centric approach (bias towards technical solutions) and a more balanced approach (emphasizing both technical capacities and contextual understanding).

We particularly focus on aspects of motivations, collaborations, longevity and process, and evaluations, because these aspects are closely influenced by funding agendas (Lin et al. , 2024 ) . 2.

2. 1. Techno-Centric Approaches A techno-centric approach, as defined by Dodson et al.

( 2012 ) , “place[s] the technology at the center of the intervention by giving prominence to technical features and capabilities. ” It is rooted in “techno-solutionism,” where complex, structural challenges are reduced to problems amenable to technical interventions (Morozov, 2013 ) , often under the false assumption that these solutions are neutral or apolitical (Feenberg, 2002 ; Winner, 2010 ) .

While leveraging emerging technology as a silver bullet to social problems appears optimistic and well-intentioned, Berlant’s concept of “cruel optimism” points to the cruel side of that optimism: promising social transformation while deepening surveillance, reproducing bias, or ignoring community needs (Berlant, 2011 ) .

These initiatives often fail due to “design-reality gaps” (Heeks, 2002 ) : a mismatch between system designers’ perceptions of local contexts and local actuality. In “tech for good”, techno-centric projects are driven by an “innovate at all costs” mentality, leading to the creation of technologies with little maintenance support or local capacity building afterward (Saha et al. , 2022b ) .

It incentivizes a top-down design approach, where the solutions are typically designed by external technologists without adequate engagement with the local communities (Walsham, 2017 ; Dada, 2006 ; Harris, 2016 ; Dodson et al. , 2012 ) , especially during the decision-making stages of projects (Saha et al. , 2022a ) .

It could also lead to the phenomenon of “bungee jumping research” (Dearden and Tucker, 2016 ) , characterized by brief, superficial engagement (Saha et al. , 2022b ; Sultana et al. , 2019 ; Dada, 2006 ) .

Development of the solutions tends to follow a linear and short-term process: building the technology, deploying it, and moving on to the next project, where researchers often leave after their work is published, making it difficult to sustain the systems they have introduced (Taylor et al. , 2013 ; Dada, 2006 ; Walsham, 2017 ) .

Last but not least, evaluations of techno-centric projects constrain assessments of success to technical dimensions, narrowly focusing on metrics such as model accuracy or the production of tangible artifacts like software (Dodson et al. , 2012 ; Lin et al. , 2024 ) .

2. 2. 2.

Balanced Approaches A balanced approach is grounded in the understanding of problems faced by specific communities, in addition to leveraging emerging technology’s capacities. Community-collaborative approaches (CCA) are central to a balanced approach. A systematic review by Cooper et al.

defines CCA as projects that “involve collaboration with community stakeholders as co-researchers throughout the research process while investigating complex problems of concern to the community (e.g., social and health-related), and developing novel technologies or sociotechnical solutions” (Cooper et al. , 2022 ) . CCA includes work in participatory design (e.g., (Delgado et al.

, 2023 ; Bondi et al. , 2021 ) ), community-based participatory research (e.g. (Wan et al. , 2023 ; Liang et al.

, 2023 ; Strohmayer et al. , 2017 ) ), and action research (e.g., (Hayes, 2011 ; Le Dantec, 2016 ; Lemaire and Muñiz, 2011 ; Balestrini et al. , 2014 ) ).

Projects that follow CCA are more likely to create lasting social impact because they have community support and input from the start (Gandhi et al. , 2007 ) and amplify positive values already present in the community (Toyama, 2015 ) . Brown and Mickelson ( 2019 ) emphasize the necessity of “the full inclusion of all relevant parties in every aspect of the project” to ensure project success.

A balanced approach emphasizes that community members’ lived experiences are a legitimate form of knowledge within their specific contexts (Erete et al. , 2023 ; Cooper et al. , 2022 ) , on par with technical expertise.

This emphasis is grounded in the concept of epistemic justice (Fricker, 2007 ) , which calls for fairness in recognizing who is seen as a knower and what forms of knowledge are valued. In “tech for good” projects, this means embracing modalities of knowledge sharing that go beyond traditional academic outputs, such as storytelling and community observations (Erete et al. , 2023 ; Harrington et al.

, 2019a ; Zegura et al. , 2018 ) . Zegura et al.

( 2018 ) demonstrate that valuing the community’s local knowledge is not only crucial for the quality of data and the subsequent data analysis, but also generates “care” among researchers and community collaborators (Zegura et al. , 2018 ) . Relationships among collaborative teams are critical for the feasibility and sustainability of community-based projects (Dell and Kumar, 2016 ; Kumar and Dell, 2018 ) .

Sustainable relationships with communities can also be cultivated through regular presence, trust-building, and iterative design processes, and they enable project agendas to emerge organically from (rather than being imposed on) communities (Dearden and Tucker, 2016 ; Carroll and Rosson, 2013 ) .

Lastly, a balanced approach extends evaluation beyond technical outputs to include community outcomes such as increased technical capacity (Zegura et al. , 2018 ) , impact on everyday practice (Taylor et al. , 2016 ) , and “empowerment of the participants and the knowledge gained by the technical designers” (Drain et al.

, 2021 ) . Prior work has started to evaluate the extent to which an AI project grants agency to community stakeholders in different stages of projects (Corbett et al. , 2023 ; Feffer et al.

, 2023 ; Delgado et al. , 2023 ) . For example, Delgado et al.

( 2023 ) develop the Parameters of Participation framework that lays out the dimensions of participatory AI work, categorizing projects into a spectrum of four modes— consult , include , collaborate , and own —and three dimensions— participation goal , scope , and form .

The authors then use the framework to analyze 80 research articles, revealing that existing participatory AI projects focus more on community consultation and less on collaboration and stakeholder ownership (Delgado et al. , 2023 ) . Prior work has also highlighted some limitations of CCA (Pine et al.

, 2020 ; Erete et al. , 2023 ; Harrington et al. , 2019b ) .

For example, nonconsensual, short-term, and non-contextual-specific participation would lead to the exploitation of impacted communities through “participation washing” (Sloane et al. , 2020 ) . Sloane et al.

argue for recognition of participation as work that is valuable for machine learning development, including data work and context-specific consultation (Sloane et al. , 2020 ) . In AI4SG specifically, Lin et al.

document the contributions that community organization staff make to AI4SG projects, including data work and participating in regular feedback meetings (Lin et al. , 2024 ) . However, this labor is often invisible or overlooked by downstream AI developers and the public (Crain et al.

, 2016 ; Gray and Suri, 2019 ; Crawford, 2021 ; Sambasivan et al. , 2021 ) . A recent body of work has started to quantify invisible labor in crowd work (e.g., managing payments and hyper-vigilance) (Toxtli et al.

, 2021 ) and online volunteer moderators (e.g., approving content and managing users) (Li et al. , 2022 ) , revealing that such unseen tasks make up a substantial portion of the overall work. 2.

3. Emerging Understanding of “Tech for Good” Decision-Makers and AI4SG Investments An emerging body of work has begun to understand the perspectives, incentives, and challenges faced by grantmakers in “tech for good. ” For example, through interviews with eight decision-makers in development projects, Saha et al.

( 2022a ) found that community participation is rarely a part of the grant decision-making process, and it is not a requirement for applicants either. Some barriers to incorporating community voices include limited access to local organizations and communities, and overconfidence in their knowledge about the local contexts (Saha et al. , 2022a ) .

Relatedly, Gardner et al. ( 2022 ) argue that a lack of internal expertise in AI and AI ethics creates barriers for funders to incorporate Trustworthy AI principles.

Studies on other influential decision-makers, such as policymakers in government agencies, point out factors influencing their decision-making process around innovative technologies, including learning from neighboring agencies’ experiences and reliance on national surveys (which omit marginalized communities’ views of their needs) (Saha et al. , 2022b ) .

Focusing on private tech companies specifically, some recent works have started to critically examine investment in AI4SG through case study analysis (Nost and Colven, 2022 ; Lukacz, 2024 ; Henriksen and Richey, 2022 ) .

For example, Nost and Colven ( 2022 ) view Microsoft’s AI for Earth program as a form of greenwashing and “philanthrocapitalism”, where the company uses the initiative to enhance its reputation as a socially responsible corporation, while simultaneously contributing to environmental harm through its broader business operations, including its reliance on fossil fuels for data centers.

Henriksen and Richey ( 2022 ) examine Google’s philanthropic efforts, similarly arguing that they serve dual purposes: advancing social causes while simultaneously expanding Google’s market influence and integrating its technologies into more areas of life.

Comparing “Big Tech” companies with the “Big Tobacco” industry, Abdalla and Abdalla ( 2021 ) examined “Big Tech” funding’s influence on the broader academic landscape (though not specific to AI4SG), shaping events and research questions that “put forward a socially responsible public image” of tech companies. 2. 4.

Document Analysis of Pre-Existing Texts about AI and Funding (beyond AI) Document analysis is a systematic procedure for evaluating printed or electronic documents to “elicit meaning, gain understanding, and develop empirical knowledge” (Bowen, 2009 ) .

Documents are pre-existing textual instruments (e.g., public reports, proposals, meeting notes) that serve as “social facts” (Bowen, 2009 ) and can be qualitatively evaluated (Morgan, 2022 ) . Document analysis can provide “contextual richness” by offering background information and inspiring new research questions (Bowen, 2009 ) .

In addition, because documents are “unaffected” by researchers’ presence and influence (unlike interview data), document analysis provides a way to evaluate stable texts that are not distorted by participant reactivity and thus offer access to information that may be difficult or impossible to gather through interactive methods (Morgan, 2022 ) .

Although pre-existing texts only contain “limited information” and reflect only what their creators (such as funders) choose to make available, they still provide a meaningful part of the larger picture and guide future research directions (Morgan, 2022 ) . HCI scholars have used document analysis to examine pre-existing texts about AI, encompassing AI ethics toolkits (Wong et al. , 2023b ) , AI policy documents (Wong et al.

, 2023a ; Sophie Bennani-Taylor, 2024 ; Bareis and Katzenbach, 2022 ) , AI risk manifesto-style statements (Oldenburg and Papyshev, 2025 ) , patents (Cheon, 2023 ) , and military funding (Widder et al. , 2024 ) .

Pre-existing documents about AI do not simply describe AI as facts; they construct AI imaginaries —“collective visions, beliefs, symbols, and expectations that individuals and communities hold about AI and its potential outcomes” (Zhong et al. , 2025 ) , rooted in Jasanoff and Kim’s “sociotechnical imaginaries” (collective visions of technology’s role in society) (Jasanoff and Kim, 2009 ) .

These documents not only define what AI is, what problems it should solve, who benefits from it, and whose perspectives remain invisible, they also “actively shape the trajectory of AI development, integration, and governance” (Zhong et al. , 2025 ) . For example, Wong et al.

( 2023b ) analyzed a corpus of 27 AI ethics toolkits to “identify the discourses about ethics, the imagined users of the toolkits, and the work practices the toolkits envision and support. ” They found that AI is imagined as a technology that can be made “ethical” through technical adjustments—better models, checklists, or audits (Wong et al. , 2023b ) .

Sophie Bennani-Taylor ( 2024 ) argue that policy documents perform “discursive infrastructuring”: they stabilize certain meanings of AI and legitimize particular sociotechnical arrangements.

Through an analysis of UK’s National AI Strategy document, the author found that AI is constructed as 1) an autonomous, inevitable force, 2) a driver of national progress and economic modernization, and 3) a domain dominated by technical expertise over social knowledge (Sophie Bennani-Taylor, 2024 ) . Closest to the corpus analyzed in our work, Widder et al.

( 2024 ) analyzed U.S. Department of Defense’s grant solicitations for the use of AI in military applications and revealed a deep entanglement between academic AI research and militarized imaginaries, where recurring rhetorical devices like “one small problem” justify continuous funding increases for military AI.

Similarly, prior research examining funding documents (not limited to AI) shows that funding languages are not neutral; they encode priorities, shape applicant behaviors, and structure downstream research collaborations. For example, Palmero et al.

( 2025 ) conducted a textual network analysis of Horizon Europe funding calls, treating these texts as “cognitive artefacts of policy design” that reveal how high-level policies are translated into operational agendas. The authors point out that funding document “predefines applicants’ behaviour, shapes their own perceptions of eligibility, and influences evaluative processes” (Palmero et al. , 2025 ) .

Existing work in innovation studies likewise demonstrates the downdream influence of funding documents: Peng et al. ( 2024 ) find that the proportion of ”promotional language” (such as “innovative” and “unique”) in grant proposals is statistically associated with a proposal’s likelihood of being funded, while Smith et al. ( 2023 ) show “how thematic directives rooted in funding calls influenced collaboration across research.

” Building on this body of work, we contribute a first document analysis of AI4SG funding documents to understand how funding language discursively constructs AI imaginaries (especially AI’s role in social good domains) and characterizes the conditions under which its positive impact might be achieved.

To examine how funding documents frame AI4SG, we collected and analyzed funding calls and grant announcements that are publicly accessible to AI4SG project applicants and the wider public audience. In this section, we detail our approach for identifying and analyzing AI4SG funding documents (Subsections 3. 1 and 3.

2 ) as well as our team’s positionality (Subsections 3. 3 ). We discuss limitations to our approach in Subsection 3.

4 . 3. 1.

Qualitative