Generative AI for Social Impact
Lingkai Kong, Cheol Woo Kim, Davin Choo, Milind Tambe · 2026
Read Generative AI for Social Impact by Lingkai Kong, Cheol Woo Kim, Davin Choo, Milind Tambe (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-04238.txt.
Introduction
Lingkai Kong, Cheol Woo Kim, Davin Choo, and Milind Tambe ask how generative AI can help social impact work scale. Projects in health, conservation, and security have scored real wins, yet growth stalls at deployment. The authors trace the stall to three linked gaps that hold projects back together.
Data are scarce or unreliable, planning must handle combinatorial choices in shifting settings, and expert knowledge is hard to fold in as conditions change. Language model agents can turn plain-language guidance into clear goals and limits for planners, easing the alignment gap. Diffusion models can make realistic synthetic data and support uncertainty-aware planning, easing the data and policy gaps. Together, the four authors argue, these tools could yield systems that are scalable, adaptable, and aligned with human judgement in high-stakes resource choices.
Catalog record
| CREATOR | Lingkai Kong, Cheol Woo Kim, Davin Choo, Milind Tambe |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 18 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-04238
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Generative AI for Social Impact Lingkai Kong, Cheol Woo Kim, Davin Choo, Milind Tambe Abstract Abstract AI for Social Impact (AI4SI) has achieved compelling results in public health, conservation, and security, yet scaling these successes remains difficult due to a persistent deployment bottleneck. We characterize this bottleneck through three coupled gaps: observational scarcity resulting from limited or unreliable data; policy synthesis challenges involving combinatorial decisions and nonstationarity; and the friction of human–AI alignment when incorporating tacit expert knowledge and dynamic constraints. We argue that Generative AI offers a unified pathway to bridge these gaps. LLM agents assist in human-AI alignment by translating natural-language guidance into executable objectives and constraints for downstream planners, while diffusion models generate realistic synthetic data and support uncertainty-aware modeling to improve policy robustness and transfer across deployments. Together, these tools enable scalable, adaptable, and human-aligned AI systems for resource optimization in high-stakes settings. 1 Introduction AI for Social Impact (AI4SI) is a distinct and rapidly growing subdiscipline of artificial intelligence research dedicated to achieving measurable positive societal impact, particularly for vulnerable and under-resourced populations [ 20 , 17 ] . This field leverages the predictive, optimization, and decision-making capabilities of AI to address complex global challenges in areas like public health [ 27 ] , environmental sustainability [ 5 ] , disaster response [ 15 ] , and poverty alleviation [ 7 ] . The fundamental objective of AI4SI is often the design and deployment of systems that can optimize limited resources to achieve maximal human benefit. Translating these objectives into practice requires addressing the central challenge of resource optimization under uncertainty. Over the past two decades, our work has focused on developing algorithmic frameworks to meet this challenge across diverse high-stakes domains, including public health (maternal and child care [ 23 ] , HIV prevention [ 19 ] ), wildlife conservation [ 28 ] , and public safety [ 18 ] . These deployments provide empirical evidence that AI-driven optimization can yield significant, verifiable improvements in real-world settings. For example, in large-scale mobile health programs in India, these strategies contributed to a 30% reduction in program drop-out rates [ 23 ] , while in national parks, they led to a five-fold increase in the detection of illegal snares [ 26 ] . Despite these notable successes in algorithmic design and localized deployment, a persistent and critical barrier stands in the way of achieving widespread, scalable AI4SI adoption: the deployment bottleneck . This bottleneck manifests across three distinct but interconnected gaps: the observational scarcity gap (data acquisition), the policy synthesis gap (modeling and learning), and the human-AI alignment gap (real-world implementation). This paper asserts that the next generation of AI4SI scalability will be unlocked by the strategic application of Generative AI to overcome each facet of this bottleneck. The rest of this paper is organized as follows. Section 2 reviews representative AI4SI deployments from our group and highlights how the deployment bottleneck arises in practice. Section 3 discusses how generative AI, via LLM agents and diffusion models, can bridge the three gaps and enable scalable, human-aligned deployment. Section 4 concludes with open directions toward trustworthy AI4SI at scale. 2 Deployment Bottlenecks 2.1 The Observational Scarcity Gap (Data) A key challenge in many AI4SI domains is the observational scarcity, which makes it difficult to have the required data for learning and inference. We illustrate this challenge via our continued effort on network-based HIV prevention. YEH in Los Angeles: Hidden Social Networks. Our team conducted one of the first large-scale applications of social network algorithms for public health, focusing on spreading HIV prevention information among Youth Experiencing Homelessness (YEH) in Los Angeles starting in 2013 [ 19 ] . This pioneering field study demonstrated the use of AI-guided interventions to achieve significant reductions in HIV risk behaviors among 750 YEH compared to traditional methods. However, this work provided a clear, early example of the observational scarcity gap . The optimal intervention strategy depended entirely on the hidden, unobserved social network structure linking the youth. Because this network was not known in advance, our influence maximization algorithm had to rely on a sampled fraction of the social network [ 25 , 8 ] . South Africa: Partially Observed Contact Networks. In continued work on HIV prevention in South Africa, our team confronted the observational scarcity gap in an even more direct form. Although network-based testing is endorsed by the WHO as a key strategy [ 16 ] , the underlying transmission/contact network is rarely observed in full in real deployments. Instead, teams act sequentially, observing infection statuses only as individuals are tested. Our Adaptive Frontier Exploration on Graphs (AFEG) framework [ 2 ] formalizes this setting and yields principled decision rules given a contact graph, but the core scarcity remains: AFEG assumes the graph is known upfront, whereas real deployments often face missing or unreliable network structure. 2.2 The Policy Synthesis Gap (Learning and Modeling) Even with sufficient data, learning deployment-ready policies in AI4SI remains challenging. Decisions often depend on high-dimensional state information and combinatorial action spaces, where naive policy parameterizations (e.g., assigning probability to every action) are computationally infeasible at scale. Moreover, policies learned purely from historical data can be brittle under environmental change and domain shift. We refer to this mismatch between deployment requirements and what standard learning pipelines can reliably deliver as the policy synthesis gap : learning policies that are simultaneously near-optimal, scalable, and resilient to nonstationarity. Patrol Planning in Wildlife Conservation. Our work on the Protection Assistant for Wildlife Security (PAWS) [ 28 ] supports ranger patrol planning in large protected areas [ 26 ] . Here, patrol teams repeatedly select a subset of regions to visit each week (or month) based on poaching-risk forecasts. Even with a modest discretization, the action space is inherently combinatorial: selecting 5 regions out of 50 yields \binom{50}{5}=2{,}118{,}760 possible subsets, making explicit enumeration or full-distribution policies impractical. Beyond scale, deployed policies must remain effective as adversaries adapt and risk landscapes shift, so nonstationarity can quickly erode performance. Bridging this gap requires methods that (i) scale to large combinatorial decision spaces and (ii) adapt rapidly to dynamics changes and domain shifts in the field. 2.3 The Human–AI Alignment Gap (Deployment) This is often the most critical and overlooked gap. A mathematically optimal policy has limited value if it is not adopted by field staff or if it conflicts with unwritten social norms, political constraints, or dynamic operational requirements . Domain experts (e.g., frontline health workers and security personnel) hold essential tacit knowledge that is rarely captured in objectives or training data. The alignment gap arises when AI recommendations clash with human judgment or when static models fail to keep pace with emerging needs (e.g., major events or shifting priorities). Closing this gap requires mechanisms that incorporate evolving expert guidance directly into the policy-generation loop. Public Safety and Operational Requirements. Our work pioneered computational game theory for operational security, beginning with the ARMOR deployment at LAX in 2007 [ 18 ] . Despite its success (e.g., adoption by the US Coast Guard and US Federal Air Marshals , and more than \$100 million in savings ), ARMOR exposed a key alignment challenge: agencies routinely faced dynamic, hard-to-quantify operational requirements (e.g.,