Artificial Intelligence for Social Good

• AI’s recent progress in basic research, data availability, and computing power creates an opportunity to shape its future development.Source: Overview · p. 5
• AI can deliver societal benefits while ethical, safety, and legal concerns remain important considerations for research and policy.Source: Overview · p. 5
• The 2016 workshop discussed successful deployments and potential AI uses in urban computing, health, environmental sustainability, and public welfare.Source: Overview · p. 5

Technology Enablers

• AI enables urban transportation systems to progress from descriptive analytics to predictive models and prescriptive optimization, game theory, and mechanism design.Large-scale instrumentation captures travel patterns, while privacy remains an important consideration.Source: Technology Enablers · p. 6
• Pilot systems demonstrate real-time traffic-light optimization and on-demand, multi-modal transportation addressing first- and last-mile gaps.Source: Technology Enablers · p. 6

Technical Challenges

• Transportation pilots must scale to multi-modal city models and account for behavioral changes that affect system efficacy.Source: Technical Challenges · p. 6

Evidence-based Policy Making

• Urban AI can support policy experiments that measure mobility responses across population segments before transportation policies are fully implemented.Source: Evidence-based Policy Making · p. 6
• 25% lower travel times, over 40% lower wait times, and 21% lower emissions were achieved by Surtrac in an initial Pittsburgh deployment.Surtrac senses traffic and allocates green time in real time; the deployment expanded to 50 interconnected intersections, with plans for 150 more.Source: Case Study: Real-Time, Adaptive… · p. 7, Figure 1 · p. 7

Sustainability

• AI sustainability work organizes ecosystem management around data, modeling, decision making, and monitoring using high-quality data and science.Source: Sustainability · pp. 7–8, Short-term Applications and Challenges · p. 8

Data

• Sensor networks, citizen science, imagery, and trained surveys collect ecosystem data that AI can optimize, classify, and use to identify individuals and species.Examples include TAHMO weather stations, IBEIS animal identification, and computer vision for freshwater macroinvertebrates.Source: Data · pp. 8–9

Models

• Species models estimate populations and distributions, characterize habitat and climate effects, and predict migration, dispersal, reproduction, and mortality.Source: Models · p. 9

Policy Optimization

• Conservation policies combine ecological and economic models with AI planning and optimization under uncertainty, including habitat purchasing, temporary field rentals, and coastal adaptation.The examples use network cascade analysis, Monte Carlo optimization, multi-scale machine learning, and partially observable Markov decision processes.Source: Policy Optimization · pp. 9–10

Opportunities and Challenges in the Medium Term

• Broader ecosystem management requires multi-scale data integration, models that represent changing climate and species behavior, reliable sensors, bias-aware measurement models, local computation, and sustainable financing.Source: Opportunities and Challenges in the… · pp. 10–11

Long Term Prospects

• Long-term sustainability requires precautionary monitoring, iterative model extension and policy reoptimization, and risk-sensitive methods that control downside risks and address unknown unknowns.Source: Long Term Prospects · pp. 11–12

Success Stories

• nEmesis uses social-media self-reports to identify restaurants associated with foodborne illness and significantly improved inspection effectiveness in Las Vegas.Source: Success Stories · p. 12
• SC2i deployed clinical decision-support tools for invasive fungal infection treatment and evaluated massive-transfusion prediction to reduce unnecessary transfusions.Source: Success Stories · p. 12
• Real-time tools integrating diverse electronic-health-record data identify severe sepsis risk earlier than standard care, enabling timely treatment associated with over 50% lower mortality.Source: Case Study: Making ‘Meaningful Use’… · pp. 12–13

• Patient variation makes optimal chronic-disease therapy difficult to choose.Source: 1) Targeted therapy decisions · p. 13
• Longitudinal clinical measurements and health records can support targeted therapy decisions.The approach uses large-scale observational data to improve decision-making.Source: 1) Targeted therapy decisions · p. 13
• Computational subtyping identifies patients who manifest disease similarly, refining disease definitions.These subtypes can support individualized estimates within a probabilistic framework.Source: 1) Targeted therapy decisions · p. 13
• Individualized disease-course estimates can support better disease management, wellness, and diagnosis.Source: 1) Targeted therapy decisions · p. 13

2) New sensors, new healthcare delivery

• Low-cost sensors, social data, telemedicine, and natural-language processing can extend healthcare delivery and synthesize medical knowledge.Source: 2) New sensors, new healthcare delivery · p. 13

Near Term Enablers

• Data sharing requires incentives, common models, cloud platforms, and shared test beds that lower barriers to health research.Non-U.S. data can also accelerate model development and testing.Source: Near Term Enablers · pp. 13–14

Longer Term Opportunity

• Personalized health integrates clinical, genomic, socioeconomic, demographic, environmental, and social data to support longitudinal treatment plans.The stated opportunity is to prevent hospitalization and address broader life circumstances, not only specific diseases.Source: Longer Term Opportunity · p. 14
• Collaborative models should let decision makers test counterfactual treatment or diagnostic scenarios while incorporating messy data and uncertainty.Source: Longer Term Opportunity · p. 14
• Observational health data can encode unrelated phenomena and care-process effects, so models must address bias and transportability.Uncorrected bias may cause harm as provider practice patterns change.Source: Longer Term Opportunity · p. 14

Barriers to Progress

• AI-health progress requires cross-disciplinary training, task-aware privacy methods, portable data policies, and secure integrated records.Performance must be assessed for accountability and reliability in changing living systems.Source: Barriers to Progress · pp. 14–15

Grand Challenge

• The grand challenge is a learning healthcare system that observes data, models health, supports interventions, captures outcomes, and adapts from feedback.Source: Grand Challenge · p. 15

Public Welfare

• AI for public welfare aims to enhance quality of life and increase the equity, effectiveness, and efficiency of public services.Source: Public Welfare · p. 15

Social Issues AI can help address

• Public-welfare applications include predicting justice-system interactions, targeting anti-blight investments, forecasting workforce skills, improving emergency dispatch, preventing police incidents, and supporting student graduation.Source: Social Issues AI can help address · p. 15

Success Stories and Case Studies

• Predictive systems are being used to target lead inspections before child exposure and to identify homes and neighborhoods likely to develop blight before deterioration occurs.Source: Success Stories and Case Studies · p. 16, Case Study: Economic Development:… · p. 17
• A police-risk system identified 10-20% more officers who later had adverse incidents while reducing false positives by 50-60% relative to the existing system.Source: Success Stories and Case Studies · p. 16

Gaps and Barriers

• Key barriers include insufficient experienced collaborators, too few visible projects, missing platforms for integrating sensitive public datasets, long evaluation horizons, and legal or regulatory hurdles.Source: Gaps and Barriers · p. 17

Fundamental Issues

• Public-welfare AI must account for privacy, transparency, traceability, bias, uncertainty, individual preferences, social dynamics, and the organizational capacity to respond locally.Source: Fundamental Issues · p. 18

Opportunities

• Research opportunities include better data collection and federation, behavioral prediction, policy evaluation, intervention support, and models robust to bias, missing data, heterogeneity, and changing conditions.Source: Fundamental Issues · p. 18, Opportunities · p. 18

Research Challenges

• Because public-good systems involve interacting domains and hidden outcomes such as engagement, they require human-in-the-loop methods and integration across multiple AI systems.Source: Research Challenges · p. 19, A Grand Challenge · p. 19

• AI research is already being applied to urban computing, sustainability, health, and public welfare, but this social-good trend needs encouragement.Source: Cross Cutting Issues · pp. 19–20

Use inspired research

• Use-inspired research grounds AI questions in actual applications, including settings where unreliable technology access creates fundamentally different data and research needs.Source: Cross Cutting Issues · p. 20

Interdisciplinary teams and new styles of research

• Interdisciplinary AI for Social Good teams require members rooted firmly in the relevant domain discipline.Source: Cross Cutting Issues · p. 20

Cross Cutting Issues

• Evaluating social-good AI requires measuring real-world impact rather than relying only on modeling efficiency or simulation improvements.Impact measurement takes time and may not fit a six-month AI conference cycle, motivating new publication venues and incentives.Source: Cross Cutting Issues · p. 20

Interpretability, transparency, accountability

• Interpretability and transparency are needed because opaque models may be rejected or may rely on biased, racist, sexist, or illegal inferences.Source: Cross Cutting Issues · p. 20

Human-AI boundary

• Human-AI applications must define an appropriate boundary, especially when decision aids operate in domains involving vulnerable populations.Respect for autonomy implies that humans should remain in charge when they have superior control, insight, or knowledge.Source: Cross Cutting Issues · p. 20

Cross Cutting Issues

• With appropriate forethought and incentives, AI can enhance quality of life at personal, national, and global levels across applications that cannot all be anticipated today.The report encourages connections between basic AI research and state, local, and nonprofit organizations.Source: Cross Cutting Issues · p. 20
• Many social-good problems and solutions are local, creating opportunities for researchers to engage communities and explain AI’s limitations and potential benefits.Source: Cross Cutting Issues · p. 21

Artificial Intelligence (AI) for Social Good Workshop

• The workshop discussed successful AI deployments and potential uses of AI for social good, with over 300 participants, including 87 females and 75 individuals from industry.An additional 3,500 viewers watched the livestream.Source: Artificial Intelligence (AI) for… · p. 25
• The workshop section provides links to its website, videos, and presentation slides.Source: Artificial Intelligence (AI) for… · p. 25

Post-workshop AI Roundtable Discussion

• The post-workshop AI roundtable included participants from universities, government agencies, industry, and research organizations.The listed participants represented organizations including OSTP, AAAI, CCC, NSF, DoD, Microsoft Research, and multiple universities.Source: Post-workshop AI Roundtable Discussion · p. 26