Proceedings of a Workshop—in Brief

Convened June 4, 2025

Artificial Intelligence for Sustainability: Maximizing Benefits for the United States
Proceedings of a Workshop—in Brief

As the world faces unprecedented sustainability challenges, including biodiversity loss, resource depletion, and ecosystem degradation, there is a critical need for innovative, scalable, and data-informed solutions. Artificial intelligence (AI) has emerged as a transformative technology with the potential to accelerate progress toward sustainability goals by enhancing efficiency, optimizing resource use, improving environmental monitoring, and enabling data-informed decision-making. While AI presents opportunities, its deployment also introduces complex sustainability challenges. Training large AI models requires significant computing power, leading to concerns about the carbon and water footprints of AI itself. Further consideration of AI’s social implications is important to maximize its sustainability benefits.

To examine AI opportunities, challenges, and pathways through a lens of sustainability, the National Academies of Science, Engineering, and Medicine’s Roundtable on Science and Technology for Sustainability (STS Roundtable), in collaboration with the Board on Mathematical Sciences and Analytics (BMSA) and the Board on Human-Systems Integration (BOHSI) convened a hybrid workshop on June 4, 2025.¹ The workshop examined how AI can be leveraged to maximize benefits for the United States at the intersection of nature, people, and this critical technology. This Proceedings of a Workshop—in Brief provides a high-level summary of key discussions held during the workshop.

WELCOME AND GOALS OF THE WORKSHOP

The day began by introducing the work of the STS Roundtable and objectives of the workshop. Franklin Carrero-Martínez (National Academies) highlighted the history of the Roundtable since its establishment in 2002 and thanked the George and Cynthia Mitchell Endowment for Sustainability Science for its support. Vaughan Turekian (National Academies) underscored how AI has become top-of-mind across domains. He noted AI’s tremendous opportunities, but also its impacts on energy use, water, and other resources.

Roundtable co-chair Cherry Murray (University of Arizona) acknowledged the sponsorship of the Mitchell Endowment to allow for discussion of cutting-edge issues such as those in the current workshop. Roundtable co-chair Klaus Tilmes (policy advisor and development consultant) said the interrelationship between AI and sustainability requires a look at emerging trends related to international competitiveness. He noted the National Artificial Intelligence Research and Development Stra-

__________________

tegic Plan recommends investments in pre-commercial research, which calls for policy making around the nexus between sustainability and AI.²

Comments from representatives of two National Academies boards reflected the workshop’s cross-divisional collaboration. Brittany Segundo said the growing interest in the mathematical and statistical foundations of AI and its applications as well as an increased awareness of its environmental impact have led BMSA and other boards across the National Academies to establish, among other efforts, the Roundtable on Artificial Intelligence and Climate Change. Emanuel Robinson explained BOHSI’s focus on the intersection of humans, technology, and the built, organizational, and natural environments.

Workshop chair Shefali Mehta (Open Rivers Consulting Associates) introduced that the Science and Technology for Sustainability Program frames sustainability at the intersection of nature, people, and technology. Mehta reviewed the workshop objectives in this context: (1) review AI applications that support sustainability goals; (2) examine challenges, risks, and considerations of AI; and (3) discuss opportunities for cross-sector collaboration, interdisciplinary research, and international cooperation to maximize AI’s benefits for sustainability efforts and our society overall.

FRAMING REMARKS: THE INTERRELATED BENEFITS OF AI AND SUSTAINABILITY

Daron Acemoglu (Massachusetts Institute of Technology) shared his perspective on AI’s effect on the labor market and the macroeconomy, particularly how AI can be “pro-worker.”

Discussion

A participant asked about the impact of AI on workers in different parts of the world and occupations. Acemoglu responded that blue-collar workers in the West have the most experience with automation. Although many lost jobs, the current workforce seems comfortable with robot-performed tasks. This suggests some costs are transitional, borne by one cohort of workers but benefiting the next cohort, he said. While AI is less used in developing countries, workers will experience indirect effects through changes in the global division of labor. Large-scale joblessness would be hugely destabilizing in any country, he noted, but a slower timeframe allows for planning. In some situations in the developing world, AI could extend skills and services. AI requires international

__________________

coordination, he said. The push toward artificial general intelligence (AGI) could result in less cooperation.

Asked about levers to drive pro-worker AI, Acemoglu posited that if 60 percent of AI researchers and companies realized such a model is both socially beneficial and could make money, the problem would be solved. Reaching that 60-percent threshold, especially when locked into an AGI-driven race, starts with changing the conversation and outlining how to achieve pro-worker AI with government incentives and regulations.

SUPPLY CHAIN FOR AI: THE INPUTS TO AI

The first panel, moderated by Roundtable member Anne Roby (Linde PPC, retired) considered the flow of resources, components, and data to develop and deploy AI technologies.

Technologies, Talent, and Ways of Working

Andrés Naranjo (OrangeSpark Industrial AI; Brydon Group) offered four points based on his decades of AI experience. First, all the attention on large AI labs ignores what other entities can do, such as smaller enterprises and academia. Second, AI comprises a “salad bowl” of technologies. Most common and well known is machine learning, followed by computer vision. Generative AI, which emerged in 2022, has three models: large-language models (LLMs), multimodal, and reasoning.

Third, a shortage of AI talent and ways of working exists. Needed inputs include talent, high-quality data, models, computing power, data center capacity, and energy. High-quality data requires experts to give feedback and models for workbenches. Refuting the belief that software engineering jobs are crashing, he said they are transforming to AI jobs. These jobs include data engineers, data scientists, ML and AI engineers, and others. Moving along the continuum from machine learning to agentic AI, talent becomes more scarce. However, it is possible to upskill, including with free or low-cost online courses. In terms of ways of working, waterfall and agile software development methods do not work for AI, which needs a “sandbox” to experiment and manage risks before release. Data costs money, but workbenches and models are available at low cost. Finally, machine learning and computer vision can already power 85 percent of use cases to advance the Sustainable Development Goals (SDGs), he said. Examples include optimizing grids, managing traffic, and more.

Developing a Proactive Strategy

Erica Fuchs (Carnegie Mellon University) noted in the 1960s, Andrew Marshall looked beyond prevailing methodologies of systems analysis to develop net assessment, a methodology to illuminate emerging problems and opportunities with time for leaders to make decisions.³ Different tools and political contexts have ensued, but the country is repeatedly caught off guard to react to crises. “It’s not that policy makers aren’t trying to address these issues,” she said. “They lack the frameworks, data, and analytic tools to achieve their desired national objectives.”

She testified in Congress about the need for an Analytic Intelligence for Critical Technology Strategy (ACTS) to help the United States identify, invest in, and secure technologies critical to national security. It should be forward thinking; leverage integrated, interdisciplinary teams; focus on the intersection of departmental missions; serve as a neutral third party; and operate in a highly flexible, distributed mode to guide science and technology investment and garner talent.

After her testimony, she was tasked to lead a National Network for Critical Technology Assessment. It brought together top analytic talent from thirteen institutions in a one-year sprint to produce a pilot to demonstrate “the whole is stronger than the sum of its parts.”⁴ The AI team found AI has the potential to increase scientific discovery, productivity, output, and employment, but the benefits are uneven. For example, while increased AI job postings correlate with increased non-AI job postings, opportunities are not evenly distributed by geographic region, company size, or scientific discipline.

Looking at AI in scientific disciplines, experts in AI and different scientific domains met at Carnegie Mellon to discuss opportunities and chokepoints. They identified scientific uses of AI such as harvesting the literature, serving as a quantitively capable research assistant, predicting properties of interest, and developing founda-

__________________

tional models to simulate complex systems. Two limitations to seizing these opportunities are (1) access to publicly available, machine-readable data and (2) a person or a pair of collaborators with both domain expertise and expertise in AI tools. Access to expensive compute and energy resources, as well as capital, are constraints, but a lot of “low-hanging fruit” is available now. Each scientific field has different needs, which can be learned through surveys, workshops, and other research. She also called for showcasing pockets of success.

Shifting from science to manufacturing, Fuchs touched on AI’s contributions to competitiveness, security, and societal well-being. She noted if access to microprocessor components were cut off, U.S. infrastructure could operate in the short term but could not expand. Many of the chips used for technologies today could be produced in the United States, but it would take 18 to 36 months to ramp up production. Getting ahead of problems strategically, especially those with longer time lags, is important, she concluded.

Creating AI Impact in Manufacturing

Larry Megan (Baldwin Richardson Foods Co.) continued the discussion about AI in manufacturing based on his experience at the intersection of manufacturing and digitalization in companies of different sizes. Megan reflected on AI’s potential to increase resource efficiency, such as conserving water and reducing waste, which benefits both sustainability and profitability. He pointed to three paths of AI value: embedding it in end-products, contributing to personal productivity, and customizing applications to meet customer needs and get to market faster. The AI supply chain requires leadership and vision, a strong technical foundation, good data, and talent.

Megan observed that small and medium enterprises (SMEs) have been left behind in the technology race. Most government initiatives do not reach them. Most heads of companies with a few dozen or hundreds of employees lack exposure to AI possibilities and R&D funding. They often have aging infrastructure and technical debt, and little experience with modern data platforms. They also have more difficulty than larger companies in accessing, hiring, and retaining talent across the AI value chain. He urged bringing AI expertise to SMEs by leveraging existing Manufacturing Extension Partnerships, incentivizing partnerships with vendors and academia, and encouraging investment through targeted research funding and tax incentives.⁵ SME-targeted solutions must be simple and geared to developing local talent. It is an issue of competitiveness and national security, he said.

Discussion

A participant reflected on the growing demand for physical AI (e.g., robots) and synthetic data (e.g., simulated data that mirrors the real world). Fuchs urged distinguishing between what is ready today and what to plan for the future. Megan said from a manufacturing perspective, lack of data is not the problem. The challenge is to contextualize and use data efficiently. Naranjo said robots are expert systems built on if/then data; there is a long way to develop robots built on non-deterministic systems, although it will happen. Some areas have reached exhaustion in terms of the amount of data, and what is needed is reinforcement learning based on expert feedback, he said. Fuchs cautioned that generalizations about AI do not work. There is no single reason for chokepoints nor ways to apply AI.

The tension between technology and realities of the shop floor was explored. A participant also reflected that AI may be shaping career aspirations and the American dream. Megan responded at the end of the day, the country has to “make stuff.” Skilled engineers and other workers are still needed. AI technology can work in partnership with them, but the overall objectives have not changed. Fuchs spoke to the future of work. It is important to be real with people about the implications of various technologies. Overall, there are opportunities, but workers are affected differently, depending on their job and location.

There are both values-based and business arguments to upskill workers, but sometimes it is hard to make the case for resources, a participant observed. Megan said training to develop workers readily exists online, in community colleges, and elsewhere. SMEs need help in the curation of that content for their workers.

__________________

THE IMPACT OF AI ON SUPPLY CHAINS: THE OUTCOMES FROM AI

AI can revolutionize the efficiency and sustainability of supply chains by optimizing processes, reducing waste and energy consumption, and improving transparency, noted roundtable member Meghna Tare (University of Texas Arlington). She moderated a panel to consider how to ensure data quality and transparency, skills needed, and how AI can support supply chains that are efficient, sustainable, and adaptable to future shocks.

AI in the Energy Sector

Feng Qiu (Argonne National Lab) explained how mathematical optimization models underlie decisions in all sectors, with the goal to find the best solution by maximizing an objective function subject to given constraints. In the electricity market, it has become increasingly difficult to use traditional optimization models as grids have become more complex.

Qiu and colleagues are looking at how AI can accelerate, but not replace, existing approaches for better and faster decision-making. It learns patterns from historical data and provides hints for solutions. The approach builds on the benefits of mathematical optimization; leverages decades of research with recent advances; and would have less difficulty to be understood/accepted by the power grid community. They have developed an open-source MIPLearn framework⁶ to help the power sector adapt AI for optimization at scale.

Qiu also discussed AI for more informed decision-making in grid management. Grids comprise a massive amount of assets, including transformers, transmission and distribution lines, circuit breakers, and more. Assets need to be replaced and restocked as they reach the limits of their life spans. Early maintenance wastes machine life, but late maintenance risks outages. In some cases, restocking takes three to five years. AI-driven Prognostics for Grid Asset Operations and Maintenance show the probability of a failure in order to make decisions about when to perform maintenance or replacement. In summary, in energy and other engineering fields, AI has potential to achieve more sustainable supply chain management through more informed decision making.

AI for Sustainable Food Systems and Nutrition

Elena Naumova (Tufts University) discussed how AI can help build more sustainable and resilient food chains, then offered examples. She began by defining a sustainable food system as one that provides food security and nutrition for all, while preserving the economic, social, and environmental foundations for future generations.⁷ Nutrition lies at the intersection of food systems, well-being, and healthcare. AI tools must be matched with the tasks to achieve sustainable food systems and nutrition.

Echoing previous speakers, she explained AI as a computer-assisted approach to synthesize knowledge, and the AI universe consists of many applications and types. Non-computer experts must be able to understand how they can use AI tools to achieve a measurable impact in their domain. AI can help with many tasks across the food chain, from production through distribution to consumption.

Naumova highlighted several examples. Nestle Research and Development uses AI tools to support employees’ business functions, respond to customers, and create new products. Danone operates an academy to upskill employees in AI, uses digital twins in decision making, and analyzes consumer data for promotions and product offerings with AI. Agreeing with Megan about the difficulty for SMEs to use AI, she highlighted several smaller companies. Brightseed built one of the world’s largest proprietary libraries of plant-derived bioactive compounds, mapped 7 million plant compounds, and identified 40,000 predicted bioactives.⁸ Agilitas developed a proprietary database for more than 100 categories to enable product development, build consumer trust, and innovate. Spoiler Alert offers scalable data infrastructure to manage access and slow-moving inventory to reduce waste. She noted that companies, especially the small ones, often do not share their datasets because their survival depends on the data they compiled.

She suggested a potential use of AI in averting or minimizing expensive and potentially dangerous food recalls. However, she added, food recall surveillance data is incomplete and hard to harmonize, which she termed a Babylonian Tower problem. She also touched on creation of the Global Nutrition and Health Atlas.⁹ It taught

__________________

lessons about how to build an interactive and open-access infrastructure and how to maintain and harmonize datasets. Despite a lot of data collected by various ardencies, some critical data related to health and nutrition are missing. While she supports the use of AI for sustainable food systems, she urged finding ways to involve multi-sectoral partners, as well as building in data integrity, analytical robustness, and methodological soundness.

AI Tools for Sustainability

William Babis (Stockholm Environment Institute [SEI]) has helped build AI-powered research and policy tools at SEI and participates in an internal task force to ensure AI use that enhances rather than compromises the integrity of the organization’s work. Babis shared several tools. They include incorporating AI into the generation, organization, and querying of a renewable energy policies database; a set of processes that was time- and resource-intensive before the use of AI. The AI Policy Reader tool used for this purpose can also—with sufficient methodological instructions and calibration—generate similar databases for other applications intended to summarize policy landscapes. William’s team has also built a Newsfeed Tracker that collects articles from around the world, and can filter, read, and answer questions about them with validation and transparency. SEI also created a tool called Trase,¹⁰ which maps the production of agricultural commodities in countries, regions, and municipalities around the world prone to deforestation. AI has been used on satellite imagery to identify deforested land and locate new soy processing facilities.

Discussion

Asked to consider what is needed to invest to and adopt AI, given the evidence from business cases, Naumova opined companies are cautious. Bigger companies may start with consumer-facing activities. She also called for improved communications between AI and domain experts. Babis described ways to validate databases and screen out misinformation. It is important to pre-define a sufficient level of accuracy and calibrate, he said. For example, for the News Tracker, humans reviewed documents, AI did the same task, then the results were compared. Both humans and machines made some mistakes, but no systemic bias was found.

A participant noted that companies conduct RAM analyses (reliability, availability, maintainability) when considering large capital decisions, but they are time-consuming and expensive. She asked whether AI could help. Another participant asked about AI’s potential in informing Black Swan events (rare, unexpected, but very significant occurrences, such as a multi-country power outage). Qiu said in both cases, AI could run many different scenarios to help in decision making. Naumova noted that the results sometimes suggest investments that decision makers are reluctant to make.

Dialogue: Inputs and Outcomes from AI

A short session encapsulated takeaways from the first two panels. Megan reflected that people sometimes look to AI as a cure-all, but Qiu’s presentation offered ideas for AI to complement, not replace, existing practices. He also noted a theme across both panels about the need for talent. Qiu commented on the possibility that company culture can prevent AI adoption, despite the advantages. Naumova said the presentations underscored the time is right for a discussion of strategic steps to move forward. A deeper understanding of how to communicate science and improve data literacy should precede pushing forward with AI, she said.

Naumova described building a global food and nutrition database. She was part of a team that created metrics of success based on four principles: evidence, efficiency, emphasis, and ethics. Overall, AI has provided adequate but not outstanding responses to complex questions. Thus, she said, she teaches students to develop skills as information-sorters, not just information-seekers, with AI as a viable assistant to the expert. Megan commented AI may not get the newest, best solutions, but it drives consistency. For example, a production scheduling problem given to seven people would result in seven different answers. Technology adapts to a world where there is not a single expert.

Roby commented on the need for culture change to utilize AI. Babis said this is a timely question at SEI, which is discussing how to take advantage of AI while mitigating risk. In the for-profit world, change does not come as fast as one would want, Megan said. People have different levels of technology adoption; in any event, they have to keep their businesses running. Change has to come both from the top, where leadership sees alignment with

__________________

strategic objectives, and the bottom, where early adopters can be proponents beyond the technology team. Naumova commented people do not like to say they are wrong, but “AI helps to see where ignorance and arrogance collide.” It requires thinking about multiple scenarios and experimentation. Change is inevitable, Megan said. Managers must be empathetic to how employees adjust to new technology, but realistic that being competitive requires change.

When dealing with classified work and AI, security is a top priority, a participant noted. Qiu agreed and described ways to advance AI without exposing security, such as releasing responses (e.g., how to respond to an outage) without releasing data. Asked about the main threats to sustainability from AI applications, the most obvious is overall energy consumption, Megan said. A more tactical link, from his perspective, is how to get confidence in answers and use them in the right way. Technology must be combined with domain expertise. Qiu brought up the pressure on the grid from data centers. Roby commented that AI can optimize, but the right metrics must be selected to drive sustainable behavior. The relationship between the complexity of models and accuracy is log-linear, Babis noted as is the relationship between time spent “thinking” about a problem (“infer-encing” is the commonly used industry term) and model accuracy. More model complexity and time spent processing queries both consume more energy. While energy consumption is concerning but not alarming today, the trajectory is alarming, he added.

SOCIETAL IMPACT OF AI ON EMPLOYMENT AND WORKFORCE

Roundtable member José Lobo (Arizona State University) moderated a session on AI’s impact on employment and the workforce. He called attention to a recent National Academies consensus study on the topic.¹¹ Building on Acemoglu’s framing remarks (see above), the panel shared U.S., global, and indigenous perspectives.

U.S. Labor Market Projections

Christine Machovec (Bureau of Labor Statistics [BLS]) discussed the BLS’s employment outlook for 2023 to 2033 and explained how BLS incorporates technology, including AI, into its 10-year employment projections (EP). EP goals are to identify structural changes to the economy and to employment over the 10-year projection period based on a range of qualitative and quantitative data. Assumptions include approximately full employment and that labor productivity and technological progress remain in line with historical patterns. She acknowledged that uncertainty about impacts related to AI remains high. Overall, the two occupational groups with the highest projected growth between 2023 and 2033 are healthcare support (15.2 percent increase) and computer and mathematical occupations (12.9 percent).

Machovec highlighted how AI might affect computer and legal occupations.¹² In computer occupations, strong demand is projected for coding, information security, and database architecture. Office and administrative support employment may decline if technology provides more automation, such as chatbots to deal with routine customer requests. In the legal field, paralegal and legal assistant occupations will grow more slowly than average because AI can sift through massive amounts of information and synthesize findings. AI will have a lesser impact on demand for lawyers, as fewer of their tasks can be automated. Machovec shared employment projections related to AI infrastructure, with large increases for software publishers, infrastructure providers, and others (Figure 1).

Different Future for AI: Indigenous Wisdom, Circular Systems and the Story We Choose

Keolu Fox (University of California San Diego) posed an alternative vision for AI that challenges what he called the conventional “pill-fed” message. He posited that everyone deserves access to AI that reflects their needs, including indigenous communities who have

__________________

prioritized sustainability for thousands of years. In his field of genome sequencing, 90 percent of tools are for health issues for people with western European ancestry. Genetic material is being sequenced and extracted from other places but not to benefit those who provide it.

Data is now the most valuable resource commodity on Earth, he continued. He called for educating community members about this resource, which is theirs. He is part of the Native BioData Consortium, the first genome-sequencing center for and by native peoples.¹³ “Indigenous data sovereignty” has motivated a number of scientific questions. For example, why is there a data center boom in the desert? Will AI expand dependence on natural gas and water consumption? For answers, Fox urged looking to indigenous communities as thought leaders. As one resource, Abundant Intelligences is an indigenous-led research program that conceptualizes, designs, develops, and deploys AI.¹⁴

“We don’t have to accept the narrative that AI must be extractive,” Fox stated. Circularity can be prioritized. Data centers can be designed with the principles of resource flow and reciprocity, self-sufficiency and closed-loop systems, and adaptive management. Older technology could be repurposed to reduce e-waste. The development of specialized small language models (SLMs) would use fewer resources and be better aligned to answer important scientific questions.

Global Perspective

Shamika Sirimanne (United Nations Conference on Trade and Development) posed whether AI will benefit all or deepen existing divides. The answer, she said, is how it is developed, deployed, and governed. Without purposeful interventions, AI will only benefit a few. On the positive side, the cost of AI development is declining. This helps developing countries access and leverage technologies. Yet, she said, AI development is market-driven by a few corporations investing billions in labor-replacing applications. Government subsidies also focus on maximizing efficiency, which she called a code word for labor replacement. When jobs disappear, social and political instability follows, she said.

A message from the United Nations is that governments should realign incentives to encourage AI that serves societal needs. She underscored that AI should empower, not replace, human workers. As examples, in countries with few healthcare providers, AI could broaden access to personalized medicine and improve medical education. In education, where one teacher has 40 to 50 students in a classroom, AI-powered adaptive learning can serve as a teaching assistant. The task for deploying AI for the social good cannot be left to the private sector, she stated. Governments have a role in guiding investments, and public funding is needed. AI potential remains largely untapped in developing economies. Left unaddressed, this will widen global inequality, which will lead to the “lose-lose situation” of instability.

Sirimanne challenged the myth that developing countries cannot use AI. Most do not intend to aim for the frontier of development but can use practical applications. They do not need state-of-the-art facilities to make this work. International cooperation is crucial to unlock the benefits. Partnerships and knowledge sharing can ensure AI serves many constituencies. Scaling up promising pilot programs is critical. Safety and ethical issues must be addressed internationally, she stressed. Without robust frameworks, AI can be and has been used for harmful purposes. The UN can shape global governance and improve transparency. Transformative change can also start from below. Activists can highlight the risks of unregulated AI and apply pressure to put human-centered innovation first. Decisions today will determine if AI becomes an instrument of inclusion or exclusion.

Discussion

Lobo concurred with the importance of education about AI. The leadership at his institution has embraced AI, so he and others are urging that students who are not computer science or math majors need to be involved. Mehta commented on the opportunities for small companies to maximize the benefits discussed. Sirimanne agreed the opportunities are immense but development must be intentional. Technology does not have a brain, it is how humans direct investment and development, she commented.

THE ROLE OF SCIENCE IN ADVANCING AI RESEARCH RELATING TO SUSTAINABILITY

Roundtable member Miguel Román (National Aeronautics and Space Administration) reflected on the foundational role of scientific research. He moderated a panel

__________________

to consider scientific collaborations, interdisciplinary research, and other elements needed to make AI for sustainability not just possible, but actionable. There are no quick fixes, he commented, but science must ground the conversation to chart a path forward.

User-Centered Design and Application

Michael Jacobs (IBM) described the IBM Impact Accelerator, which moves innovation out of the lab and into the field. Five projects are selected annually to participate in a two-year program to implement technology to solve real-world problems. The Accelerator also is a call to action to increase sharing between researchers and those who might apply research. He envisions something akin to a dating app to share ideas and needs in the digital marketplace.

His own experience shows the importance of user-centered design, he continued. He came to IBM from the New York subway system where he managed reconstruction after Hurricane Sandy. Long-time workers knew the system and did not need an AI model to know where flooding occurred, but workforce demographics meant many experienced professionals were nearing retirement and the situation was going to worsen. There was a need for knowledge management and forecasting to bridge gaps with digital tools. A key lesson was how to take innovation from lab to application: as an example, a digital solution would need to be compatible with the thick gloves that workmen use.

Future of AI Research

Eric Horvitz (Microsoft) discussed findings of a presidential panel convened by the Association for the Advancement of Artificial Research.¹⁵ He co-authored the work on AI and Sustainability, one of seventeen areas. He shared its top-level points: AI is transforming industries with the potential to drive sustainability progress to net-zero energy and climate resilience. Challenges include energy and water demands. Proactive efforts are needed to enhance progress.

AI computation represents a very small share of global energy and water consumption, but can strain local grids and resources, he said, which requires investments in local capacity and innovations. The most significant impacts, positive and negative, will be AI deployment.

Resource use is growing, but even in high-growth AI scenarios, electricity use remains a relatively small share of global energy demand. AI is becoming more energy-efficient and renewable-powered infrastructure is increasing. While GPUs (graphics processing units) consume more power than CPUs (central processing units), they are becoming more efficient per unit of computation. SLMs are emerging as energy-efficient alternatives to LLMs. Innovations in cooling are reducing energy use. AI has the potential to enhance efficiency in many sectors, with sustainability applications related to core capabilities, systems optimization, climate and water resilience, wildlife conservation, and more.

To take advantage of potential benefits, Horwitz called for scenario modeling, from minimal use to widespread deployment, and with best-case and worst-case outcomes. This would guide sustainable AI innovation, mitigate unintended consequences, and assess risk.

Two Questions for AI

Owen Gaffney (Nobel Prize Outreach) reflected on two questions about connections between AI and sustainability: how AI will change research and how to ensure that AI will stabilize, not de-stabilize, the planet. In his view, a watershed AI moment occurred in 2024 when the Nobel Prizes in physics and chemistry were awarded for discoveries that used AI. However, most existing models are owned by companies, and incentives about priorities and control of the compute are mostly driven by market forces. He reflected on previous comments about how pro-worker AI can serve the common good. While AI could address sustainability challenges, a lot of its resources are being used elsewhere.

Gaffney, Carrero-Martinez, and Amy Luers were part of a team to develop the Earth Alignment Principle for AI, defined as “the principle of aligning the development, deployment and use of AI to promote planetary stability and stewardship for the benefit of humankind.”¹⁶ It has three pillars: consumption and production within planetary boundaries; power structures; and social cohesion. Operationalizing the principle requires risk assessment, long-term sustainability research, knowledge exchange, and investments in domain-specific models. Solu-

__________________

tions, many of which already exist, need combinatorial complexity, high-quality datasets, straightforward testing, and receptive markets. Immediate scaling of market-ready solutions and investments in blue-sky research are both needed, he said.

Critical Role of Science

Amy Luers (Microsoft) reflected on the good news that AI holds transformational potential and examples are unfolding, but intention and collaboration are needed. In her view, AI has three capabilities to become a sustainability game-changer. First, it can help understand, predict, and optimize complex systems to develop tools in different systems. Second, it can accelerate and discover new materials and processes in record time. Third, it can empower the workforce. Science is needed to unlock this potential, Luers said. Microsoft developed a playbook with five points to accelerate sustainability with AI: (1) Invest in AI intentionally; (2) Develop inclusive, representative digital and data infrastructure; (3) Minimize resource use; (4) Advance AI policies and governance for sustainability; (5) Build capacity not just in technical skills, but to integrate AI. Science is not just a contributor but a compass to navigate risks, unlock opportunities, and ensure that AI serves people and the planet, she said.

Discussion

Román reflected on comments throughout the workshop about how AI can generate usable knowledge. He expressed concern about potential divestments in data streams to feed AI capabilities as well as competition for resources across scientific fields. Horvitz responded that AI methods can help define goals and synthesize knowledge in digital simulations, which can then be integrated with experimental methods such as laboratory-based research and other tools. It is important to take stock of datasets, he urged. Many are interwoven and the lack of one will affect others. In the climate area, for example, work is needed on data provenance to understand where gaps may show up after funding changes or shutdowns. Expertise is distributed and fragile, he also commented.

To put the Earth Alignment principle into practice, Gaffney urged deeper thinking on how to frame and resource AI tools to deliver solutions, as occurred with alpha folding.¹⁷ He reflected on the need to rethink economic models and incentives to build the next generation of technology. To move research out of the lab and into communities, Jacobs called for a whole-of-society approach. Foundational to IBM work is good design, which provides a massive opportunity to consider sustainability from the outset. Best practices are to ensure a project is inclusive for all living things; easy to use and learn for all people; and efficient for users and power consumption. These design practices can only be achieved with diverse and empowered teams, he said.

Luers called for bringing AI into the co-design of research for sustainability. In addition to infrastructure, training and capacity-building to use tools and co-develop models are needed. Román suggested the National Academies could help inform federal agencies to move forward together, for example in harmonizing data plans. Gaffney noted the existing high level of government investment and support for AI, which could be used to ensure that systems are not just market-driven.

To measure success and protect budgets, Jacobs said his portfolio’s topline KPI (key performance indicator) is the number of people directly benefited. Secondary are environmental and technical impacts. Demonstrating that these are IBM values helps in recruiting and retaining talent company-wide. Luers explained she connects science and technology to guide work on Microsoft’s sustainability investments. For Microsoft to achieve its sustainability goals, the world must advance, she said, given the company’s reliance on supply chains and purchased energy. Measures of success include whether the company is meeting its commitments and leveraging further investments.

A participant observed that AI can create problems, for example with misinformation. It is hard to build trust but very easy to destroy it, she noted. Gaffney said this is why social cohesion is a pillar in the Earth Alignment principle. Different business models can increase or prevent misinformation, and hard work is needed so incentives move in the right direction. In terms of the policy environment, Luers posited that policies in the European Union will have ramifications globally. She

__________________

predicted more legislation will eventually play out over time.

SYNTHESIS: FUTURE NEEDS AND OPPORTUNITIES

Roundtable members and participants reflected on the discussions and brought up several areas to explore further.

To Roundtable advisor E. William Colglazier, the biggest short-run challenges to sustainability are conflicts and wars, along with misinformation aggravated by social media. He suggested social scientists look for ways to use AI to understand these problems. He also noted natural disasters along with human disasters where AI might assist social scientists in helping humanity to overcome challenges to sustainable development. Naumova said each field has its own code of ethics, including ethical implications of AI. She urged a multidisciplinary conversation for responsible AI development. Román pointed to the work of Gary Machlis on warfare ecology.¹⁹ He agreed on the need for open discussion across disciplines.

Opportunities and Change

Mehta highlighted several themes she heard throughout the workshop. She noted panelists touched on the immense opportunities of AI and reflected that she heard more agency and voice about AI development than in previous meetings on the subject. For example, she noted the focus on SLM to provide domain-specific, hyper-local, and user-centered value that offers many benefits, while mitigating the environmental impacts. Several presenters commented on AI to enhance, not replace, human workers, which is where pro-worker and inclusivity comes in. She also noted that AI requires immense change, which is difficult for people in any setting. AI requires us to change in all contexts—at the individual level, in our work and with our teams, and in how we engage and connect with society. Many participants suggested the importance of the National Academies continuing discussions on the future of AI for sustainability, as tools and applications are evolving rapidly.

__________________

DISCLAIMER This Proceedings of a Workshop—in Brief was prepared by Franklin Carrero-Martínez, Paula Whitacre, and Emi Kameyama as a factual summary of what occurred at the workshop. The statements made are those of the rapporteurs or individual workshop participants and do not necessarily represent the views of all workshop participants; the planning committee; or the National Academies of Sciences, Engineering, and Medicine.

PLANNING COMMITTEE Shefali Mehta (Chair), Open Rivers Consulting Associates; José Lobo, Arizona State University; Nebojsa Nakicenovic, International Institute for Applied Systems Analysis; and Giulio Quaggiotto, Prime Minister’s Office, United Arab Emirates. The National Academies’ planning committees are solely responsible for organizing the workshop, identifying topics, and choosing speakers. Responsibility for the final content rests entirely with the rapporteurs and the National Academies.

REVIEWERS To ensure that it meets institutional standards for quality and objectivity, this Proceedings of a Workshop—in Brief was reviewed by Michael Jacobs, International Business Machines Corporation; and Elena Naumova, Tufts University. Marilyn Baker, National Academies of Sciences, Engineering, and Medicine, served as the review coordinator.

SPONSORS This activity was supported by the National Academy of Sciences George and Cynthia Mitchell Endowment for Sustainability Science. Any opinions, findings, conclusions, or recommendations expressed in this publication do not necessarily reflect the views of any organization or agency that provided support for the project.

STAFF Franklin Carrero-Martínez, Danielle Etheridge, Emi Kameyama, Emanuel Robinson, and Brittany Segundo.

SUGGESTED CITATION National Academies of Sciences, Engineering, and Medicine. 2025. Artificial Intelligence for Sustainability: Maximizing Benefits for the United States: Proceedings of a Workshop—in Brief. Washington, DC: National Academies Press. https://doi.org/10.17226/29267.

For additional information regarding the workshop, visit https://www.nationalacademies.org/event/45014_06-2025_artificial-intelligence-for-sustainability-maximizing-benefits-for-the-united-states-a-workshop.

Policy and Global Affairs Division of Behavioral and Social Sciences and Education Division on Engineering and Physical Sciences Copyright 2025 by the National Academy of Sciences. All rights reserved.