Artificial intelligence is increasingly functioning as core societal infrastructure, underpinning scientific research, public services, education, healthcare, finance, and digital platforms at global scale. As AI systems grow in size, complexity, and reach, concerns about their sustainability have become increasingly urgent, spanning environmental impact, long-term maintainability, accountability, and governance. Yet current approaches to AI sustainability remain fragmented, often addressing efficiency, fairness, safety, and oversight in isolation rather than treating AI systems as end-to-end, evolving socio-technical artifacts.

This Shonan Meeting convenes an interdisciplinary group of researchers and practitioners to rethink sustainable AI as a lifecycle-wide, systems-level challenge. Moving beyond narrow notions of “green AI” or isolated technical optimizations, the seminar will explore how sustainability considerations can be embedded throughout the AI lifecycle, from data collection and model development to deployment, adaptation, reuse, and retirement. Discussions will be guided by questions around lifecycle thinking, trade-offs and co-optimization, measurement and visibility, reuse versus retraining, and lessons from other long-lived digital infrastructures.

Grounded in the Shonan tradition of interactive, reflective exchange, and informed by Japanese perspectives on long-term stewardship and responsible reuse, the meeting aims to articulate a shared interdisciplinary research agenda for sustainable AI systems. Expected outcomes include a conceptual map of key challenges, a community-oriented manifesto, and seeded collaborations that advance sustainable AI as a foundational concern for informatics and AI research.

Artificial intelligence is increasingly functioning as core societal infrastructure, underpinning scientific research, public services, education, healthcare, finance, and digital platforms at global scale. As AI systems grow in size, complexity, and reach, concerns about their sustainability have become increasingly urgent. Large-scale models and data-intensive pipelines place unprecedented demands on compute, energy, and hardware resources, while also raising questions about the long-term maintainability, accountability, and governance of systems that are expected to persist and evolve over time.

Yet current approaches to AI sustainability remain fragmented. Environmental impacts are often addressed separately from questions of fairness, safety, robustness, or transparency. Technical optimizations tend to be local, focusing on individual models, training runs, or hardware components, rather than treating AI systems as end-to-end, evolving socio-technical artifacts. At the same time, policy and governance efforts frequently lag behind technical practice and lack shared abstractions and evaluative frameworks that would support meaningful oversight across stakeholders and institutions.

This Shonan Meeting proposes to convene an interdisciplinary group of researchers and practitioners to rethink AI sustainability as a lifecycle-wide, systems-level challenge. Rather than focusing narrowly on “green AI” or isolated efficiency gains, the seminar will explore how sustainability considerations can be embedded throughout the AI lifecycle, from data collection and model development to deployment, adaptation, reuse, and eventual retirement, while remaining aligned with broader societal goals and responsibilities.

The Shonan setting provides a particularly fitting context for this discussion. Japanese traditions of long-term stewardship, careful maintenance of shared infrastructure, and responsible reuse foreground sustainability as a question of durability and care over time, rather than short-term optimization. These perspectives resonate strongly with the need to reconceptualize AI systems not as disposable technical artifacts, but as enduring infrastructure that must be designed, evaluated, and governed with their full lifecycle in view.

The objective of the seminar is to develop a shared, systems-level understanding of AI sustainability that spans technical, organizational, and governance perspectives. Rather than advancing a specific solution or framework, the meeting is organized around a set of guiding questions that frame sustainability as a lifecycle-wide concern for AI systems:

• Lifecycle thinking. How should sustainability considerations evolve as data, models, and pipelines are reused, adapted, or repurposed over time, and what responsibilities persist as systems change?

• Trade-offs and co-optimization. How can performance, efficiency, fairness, robustness, transparency, and environmental impact be reasoned about jointly rather than optimized in isolation?

• Measurement and visibility. What metrics, documentation practices, or artifacts are needed to make sustainability properties legible to developers, operators, auditors, and policymakers?

• Reuse versus retraining. Under what conditions is reuse more sustainable than rebuilding from scratch, and how should such decisions be evaluated across technical and institutional dimensions?

• From tools to infrastructure. What lessons can AI research draw from other long-lived digital infrastructures that have grappled with scale, efficiency, and governance?

Together, these questions aim to surface gaps in current practice and help articulate a shared, interdisciplinary research agenda for sustainable AI systems.

The seminar will follow the Shonan meeting style, emphasizing interaction, openness, and collective sense-making over formal presentations. The format will combine short framing talks, moderated plenary discussions, focused breakout groups, and regular synthesis sessions.

The seminar will bring together approximately 25–35 participants from diverse backgrounds and career stages, including senior researchers, early-career scholars, and selected practitioners. Participants will be chosen to ensure diversity across disciplines, geographies, and perspectives, with particular attention to bridging communities that do not typically meet.

Rather than producing a single technical artifact, the seminar aims for outcomes that reflect its exploratory and integrative nature, including:

• A shared conceptual map of AI sustainability challenges and research opportunities
• A community-oriented manifesto summarizing key insights and open questions
• Seeded collaborations that can mature into workshops and joint projects

Ultimately, the goal of the meeting is to help reposition AI sustainability as a foundational concern for informatics and AI research, one that demands lifecycle awareness, interdisciplinary collaboration, and sustained attention as AI systems continue to scale and embed themselves in society.