Latest News

Researchers Eya Ben Chaaben and Janin Koch from INRIA and Université Paris-Saclay presented a forward-looking paper at the SIGCHI 2023 Conference on Human Factors in Computing Systems, held in Hamburg, Germany. Their study, “Addressing Sustainable ML Life-cycles through Human-Centered Design”, advocates for a human-centered rethinking of machine learning (ML) development practices, emphasizing sustainability not only during training but across the entire AI lifecycle.

 This work reflects INRIA’s continued leadership in Sustainable AI research and directly supports initiatives such as the EU Horizon-funded project SustainML, which calls for application-aware and lifecycle-conscious AI design.

From Efficient Models to Sustainable Pipelines

The paper challenges the narrow focus of traditional sustainable ML efforts, such as carbon-aware training, and urges the community to adopt a holistic lifecycle perspective. Through a critical synthesis of literature and lifecycle modeling, the authors examine how ML developers can embed sustainability into every stage of development:

• Project Framing. Developers often rely on personal experience and well-documented libraries instead of articulating explicit goals or trade-offs. This can obscure sustainability concerns from the outset. The authors recommend tools that support structured exploration of constraints like training time, interpretability, and energy use.
• Data Extraction. Data storage, cleaning, and labeling are frequently overlooked sources of energy and ethical cost. The paper encourages developers to collect only relevant data, integrate societal and environmental indicators into quality assessments, and explore alternatives such as human-in-the-loop labeling.
• Model Training. Despite the existence of efficient techniques like transfer learning and low-precision training, awareness among practitioners remains low. The authors advocate for visual tools that can expose these energy-saving opportunities and promote prototyping methods that support early-stage reuse and recycling decisions.
• Deployment & Monitoring. Sustainability is often sacrificed post-deployment due to inadequate feedback loops or retraining from scratch. The paper recommends including user feedback, monitoring for energy inefficiencies, and extending model longevity through better configuration and interface design.
• Model Retirement & Recycling. Few current workflows include structured model end-of-life protocols. The authors call for policies and systems to enable transparent model retirement, secure data deletion, and reuse of prior artifacts to prevent unnecessary retraining.

Challenging AI Development Assumptions

The study critiques several dominant assumptions in current ML practices that hinder sustainability. The belief that a single model should serve all contexts leads to oversized and inflexible architectures, while the default mindset that more data is always beneficial results in increased computational and environmental overhead. Likewise, prioritizing precision and recall above all else often marginalizes critical factors like usability, energy efficiency, and model adaptability.

In contrast, the authors advocate for development approaches that are contextual and user-centered. By designing smaller, more adaptive models that reflect specific needs and constraints, developers can deliver effective AI systems that also minimize ecological impact.

Toward Human-Centered Sustainability in AI

The authors envision a future where ML tools are democratized through intuitive design, structured sustainability education, and lifecycle transparency. Their modified lifecycle model explicitly incorporates sustainability checkpoints from data extraction to model retirement.

Key calls to action include:

• Educational Outreach
  Train developers on the long-term environmental and social impact of their decisions.
• Lifecycle Transparency
  Encourage better documentation and metrics to support accountability at each development stage.
• Multi-Stakeholder Governance
  Embed community feedback and ethical oversight throughout the ML pipeline—not just during deployment.

A Milestone for SustainML

This work reinforces SustainML’s core goals: aligning technical development with environmental stewardship, ethical accountability, and lifecycle-aware system design. By spotlighting how human-computer interaction (HCI) practices can bridge sustainability gaps in ML, INRIA advances a paradigm where responsibility is embedded in every decision, starting with design, not just data.

MORE INFORMATION ABOUT SUSTAINML:

If you have any questions, please get in touch with This email address is being protected from spambots. You need JavaScript enabled to view it..