Extending the SCI to AI, addressing the challenges of measuring AI carbon emissions
Extending the Software Carbon Intensity specification to AI systems
AI is scaling fast. But how do you actually measure the carbon footprint of an AI system? Training runs, inference workloads, fine-tuning, data pipelines. The existing SCI specification gives us a solid foundation for measuring software emissions, but AI introduces unique challenges that the current spec wasn't designed to address. This assembly will tackle that head-on.
Over six 90-minute sessions, a small group of selected participants will work collaboratively to define the software boundaries for AI systems, identify appropriate functional units, and draft a baseline "SCI for AI" specification. You'll compare existing measurement approaches (Energy Score, Green AI Index, EcoLogits, and more), debate what should and shouldn't be in scope, and seek consensus on hard questions like whether and how to include training emissions.
This isn't a lecture series. It's hands-on, collaborative work using Miro boards, GitHub, and structured facilitation. By the end, you'll have contributed to a rough draft specification that will transition into the Software Standards Working Group for formal development. You'll be part of the group that shapes how the industry measures AI's environmental impact.
If you have expertise in AI systems, sustainability measurement, or standards development (and you're ready to roll up your sleeves), this is your chance to help define the rules before everyone else is playing by them.
