Refining how AI impacts are measured

• The group explored alternatives to token‑based allocation of training emissions, particularly for video generation, and discussed more meaningful functional units such as pixel‑seconds for visual outputs.

Spotlighting underexplored areas

• While much existing research focuses on text‑based models, participants highlighted the need to address video, audio and multi‑modal generation, where model behaviour, output quality and compute intensity vary significantly.

Clarifying scope and boundaries

• Consensus emerged on including full data‑centre energy impacts (including cooling and auxiliary loads) within scope, while noting related issues such as water use, rebound effects and increasing on‑device computation.

Constructive industry engagement

• The importance of balancing transparency about limited public disclosure with a collaborative tone toward model developers was emphasised.

Positioning against existing frameworks

• The group agreed the methodology should clearly state where it aligns with established approaches and where it introduces new thinking, particularly for commercial video generation.

Grounding the work in real use cases

• Members were invited to contribute detailed examples of GenAI workflows across media, including available metrics and counterfactuals, to ensure the methodology reflects real‑world practice rather than theoretical models.

The discussion reinforced the value of shared, practitioner‑led insight in developing credible, decision‑useful guidance on AI sustainability, with further data‑gathering and collaboration planned ahead of the next phase.