Green AI Model Cards
With this new model card the environmental factors of AI models are as easily comparable as a card game.
Challenge
AI Model Cards - a new concept to standardize environmental reporting for AI models
The challenge is to propose clearly defined sustainability indicators that could be included as part of a standardization drive for AI model cards, which can provide regulators with comparable benchmarks that measure sustainability practices of AI models across regulatory environments
Team
Kerstin Forster, Sarah Grotkasten, David-Peregrin Hain, Kristina Pavlovic, Suchithra, Negash Desalegn Weldekiros
About the prototype
Our findings, while researching the current situation of AI model cards, showed a lack of sustainability indicators concerning environmental factors used by companies. The theoretical framework around environmental indicators exists though and allows to find around 60 indicators in a short timeframe. We developed two life cycle models to display the levels of environmental impact: The AI hardware life cycle focuses on the areas of production, transport, operation, further development and beyond-life/end-of-life, while the AI software life cycle displays data acquisition, model training, model deployment and beyond-life/end-of-life. The list of indicators along these life cycles can be used for creating AI model cards for which we created a design which allows simple comparability between different AI models. As a result our Green AI Model Cards are enabling citizens as well as companies and states to make informed decisions when it comes to the environmentally sustainable use of AI.
Outputs
Further information about the progress, milestones, and roadblocks.
Impressions
