Measuring environmental impacts of AI

How to measure the environmental impact of AI? This was another significant topic discussed at the AI Standards Hub Global Summit, and here is what I learnt:

How to define boundaries concerning training runs and telecommunication costs?

What constitutes a training run? Is it a single iteration of an algorithm? A complete training cycle until convergence? Or the entire lifecycle of model development, including experimentation and hyperparameter tuning? Each definition will result in different energy consumption figures.

AI applications frequently involve massive data transfers between devices (e.g., smartphones, sensors) and data centers.

Measuring the energy consumed during these transfers is difficult, as it depends on many factors.

Practical challenge lies in obtaining the granular, accurate data needed for precise measurement:

Some data may be proprietary (for example, energy used to transmit data for a specific AI app) or impossible to obtain (like environmental impact of the mining and manufacturing of all the hardware used in an AI system).

AI is deeply intertwined with broader digital technologies, and it is difficult to isolate the specific environmental footprint of AI from the overall impact of digitalisation.
Indirect environmental impact you may not be considering - AI-driven changes in consumers’ bevaviour:

you may decide to replace your older but still functioning phone with a newer one that has AI-powered features sooner that you would have if not for the AI development. This increased consumption cycle adds to the overall environmental burden.

Electrification of transportation is increasing demand for electricity alongside AI.

This puts a strain on existing electric grids and raises concerns about the environmental impact of electricity generation, especially if it relies heavily on fossil fuels.

Some of the proposed solutions?

Accelerate and improve the environmental impact assessment of AI: empower standardisation experts and AI and other researchers (bringing their research methodology knowledge to the table) to work together.
Use the increasing digitisation of standards to enable faster research (for example, through cross-referencing and using existing dictionaries)

Many thanks to Dr Valerie Livina CMath FIMA, Juliette Fropier, Arti Garg and Uzma Chaudhry for this fantastic discussion!