5a. Take a green approach to AI used by the service
The use of Artificial Intelligence on a service could include analysing datasets, such as satellite or weather data. It could also include generative AI to create new content, such as text.
These technologies may bring about benefits as well as costs from an environmental sustainability standpoint.
Sub-actions
5a. (i) Be clear on use case for AI
5a. (ii) Choose the appropriate size model
5a. (iii) Reuse existing models where possible
5a. (iv) Carefully choose supplier and region
(i) Be clear on the use case for AI
Be clear on the use case for generative AI and whether an alternative non-generative AI approach such as image recognition, natural language processing would be more appropriate.
AI Playbook for the UK Government: ‘Assess the environmental impact of training and/or deploying your AI system before commencing development. Consider whether the impact represents a reasonable trade-off between benefits and energy consumption, and whether a less energy-intensive system might be able to achieve the same or similar results.’
Environmental benefit:
Being clear on the use case for AI, and likely environmental impacts, helps ensure the right choices are made.
Read more
(ii) Choose the appropriate size model
Services will use AI for different tasks. Small AI models can be trained for a specific task, are cost effective, quick (as there is less data for them to reference) and they can be run on basic hardware (such as a PC or phone).
Having good AI evaluation/performance metrics will allow teams to make effective choices about the smallest model they can ‘get away with’ using.
Environmental benefit:
Smaller AI models consume less resources to train and operate compared to larger ones.
Read more
(iii) Reuse existing models where possible
Consider building on an existing model rather than starting a new one from scratch.
Environmental benefit:
Reusing an exisitng model saves on the impacts assoicated with developing a new one.
Read more
(iv) Carefully choose supplier and region
Choosing the right AI supplier and location is important. You can reduce the carbon footprint by training models in low-carbon regions or times of day.
Keep in mind the other environmental impacts of AI, especially water use. Always consider regional differences. For example. water use will have a different impact in Scotland than Suffolk.
Use third-party auditors to help select the most environmentally-friendly supplier.
Teams might be able to swap a well-known model for a more sustainable and similarly-effective option (e.g. DistilBERT vs BERT) - especially if they have the service performance metrics to support their decision
Environmental benefit:
Choice of supplier and region for AI models can have a significant impact on carbon footprint.