Spotlight on … limiting the environmental impacts of AI

In this spotlight on article Hannah Smith, Director of Operations at The Green Web Foundation, explores how the increasing use of AI (Artificial Intelligence) is impacting digital carbon emissions, and the different steps organisations can take to limit their impact.

UK AI adoption is accelerating, with nearly two-thirds (64%) of organisations reporting that they now use AI in some form, according to a 2026 Amazon Web Services’ report. While the proposed benefits of efficient workstreams sound appealing, it’s important to weigh up the pressure this rapid increase in demand imposes on the environment, with uncontrolled AI practices posing a threat to our net zero ambitions.

The problem

The rise of AI use has many environmental implications. Often focused on is the immense water consumption required to run AI, due to the expansion of data centres and the water needed to keep them cool. The UK already faces a projected daily water deficit of nearly 5 billion litres by 2050; data centre expansion will only exacerbate this. The daily water demand for a 100MW data centre is 2 million litres and we currently have around 5.3 GW (5300MW) of data centre expansion planned, much in already water-stressed areas, such as the south-west of England.

The demands from data centres and AI go beyond water. They are hugely carbon intensive with energy consumption from data centres predicted to skyrocket. AI hardware requires more energy to run than other digital technology and increasing AI demand is a key contributor to the leading energy think tank, IEA, estimations that digital infrastructure will become the biggest consumer of energy in years to come.

Increased AI use is also accelerating extractive industries with the chips used requiring more resources to manufacture. AI systems themselves are also being utilised to speed up fossil fuel extraction processes.

It’s important to remember that the impact of AI goes beyond environmental implications. Responsible and ethical AI use requires consideration of social, political, creative thought and privacy issues. These are important and require careful review, but this article will be focussing primarily on carbon emissions.

Breakdown of emissions

AI emissions arise from:

• Manufacturing: building the hardware that allows the use AI.
• Training of AI: the data analysis required to build and train AI is very intensive but often overlooked in emissions reporting.
• Use / interface of the AI model: including the energy required to host the data centre, network emissions from data transfer, and the energy from using the device that allows us to interact with the AI data.

To be truly sustainable when it comes to AI, we would abstain from using it entirely, but in a digitised world focused on efficient systems, this is not always realistic. So, how can your organisation limit your carbon impact if you do choose to use it?

Step one: question your need to use AI, and know when and how to use it responsibly

Not all AI tasks have the same carbon impact. The main AI tasks, in order of impact, are:

• Extractive QA = pull verbatim answers out of text; a search function.
• Text generation = create new text based on patterns it has learned.
• Summarisation = condense text into a more manageable form.
• Image captioning = identify what the image is about and generate text that describe in words the different elements that make up the full picture.
• Image generation = create images from textual descriptions or prompts.

For comparison; the energy required for a single extractive QA is 0.003 KwH, for image captioning 0.063 KwH, while image generation greatly outweighs all other AI tasks at 2.907 KwH per action.

Step two: optimise your AI use – if you have to use it, choose responsible suppliers

All AI code can be run in ways that reduce its footprint. Alongside different AI requests having differing impacts, different models will use different amounts of energy to do similar tasks.

When selecting a model you can use:

• Task-specific; fine-tuned models that carry out a single task.
• General purpose; those trained for multiple tasks.

Task-specific models will have a lower carbon impact, as the more specific the AI model is, the more efficient it will be, reducing emissions. However, while you can limit your impact through being selective with models, across all servers’ image generation will still be the most energy intensive task, and text classification the least.

Learn more through this study.

Step three: prioritise getting estimates of AI’s footprint

Ensure your supplier is being the most environmentally friendly that they can be by asking for emission data, making sure to scrutinise the data you are being provided. Accurate data should consider the whole lifecycle of an AI model including manufacture, training and use. Carry this out by:

• Contacting your supplier directly. Some suppliers, such as Microsoft, will provide platforms to access this environmental information.
• Checking the environmental credentials of AI platforms through their model card; information files that accompany AI models including uses, limitations, biases and ethical considerations. This can be done through Hugging Face.
• Please note, including AI carbon emissions on the model card is currently an optional component, and all information is self-reported, leaving room for errors.

It is worth noting that the industry is still in the early stages of making environmental information of AI platforms readily available, with environmental credentials often being hard to access and understand.

Step four: provide direction for staff

With the rise of AI use and the discourse around it, organisations are starting to implement policies to guide staff on responsible use. Measures to consider include:

• Outlining where AI is appropriate to use and when it should be avoided, such as carbon-intensive image generation.
• Stipulating what data and which platforms can be used. If you are using a generative open-source AI platform your data can become publicly accessible or be used to train future AI models, which can pose data privacy implications for your organisation.
• Implementing regular reviews of AI use within your organisation and its environmental impacts. Consider designating a staff member or team to oversee AI procedure development.
• Provide staff AI training to raise awareness of the policy and best practice.
• Going beyond environmental impacts – such as the inaccuracies of AI results and social implications.