Responsible AI use and the environment

One of the questions we are asked most frequently at the moment is whether using AI is bad for the environment. And what should an AI policy say about environmental impacts?

This is an area that can get complicated. Robust numbers are very hard to find, and numbers without context are hard to make sense of. We’ll keep refining our thinking as more reliable data becomes available. Here is where we are up to in September 2026. Alongside environmental considerations, responsible AI use always starts with asking why you’re using AI, and if the value it brings to a task is a worthwhile tradeoff.

Individual AI use is a tiny fraction of your overall environmental impact

We don’t have precise, robust numbers on AI energy impacts. More on this shortly! But well-researched estimates suggest that a single AI text query uses less than a watt of energy. We should treat this as an order of magnitude, rather than a definitive number. For now, it’s worth noting that a watt is small. A laptop uses about 1W per minute, and your brain uses 12W per minute. In the UK, we use about 11.5 kWh of electricity per person per day.

So running 300 AI queries might amount to 3% of your daily energy use. The bottom line for individuals is – using AI will not be amongst your major environmental impacts, even if you use it quite often.

Another way to put this in context is to look at How Bad Are Bananas, The Carbon Footprint of Everything by Mike Berners-Lee. In this book, digital and IT activities all appear in the first chapter, which means that they are considerably smaller than many of our other environmental impacts. Other activities like eating, heating and transport consume much more resources than digital activity. Hannah Ritchie has also tried to put AI use in context in a recent article.

If you’re concerned to minimise your own environmental impacts, it’s important to look at the bigger picture. For most people, transport, heating and what we eat are where we see our largest environmental impacts, and where the biggest savings can be made. Technology use has an environmental impact too, but it is fairly small by comparison.

The AI industry has a numbers problem

Numbers around AI environmental impacts are contested. This is because the industry is not sharing robust and comparable numbers. So it’s hard to compare product choices, evaluate lifetime impacts or push for more sustainable products. For a brief and insightful reflection on these issues, see Chapter 6 of Slow AI by Sam Illingworth.

The team at Epoch AI have attempted ‘bottom up’ estimates based on the computational processes involved in a typical AI query, and typical model sizes. Their estimate is likely to vary considerably based on assumptions about model sizes. Other researchers including Alex de Vries-Gao have looked at macro market trends to quantify AI’s rapidly-growing energy and resource demand.

Hyperscale data centres also consume large amounts of water – around 2 million litres per day for a 500Mw data centre. For context, this is around 10x the peak water demand of a golf course. As organisations such as Nature Finance have pointed out, lack of transparency means that precise numbers are hard to come by.

Three challenges come up in connection with water consumption. Firstly, many of the largest hyperscale data centres are located in areas that are already experiencing water stress – where supplies are being consumed faster than they can be replenished. Secondly, peak water demand from data centres is typically during hot weather, when the local and regional water stocks are depleted. Finally, there is a trade-off between power consumption and water use. Reducing power consumption often comes at the expense of increased water consumption.

This work suggests that: a) basic AI use is on a par with other digital activity such as streaming video, but b) AI development is fuelling a runaway spike in energy demand. Both pieces of research highlight the fact that real numbers are not being disclosed by the industry.

More transparency and openness from the AI industry might help people make informed choices about AI use, and allow policymakers to have enough evidence to make clear decisions around the longer-term energy and environmental tradeoffs.

What does this mean for your organisation’s AI policy today?

While it’s valuable to be intentional and strategic in your use of AI, it’s probably not your biggest environmental impact as an individual. What about organisations? How might you reference sustainability and environmental commitments in an AI policy? With limited data, it can be hard to draft a policy. But here are three points you could include:

• Recognising the real and growing impacts of AI development and deployment, we commit to seeking out reliable data about environmental impacts
• Where we have clear information about the energy and resource impacts of AI tools, we will choose the lowest impact tools that can do the job
• We’ll commit to minimising unnecessary AI use, since this incurs costs with no real benefit

These open-ended commitments will send a useful signal about your priorities, and will mean you are ready to choose between tools and platforms when sufficient information is available. 

It’s not all on the voluntary sector

Charities are values-led and working for a better world for everyone. This can sometimes create extra pressure on hard-pressed groups to reduce their carbon emissions. But Nick Addington and others have pointed out that charities probably already have a small emissions profile, relative to other organisations.

The real challenge for charities in a rapidly-changing climate is about ensuring that our mission and priorities are appropriate for a rapidly-warming world, where climate impacts are growing day by day. And if digital and AI tools help make your work more effective, this could be a worthwhile tradeoff that you shouldn’t shy away from.

What does it mean for society more widely?

Looking at the bigger picture, the rapid growth of AI is driving a race to build out hyperscale data centres. These use a city’s worth of energy each (hundreds of megawatts). Some are positioning Scotland as an ideal spot for ‘green data centres’. But this seems to be based on assumptions about Scotland’s abundant renewable energy supplies. Scotland does generate a huge amount of wind energy. But wind energy is intermittent, while data centres typically require 99.999% uptime (US providers highlight how they stayed running during a state-wide blackout in Texas). Hyperscalers are going to need a secure supply of electricity 24/7/365, not just when we have excess wind energy.

Organisations like Action to Protect Rural Scotland have also pointed out that just using renewable energy most of the time would mean anything built in Scotland would count as ‘green’. That’s a trivial and fairly meaningless definition. ‘Green Data Centres’ is a concept that needs much more clarity. APRS have added up all the hyperscale data centres in the Scottish planning system. If they were all built, that would add up to 1.75x Scotland’s current peak energy demand. 

So environmental organisations and local communities are challenging the rush for hyperscale data centres. Are there real economic benefits? Are these durable? And what are the long-term local and wider environmental costs? What would we be giving up?

This means data centre buildout will be a live social, environmental and political issue for a while in Scotland, and other parts of the world. In September, the Scottish Government announced that data centres in planning would stay on hold until planning guidance and environmental assessment processes are clarified.

Looking globally, major tech companies who had ambitious climate plans about five years ago (eg net zero by 2030) have now dropped these due to the rush to build out hyperscale data centres.

Charities and people working in them aren’t in a position to directly influence all of these issues. But it’s an important context we should be aware of. There’s a balance to be struck: using AI in line with our vision and values to be more effective, while demanding more transparency and accountability on the environmental impacts so we can make sustainable choices. Finally, asking ‘why AI?’ is always a useful question.