Generative AI and carbon emissions in the Oxford context

A lone tree stands amid glowing digital lines on a dark surface symbolising nature meeting technology (©Getty Images)

The growing use of generative AI tools has sparked questions about its environmental impact at Oxford and beyond. This article by the Digital Transformation team explores the ongoing debate about this subject, and how the University is responding.

As the widespread adoption of generative AI continues, there is an increasing focus on its environmental impact. Questions include:

• Are we increasing our carbon footprint by using AI tools?
• Is it possible to balance innovation with our commitments to net zero and sustainability?
• What can we do to minimise the impact of the technology?

There are a wide set of socio-environmental issues surrounding increased generative AI adoption, including water usage and rare earth metal consumption. This article focuses on energy consumption and carbon emissions.

Researching the real impact

One of the challenges of considering this field is getting a real understanding of the energy usage and environmental impact of generative AI.

The International Energy Agency estimates that data centres account for around 1.5% of global electricity consumption, which is due to rise to 3% by 2030 following the trend of growth in large-scale facilities caused by AI demand. However, other studies have suggested different projections such as the UN’s International Telecommunication Union (ITU) and World Benchmarking Alliance (WBA) 2025 report on Green Digital Companies which further explores how providers are working to improve their footprint and reduce this impact.

Some of the leading research in this area is being undertaken by researchers here in Oxford. Dr Yee Van Fan, Senior Researcher in the Environmental Change Institute, said:

It is challenging to provide a definitive range for the energy consumption of AI or GenAI, as it depends heavily on the type of application and the model used. Several studies have attempted to estimate this, but the numbers vary widely due to a lack of transparency and the reliance on estimates, rather than actual consumption data.

The technology is also changing over time. Professor Charlie Wilson, Senior Research Fellow in the Environmental Change Institute, pointed to Deep Seek as one example:

Deep Seek was one of several generative AI models trained using a fraction of the energy-hungry chips compared to the big ChatGPT type models. There is lots of potential to improve the energy efficiency of generative AI model software, the hardware on which this runs, and the data centres that house the hardware.

Further examples of how the industry is working to decrease the energy consumption of model training have been explored by IBM and include hardware improvements and efficiency, smarter model training and collaboration between developers. 

Balancing impact and innovation

At the University, generative AI tools are increasingly being used across teaching, research and professional services. There are a range of tools and applications in place, and there is a growing focus on how innovative approaches can be balanced with the impact of these technologies.

The AI and Machine Learning Competency Centre is at the forefront of this work and has developed a pilot of premium subscriptions to generative AI tools such as ChatGPT Edu and Microsoft 365 Copilot.

The pilots have allowed for more accurate monitoring and supported more secure use. One of the focuses of these pilots is whether AI can improve people’s work and productivity with widespread usage. They are also helping to inform University investments in secure, energy-efficient AI tools.

Alwyn Collinson, Head of the AI and ML Competency Centre, said:

The AI Centre is trying to understand the impact of AI use on Oxford's ambitious environmental targets. This includes learning how many people are using generative AI, and what for, as there are big differences in the energy costs of different AI tools. That will help us understand how we can deploy generative AI appropriately and mitigate the carbon emissions that are an inevitable consequence of any new technology being widely adopted.

Estates and buildings

Work is also underway in other parts of the University to consider how AI may reduce Oxford’s carbon footprint. Estates Services, for example, is exploring opportunities to use technology and trialling sensors to manage buildings as efficiently as possible through the Smart Building project. This will use AI alongside other technologies to optimise how we manage our buildings and use space more efficiently in terms of cooling, heating and lighting areas.

Navigating complexity

A key challenge for colleagues considering GenAI use is the sheer complexity of the issues involved.

The team at the University’s Gardens, Libraries and Museums (GLAM) is working on a project to understand its carbon use across its digital activities, including the impact of AI.

Jon Ray, the GLAM Environmental Sustainability Manager, noted that the picture remains unclear due to the fast-moving nature of the field and a lack of transparency from technology providers. Nevertheless, he emphasised that AI offers clear benefits and should be used considerately rather than avoided entirely:

Emissions associated with AI are in addition to our existing activities. However, different activities have different impacts and we need to break down the problem into different scenarios. From colleagues building an AI model for research purposes, to desk-based colleagues using AI for day-to-day querying. Across all these scenarios, for now at least, our best approach is to be considerate, continually build our knowledge and ensure sustainability features in our ongoing discussions.

Holding partners to account

Most of AI’s energy usage comes from training models, rather than individual queries. While the use of AI is a factor in the environmental impact, most of its energy usage comes from training models, rather than individual queries.

The University is therefore engaging with vendors like OpenAI and Microsoft to push for sustainable practices. Oxford’s close relationship with these companies means it can directly question providers on how they are delivering their net zero targets.

Stuart Lee, Acting Chief Information Officer, said:

We are constantly talking to the vendors about how they are limiting the impact of these tools and what their targets are around net zero carbon emissions. This conversation varies from vendor to vendor.

Many of our partners have their own statements on their green technology ambitions, such as Microsoft’s commitment to becoming carbon negative by 2030. 

The Bodleian Libraries working in partnership with OpenAI has reflected their sustainability goals in the agreed Statement of Work, inviting OpenAI to support the University’s environmental ambitions.  

Towards a strategic approach

The University has an existing commitment to sustainability – as laid out in the Environmental Sustainability Strategy. Approved in 2021, prior to the current surge in generative AI use, it sets out targets to achieve net zero carbon and biodiversity net gain by 2035.

In light of these emerging technologies, we spoke with the Environmental Sustainability team to understand how AI is affecting these aims.

Paul Cross, Head of Energy and Carbon in the Environmental Sustainability team, said:

The Environmental Sustainability Strategy will be reviewed over the coming year, and the impact of new technologies and AI will form part of the discussion. These tools use more processing power, more energy and therefore more associated carbon emissions. Used in the right ways, they can offer benefits and efficiencies.

What’s next for AI and sustainability at Oxford

Understanding how AI will affect not only the University’s carbon footprint, but also everyday research, teaching and administration is an ongoing process. The University is actively monitoring and adapting to this fast-changing field and building a clearer picture of how AI is being used and how this aligns with sustainability targets.

In the meantime, Professor Charlie Wilson highlights that colleagues who are concerned about AI should consider how they are using it, as well as their wider energy consumption:

Like with any energy-hungry activity, use AI judiciously. But if you’re worried about your carbon footprint, look first to your travel, diet and home energy performance – they’re the really big contributors to household emissions.

While there are no easy answers to the debates about AI and the environment, there is a clear commitment across Oxford to consider the risks and benefits of its use, and its relationship with the environment.

Through considering our own use of AI, working closely with vendors, supporting smarter energy use and helping staff make informed choices, sustainability considerations will remain a key part of our digital future.