Environmental researchers know the impact of AI – so why do they still use it?

When it comes to AI, much of the discussion within universities is focused on students’ use of generative AI tools and the energy demands of large language models. But this narrow focus risks overlooking a significant part of the picture: academics themselves are increasingly using AI for research.

From collecting and managing large datasets to generating synthetic data, analysing data with machine learning techniques and using LLMs to support literature reviews, coding and writing, AI is becoming embedded in everyday academic research.

• Greener labs don’t need bigger budgets – just better habits
• Go green, AI!
• Start your university on a path to resilient net zero

At our university, we’ve been conducting research into how environmental researchers reconcile the use of AI with their values. We found that researchers are generally aware of AI’s environmental impacts, but that this rarely translated into sustained changes in practice. This was surprising, given researchers’ commitments to environmental values and their work to try to address environmental problems, so we explored what was going on.

While a small number of researchers tried to reduce their environmental footprint by using smaller, locally hosted models, most managed the tension between AI methods and their environmental research through a range of strategies that shifted responsibility for the impacts of their own AI use elsewhere.

Most commonly, responsibility was shifted upwards to universities, funders and institutional systems. Decisions about cloud storage, computing infrastructure and access to servers were seen as beyond individual researchers’ control, while growing pressures to publish and demonstrate productivity made it difficult to prioritise slower, less computationally intensive methods.

Responsibility was also shifted sideways to other researchers who were perceived to be using larger and more resource-intensive AI systems. Yet what counted as a larger model was rarely defined, creating blurred boundaries between acceptable and excessive AI use.

Others shifted responsibility further afield, pointing to technology companies, data centres and global supply chains as the real source of environmental harm. Some researchers looked to policymakers to address the problem, while others framed AI’s environmental impacts as an issue for the future, to be tackled when better technologies, more efficient systems or stronger governance emerge.

Together, these ways of shifting responsibility enabled researchers to continue using AI while avoiding the uncomfortable knowledge that their AI use was negatively impacting the environment.

However, our findings suggest that placing responsibility for these impacts solely on individual researchers is neither realistic nor desirable. Researchers occupy the end of the computational supply chain and have little influence over the design of AI systems, the sourcing of energy, the procurement of computing infrastructure or institutional incentives that shape research practices.

University leaders, by contrast, are in a position to take responsibility for the environmental consequences of research using AI. If they are serious about reducing the environmental footprint of research, responsibility must extend beyond individual choices and be addressed at the institutional level.

We suggest four actions that universities can take to reduce the environmental impacts of research using AI:

1. Measure and make impacts visible

University leadership teams should integrate carbon- and compute-tracking tools into research computing systems and provide dashboards that allow researchers, departments and institutional leaders to understand the environmental impacts of their work. 

Environmental costs are often invisible, making them easy to overlook. Measuring and reporting these impacts is a crucial first step towards informed decision-making and institutional accountability.

2. Make low-impact options the default

The researchers in our study who actively sought to reduce their environmental footprint often had to do so on their own initiative. Universities can make sustainable choices easier by providing access to smaller open-source models, locally hosted AI systems and shared energy-efficient computing infrastructure. AI additions should not be the default option in provided software. Procurement policies should also prioritise providers that are transparent about energy use and emissions. The goal should be to make lower-impact approaches the easiest and most accessible option, rather than leaving researchers to navigate these choices alone.

3. Embed AI sustainability into research governance and incentives

The environmental impacts of AI should not sit in a separate AI policy silo. Instead, universities should integrate consideration of the environmental impacts of AI into existing policies and processes such as net zero, sustainability and climate strategies, procurement decisions, research integrity frameworks, research ethics review and data management planning.

Consider how funding, performance and promotion criteria may unintentionally reward increasingly compute-intensive forms of research. Embedding sustainability into governance and incentive structures helps to ensure that it becomes a routine part of decision-making, rather than an optional add-on.

4. Create clear institutional ownership

Responsibility for AI’s environmental impacts is often dispersed across research support teams, IT services, procurement offices, sustainability teams and academic departments. Establish a cross-working group with the authority to coordinate action and set policy across these areas. Without clear ownership, responsibility can easily fall through the cracks; with it, universities can develop a coherent and responsible approach to sustainable AI.

The challenge is not simply helping researchers to make better choices. Researchers sit at the end of a complex computational supply chain and have limited influence over many of the factors that shape AI’s environmental footprint. If universities want researchers to use AI responsibly, they must create the institutional conditions that make responsible use possible, visible and rewarded.

Sarah Hartley is professor of technology governance at University of Exeter Business School and director of the Centre for Responsible Innovation; Emily Robinson is a researcher in emerging technologies, AI, society and sustainability; Mayra Rodriguez is a research fellow specialising in AI and data science for environmental and health applications, all at the University of Exeter.

If you would like advice and insight from academics and university staff delivered direct to your inbox each week, sign up for the Campus newsletter.