The increasingly intense public debate around AI highlights growing demands for the technology's potential to be realised in ways that align with sustainability, security and democratic values. Research on AI at Linköping University spans a wide range of disciplines and is rooted in interdisciplinary methods, generating strong research environments and research with high societal relevance. Against this backdrop, the university's Centre for Sustainability Transformations and the profile area Visual Digital Futures recently brought together researchers from computer science, machine learning, physics, media and culture studies, and environmental social sciences, for a cross-disciplinary discussion on the future of sustainable AI.
Governance and responsibility
The researchers united in a call for a more responsible approach to AI development, stronger political governance and more focus on long-term sustainability: "AI holds enormous potential, but the technology must be used for the benefit of humanity," said Björn Alling, Professor of Theoretical Physics and Deputy Director of NAISS, Sweden's National Academic Infrastructure for Supercomputing. "If we get this right, AI could help us achieve major advances in fields such as medicine and reduction of working hours."
Per Wistbo Nibell
Marie Francisco, Postdoctoral Researcher at Thematic Studies Environmental Change who studies AI governance, stressed the importance of the European Union taking a more active role as a regulator:
"The EU can play an important role by setting requirements for responsible and ethical AI. With stronger regulation, companies should not be able to access European markets without adequate safeguards," she said.
"At present, many aspects of AI systems are still ungoverned. Everything in society can be regulated in one way or another, since that is ultimately the role of politics. We need to place greater demands on politicians and other societal actors to become actively involved in discussions about the kinds of AI systems we want, and which types of AI may need to be delayed. Decisions about safety and security cannot be left to developers alone."
She also calls for independent groups of AI experts that could provide guidance free from economic incentives.
Today, many aspects of AI are still ungoverned, despite the fact that virtually everything else in our society is governed in one way or another.
Making AI's environmental footprint visible
But fundamentally, reducing the environmental impact of AI starts with understanding and identifying this impact:
"Sustainable AI is not AI that is efficient, it is AI that is accounted for," says Sebastian Mair, Assistant Professor at the Division of Statistics and Machine Learning, Department of Computer and Information Science. "A PhD student today can tell you to three decimal places how accurate their model is, but not what it cost to produce."
Per Wistbo Nibell
His research explores how different forms of data reduction can form the basis of more sustainable machine learning, whether through selecting smaller representative subsets of training data, more efficient model training, or changes in the way models are developed. For Mair, greater transparency is an important part of the solution. Researchers and developers need to become better at measuring and reporting the resources consumed by AI systems during both training and everyday use:
"I would like AI-driven research computations and experiments to automatically track energy consumption and emissions, ideally built into the computing infrastructure itself," he said. "In the same way that consumer products can carry environmental labels, we should be able to see the carbon footprint associated with different computational processes.”
What makes an AI system sustainable?
AI's growing energy consumption has attracted considerable attention in recent years. But the researchers were united in the view that sustainability cannot be reduced to electricity use alone.
Miriah Meyer, Professor at the profile area Visual Digital Futures, noted that AI has been at the centre of an unusually dramatic news cycle in recent weeks. Media reports have ranged from a major mathematical breakthrough to growing opposition in the United States to the construction of new data centres, as well as an intense debate about AI safety following warnings from researchers associated with Anthropic that rapidly advancing self-improving AI could eventually pose an existential threat to humanity.
"Many people feel powerless in the face of large technology companies," said Miriah Meyer. "By discussing both the opportunities and the challenges of AI in interdisciplinary settings, we can better explore what alternatives exist and which values should guide AI development in the future."
We need to agree on the values that should underpin AI development.
Per Wistbo Nibell
Using the Brundtland Report's classic definition of sustainable development, meeting present needs without compromising the ability of future generations to meet their own, the researchers agreed that many different factors determine whether an AI system can be regarded as sustainable. Marie Francisco emphasised that sustainable AI is also about resource use and social consequences:
"A sustainable AI system must also be ethical. We need to ask how people and local communities are affected," she said.
Julia Velkova, Professor in the Department of Culture and Society, has been researching data centres for more than a decade. She warned against oversimplifying debates about AI, by for example reducing sustainability to a checklist.
"We need to include the temporal dimension in all discussions about sustainability in technological systems," she said. "Technologies are constantly changing, which means that what is sustainability is always in flux."
Can AI be sustainable if it creates negative social impacts?
Per Wistbo Nibell
Velkova argues for a broader perspective on AI sustainability. Rather than focusing exclusively on technical systems, research should also examine the people and places affected by expanding digital infrastructure. She pointed to examples where data centre developments have been promoted as sustainable investments despite creating environmental or social costs for nearby communities.
"We often see policymakers willing to give away too much to major technology companies far too easily," she said. "This has happened in places such as Ragunda, where the establishment of a large wind farm financed by Google gave Google valuable green credit points, while degrading the local environment for citizens and a substantial part of the winter reindeer grazing grounds of several Sami villages. It demonstrates why researchers need to shift perspective, from studying systems alone to understanding how people and places are affected, from documenting gains to recognising losses."
According to Velkova, terms such as quality of life and livability may sometimes be more meaningful than sustainability as a broad umbrella term.
None of the researchers considered the kinds of existential risks recently discussed in the media to be obvious concerns. However, Björn Alling pointed to the potential risks that could arise if AI-driven cyberattacks are combined with physical weapons systems:
"We can introduce safeguards against cyberattacks. But if physical weapons are added to the equation, we are in dire straits," he said.
Scaling up Europe's AI infrastructure
Per Wistbo Nibell
Another aspect of sustainability concerns Europe's dependence on AI infrastructure developed and controlled by major US technology companies and, to a significant extent, powered by fossil energy sources. The panel discussed how Europe might strengthen its own infrastructure and develop its own AI models.
Björn Alling argued that Europe already possesses much of the expertise and infrastructure required to increase its technological independence:
"We are unlikely to outcompete the largest American technology companies," he said. "But we can build foundational European and Swedish AI systems and avoid some of the lock-in with US tech companies."
