If generative AI is increasingly becoming part of artistic practice, its environmental cost remains stubbornly absent from contemporary discussion. This is despite OpenAI CEO Sam Altman’s admission earlier this year that, “without a breakthrough,” global energy systems will struggle to cope with the demands of new AI systems. During the NFT explosion of 2021, digital artists played a vital role in raising critical awareness of the environmental cost of crypto. The problem now is whether and how artists can harness the creative potential of machine learning in a sustainable manner.

A surge in the use of generative AI to produce new text, image, and video is fueling demand for AI-accelerated servers that is increasing water consumption and resource extraction. According to a recent report by TechInsights, AI GPUs (graphical processing units) are forecast to consume 1.5% of the world’s electricity over the next five years — an increase of 300% — generating 1.1 billion metric tons of carbon emissions in the process. Meta, Microsoft, and Google have all reported increased emissions from their data centers, which, coupled with significant emissions from semiconductor manufacture, is challenging their shared ambition to reach net zero by 2030. 

The fact that image generation is the highest emitting task for machine learning models to perform only reinforces the urgent need to develop a more sustainable framework for artists. To this end, Diane Drubay brought together Elisabeth Gravil, a specialist in AI and digital transformation of cultural institutions, Alistair Alexander of Reclaimed Systems and Gallery Climate Coalition, whose work focuses on the impact of technology on ecological systems, and poet and artist Sasha Stiles, who has embedded nature-driven concepts into her practice from the very beginning. 

‍Diane Drubay: AI has become integral to many artists’ practices, whether to generate visuals and soundscapes or else to explore text and code. Over the past decade, artists such as Memo Akten, Sougwen Chung, and Mario Klingemann have all explored new forms of co-creation with machines. As AI rapidly reshapes artistic expression, what deeper insights or challenges is it introducing into your creative workflows?‍

Sasha Stiles: Artists today use AI in incredible ways, from visual explorations to sonic experiments to literary endeavors; from generative text to writing code. At the heart of it all is the idea of co-creation with a system that augments our human capacities, which, of course, has deep roots in human culture; automata are present in ancient myth, and artists have been exploring the power and meaning of algorithmic expression for ages. Artists have always been early adopters of such seismic technologies because inventions of this order of magnitude — literacy, linear perspective, the printing press, cameras, the personal computer, the internet — evolve consciousness, reshape self-awareness, our perception of the world around us, and our place in it. 

Since I was a student, my poetry has reflected on language in such contexts, and on the relationship between humans and our technological offspring. I started integrating natural language processing into my poetry in 2018, using large language models like GPT-2 and GPT-3 that I fine-tuned on my own work to generate text, serving alternately as co-author, editor, and muse. It has continuously challenged my understanding of language itself, of imagination and originality, forcing me to reconsider what I thought I knew about inspiration and the creative process, and taking my poetry into uncharted territory. 

To me, AI is a way of thinking and seeing, a lens that allows us to gaze upon and examine aspects of ourselves and the world around us that were previously invisible, inaccessible, unthinkable.

‍DD: Artists have always engaged with emerging technologies, but AI presents unique challenges and polarizes opinion. Is the art world moving toward or away from AI?‍

Alistair Alexander: The art world appears trapped in the “reality distortion field” around AI, responding to it either as a genuinely transcendent new technology with limitless potential or as one that presents existential risk — in other words, in precisely the same terms as the tech industry is presenting it. I feel that the art world needs to define its own terms for engaging with new technologies. Generally, it looks as if the environmental impact is not fully understood or acknowledged as a critical factor when considering AI, or any other technology for that matter. In our current moment, we need to consider the ecological impact of everything we do before pretty much everything else.

To be relevant, the art world has to offer some kind of reflection on the world around us. We are long past the point where we can embrace new digital technologies without any consideration for their wider implications, environmental or otherwise. We can see that unaccountable digital platforms have had huge negative (as well as positive) impacts on our informational, political, and economic spheres, and so it is also abundantly clear that however AI technology develops in the future, it is highly likely to accelerate us further on those trajectories. 

The art world is perhaps the most effective counterpoint we have to breathless technological accelerationism. Art is perfectly positioned to slow things down, to invite us all to reflect, and to ask: “Can we pause a minute? Are we really sure about this?”

‍DD: Yet while artistic applications of AI are advancing rapidly, its environmental costs remain overlooked.‍

Elisabeth Gravil: There are those who see AI as an ecological disaster and others who regard it as the magic wand that will solve everything, even our energy crisis. The truth, as usual, lies somewhere in between. 

Talking about “the impact of AI” in a global way doesn’t really make sense. In AI, we talk about machine learning, deep learning, and generative AI. Each has a different impact depending on the amount of data trained as well as where the data centers are [located], and the algorithmic model structure. 

The carbon footprint also depends on the source of energy used by the data centers and transportation infrastructures — if it uses coal-fired or nuclear power stations, or else renewable energy.

‍DD: Sasha, what do you see as the relation of natural systems to computational or AI systems?‍

SS: A lot of my interest in so-called nonhuman intelligence in the form of AI stems from a broader curiosity about diverse spectrums of intelligences — plant intelligence, microbial networks, mycelial webs, animal cognition, geological rhythms. This has been core to my work for a long time, for example in poems like Humanature (2020) and Every poem starts with a seed (2023), and in my series: “Analog Binary Code” (2019-ongoing), a digital language written in elements found in nature; “Regenerative Poems” (2021) coded in sunlight and wind; and “EcoPunks” (2021-ongoing).

Recognizing intelligence across these natural systems invites us to see AI not as “artificial” but as part of a wider ecosystem of perception and responsiveness. 

The paradox of technological advancement is that human progress goes hand in hand with ecological and other kinds of destruction. While AI offers tools to address urgent societal issues, its energy consumption evokes industrial-era resource exploitation. This tension is implied by the name of my first book, Technelegy (2022), fusing the exhilaration and momentum of technology with the elegiac mode.

‍DD: The environmental cost of AI is often discussed in vague terms, making it difficult to associate tangible actions with measurable results. How can we drive a more knowledgeable conversation about the ecological and social costs of AI?‍

EG: Cooling all the servers that run AI requires an enormous amount of water, along with resources to manufacture hardware, including chips such as NVIDIA GPUs, computers with new AI-integrated keys, or our own smartphones or wearables — all of which add to the environmental footprint. As most large models are proprietary, it is very difficult to know how they are designed and to verify their energy source. Data centers are not only dedicated to AI, which complicates calculations. 

We must not forget that the impact of AI is not only environmental, but also social. Underpaid “clickworkers” in countries like Turkey, Kenya, and the Philippines, are often forgotten in the process.

In the case of generative AI models, we need to distinguish between two phases: upstream model training and downstream generative use (or inference). On the training side, large models such as ChatGPT and Stable Diffusion, with their highly complex neural networks, require astronomical amounts of data and months of training on specific GPU chips that are neither cheap nor easy to obtain. As for inference (generating the response), according to a study by Sasha Luccioni, 1,000 prompts in ChatGPT require around 0.042 kWh of electricity. However, the variation in CO2 per kWh is x10 between where the data center is located and if it is powered by coal or nuclear energy. If we look at image or video generation, it is worse, but there aren’t many figures and few papers have been published on the subject.

‍DD: Now that AI is accessible to everyone and integrated into many everyday tools, its appeal can sometimes lead artists to adopt it without considering its full impact. What are the most practical and ethical ways for artists to start approaching AI?‍

AA: All leading generative AI models rely on massive training datasets, requiring weeks on constellations of advanced GPUs, which results in significant carbon emissions. However, the most notable carbon emissions occur during inference — using the model to generate responses. Here, users can make choices to minimize emissions.

The carbon impact of a task varies by type. Shorter text queries are less compute-intensive than longer reports. Image generation is, on average, 60 times more intensive than text, and video generation requires even more. 

Moreover, model type affects emissions — smaller versions are more efficient, designed to run on personal devices. As an artist, you need to consider how your work will be used. If it is only you taking the results and then applying them in your work, that is clearly going to have a far lower carbon impact than, say, a digital artwork on a very popular website inviting hundreds of thousands of visitors to enter their own prompt. Tracking energy use is key for model training, and using a downloaded model on personal hardware provides control over emissions. 

‍EG: When launching a project using AI, it is important to take a step back and start by asking yourself: “Do I really need AI for this?” It may seem obvious, but AI is often used simply for the buzz when a simpler solution might suffice. If you realize that AI is essential, turn to “lightweight” solutions like machine learning or non-generative deep learning. If, in the end, you need to go down the generative route, it’s essential to look at which model has the smallest carbon footprint. 

Open-source models often allow you to work locally rather than in the cloud. If the cloud is necessary, choose one that is geographically close, and plan for slower training on more traditional chips with servers that are less busy. It may seem counterintuitive, but it’s a good way of not encouraging the mass production of new chips, which requires the extraction of rare earths. For public projects such as live generative installations, remember that the impact of AI is not limited to the model training phase, but also the footprint of each end user. Finally, if you have large quantities of data to annotate, try to find transparency about worker conditions, especially with RLHF (reinforcement learning from human feedback) outsourcing.

‍SS: I have long gravitated toward smaller, bespoke approaches to AI, inspired by the work of Stephanie Dinkins. My ongoing research engagement with Bina48 began with a highly qualitative approach to curating datasets that I chose to call “mentorship” rather than “training.” I try to resist the impulse to generate heedlessly even when I’m experimenting with a new platform, but it can be very addictive. 

I hate the thought of “wasting” outputs or regarding them as disposable, and I recycle or repurpose much of what I create with AI. 

For instance, my Technelegy 2.0 project uses variations or iterations of synthetic poems generated during many AI writing sessions as training data for a new edition of my first collection that continuously writes itself. It’s eye-opening and infuriating how full of junk they are; how unintentional they mostly are; and how wasteful of energy — human, computer, planetary — they can be. But, in a way, that’s what makes them valuable as a lens through which to investigate ourselves. I see the role of artists using AI as helping to strike a meaningful, thoughtful balance between vastness and vision, as Holly Herndon, Mat Dryhurst, and the Spawning team are doing.

‍DD: How can artists concretely measure their environmental footprint?‍

AA: There is an excellent and comprehensive artist toolkit recently published by the Gallery Climate Coalition (GCC) that looks at many aspects of artists’ carbon impact, including research, production, materials, and travel. It also explores the wider impact you can have as an artist, for example with advocacy and campaigning, and how you can measure that in comparison to the carbon emissions of your other activities. 

GCC also has a carbon calculator for artists and galleries that provides an easy-to-use guide to assessing your carbon emissions and creating your own report. The area of digital emissions for artists is less well developed, but we have added a section on that to the latest carbon calculator. A key area for most artists will be buying new devices. The calculator also has a session assessing your use of cloud services such as data storage, and further estimators for video meetings and email. 

‍DD: Despite the challenges raised by the environmental and social costs of AI, encouraging developments are emerging. What changes in the AI industry are most inspiring you right now?‍

EG: Techniques are emerging to make AI more frugal in consumption, aiming for the same efficiency with less data and fewer resources — like pruning models, quantifying them, and flash attention. The goal is lighter models in training and use. For example, Liquid Foundation Models consume far fewer resources by adapting dynamically with less computing power. BitEnergy AI’s L-Mul algorithm replaces complex multiplications with additions, cutting energy use by 95% in certain neural network operations. More specialized models also consume less energy by focusing on single tasks.

There’s [currently] a shift to non-carbon energies, with nuclear resurging in the US and China, while solar and geothermal energy are gaining ground in Kenya. Other infrastructure improvements include carbon-free concrete for data centers and better placement to reduce the need for cooling water. New chip generations, like photonic chips that use light instead of electricity, increase efficiency a thousandfold. Of course, we mustn’t forget regulatory initiatives. The EU’s AI Act requires sharing models’ energy consumption, while the US is discussing a bill on AI’s environmental impact, [which is] essential for transparency.

‍DD: AI tools often prioritize scalability and speed at the expense of sustainability and purposeful design. What would an ideal AI tool for sustainable art practices look like to you? ‍

SS: In a way, Technelegy is an ongoing attempt to prototype my dream AI tool — a generative writing system designed to augment imagination and empower multimodal authorship and readership without negative inputs or impacts. [This is] less a one-size-fits-all tool of mass production or mass creativity and more of an adaptable, efficient partner — a modular array of energy-aware algorithms that can work together or divvy up, switch seamlessly from task to task, from small to large models, from localized to broader systems, and optimize for efficiency on a case-by-case basis. 

This ideal tool should prioritize a user experience that incentivizes a sense of purpose. Environmental sustainability goes hand in hand with human behavior, using this technology in ways that are for our own human energy and resources, livelihoods, mental health, and our grasp on reality. I’m often reminded of that when I spend time in the small grassy meadow near my house, a place that often finds its way into my writing. It’s equally wild and cultivated, with paths mowed by my husband interweaving overgrown beauty. To me, that feels like a nice metaphor — respectfully carving out space in a vast, untamed landscape for human imagination to wander, roam freely, and discover itself anew.

My upcoming show with Nathaniel Stern, “Generation to Generation: Conversing with Kindred Technologies,” which opens at Krasl Art Center in April 2025, is all about the need for greater attention to and care for the planetary resources that power modern life, and orbits around a central poem called Mother Computer, with its echoes of Mother Earth.