{"id":5249,"date":"2025-09-29T12:48:17","date_gmt":"2025-09-29T10:48:17","guid":{"rendered":"https:\/\/hacnumedia.org\/creating-with-ai-ecological-alternatives-do-exist\/"},"modified":"2025-12-09T18:34:33","modified_gmt":"2025-12-09T17:34:33","slug":"creating-with-ai-ecological-alternatives-do-exist","status":"publish","type":"post","link":"https:\/\/hacnumedia.org\/en\/creating-with-ai-ecological-alternatives-do-exist\/","title":{"rendered":"Creating with AI: ecological alternatives do exist"},"content":{"rendered":"\n

Tiny models, solar-powered models, residencies exploring AI and ecology\u2026 While most artists today rely on conventional Big Tech models, another segment of contemporary creation is exploring more frugal and critical approaches. Less energy-intensive\u2014sometimes at the cost of a slower creative pace\u2014these practices shift our imagination and challenge dominant representations. This article analyses these alternatives, showcasing emerging models and highlighting significant artistic contributions. <\/strong><\/p>\n\n

In the background lies generative AI and its large language models (ChatGPT, Claude, Mistral\u2026). Their environmental cost\u2014now better documented <\/a>despite the lack of transparency<\/a> from tech companies\u2014is immense: staggering energy consumption tied to training, colossal infrastructures, and the continual production of increasingly powerful hardware. This capitalist and extractivist system is powerfully illustrated by artist-researcher Kate Crawford and artist Vladan Joler in their now famous Anatomy of an AI System<\/a>. In this context, AI-assisted creation can quickly feel guilt-inducing, even at odds with the ecological and social values upheld by many contemporary artists. Let us be clear: \u201cfrugal AI\u201d is likely more of an oxymoron than a concrete reality. Yet AI creation is not condemned to rely solely on the models powered by tech giants. There are as many ways of working as there are artists, and some deliberately choose more restrained, more ecological paths. In a world of \u201cfor or against\u201d\u2014AI included\u2014where viewpoints tend to polarize, there remains space for alternative practices. Hence the need to make these dissenting initiatives visible. <\/p>\n\n

<\/p>\n\n

\"\"
Anatomy of an AI System – Kate Crawford, Vladan Joler<\/figcaption><\/figure>\n\n

<\/p>\n\n

Creating on the Margins of LLMs<\/strong><\/h2>\n\n

Very schematically, the functioning of AI rests on three pillars: a model (with a greater or lesser number of parameters), a dataset, and the computing power required for training. The best-known LLMs and image or video generators (Sora, Midjourney\u2026) rely on billions of parameters, draw on massive quantities of web-sourced data, and require colossal computational power, made possible by GPUs\u2014graphics cards over which NVIDIA now holds a near-global monopoly.<\/a> By nature, these models are therefore highly energy-intensive and difficult to reconcile with ecological concerns: climate, biodiversity, water cycles\u2026 And the issue begins, first of all, on the technical level. \u201cIt\u2019s the number of parameters that remains decisive, because the relation is fairly linear: fewer parameters mean faster inference. The architecture of the model directly influences execution speed, memory usage, and energy consumption,\u201d explains Charles Bicari, multimedia integrator at the SAT <\/a>in Montreal, who has spent several months working on AI programs applied to artistic creation. <\/p>\n\n

Another defining feature: their generalist vocation. These generative AI models can accomplish a multitude of tasks without being limited to a specific use. But in the field of artistic creation, needs differ substantially. \u201cIs \u2018the bigger, the better\u2019 really a reliable model for creation?\u201d asks Baptiste Caramiaux,<\/a> researcher at the Institute for Intelligent Systems and Robotics (ISIR) at Sorbonne University in Paris. \u201cIt is entirely possible to avoid multitask models and instead use small models. Their advantage is precisely that they are oriented toward a specific task. Artists can calibrate and personalize them, and they become a kind of instrument.\u201d Historically, many artists did not wait for the 2022 AI boom to explore machine-learning models. For decades, some have experimented with them\u2014for instance at IRCAM<\/a>, through research on gesture recognition and sound interaction. \u201cThis is the case for Marco Donnarumma<\/a> or Laetitia Sonami<\/a>,\u201d notes Baptiste Caramiaux. \u201cFor others, like Atau Tanaka,<\/a> it is even a way of performing, since the model is trained live.\u201d These small, decentralized models also offer artists direct control over their datasets, which are relatively lightweight compared to LLM-based workflows (see our article \u201cThe Effects of the Dataset<\/a>\u201d). Their energy consumption remains modest, and they can be trained on local servers rather than in the cloud. A paradigm that runs counter to the technomaximalism of AI giants. <\/p>\n\n

The Tiny Models Alternative<\/strong><\/h2>\n\n

In this context, several actors in the digital arts ecosystem are experimenting with new models, with the ambition of making them available to the artistic community in the medium term. This is the case of the MUTEK AI Ecologies Lab<\/a>, which launched in 2025 a residency call on the theme of AI and ecology. \u201cWe had many conversations with artists eager to work with AI without compromising their values or ethical commitments. The idea was less to create artworks than to develop tools that cultural and creative industries could appropriate, and release them as open source,\u201d explains Sarah MacKenzie, director of MUTEK Forum. Six projects were supported with technical assistance from the SAT for implementing small models running on microprocessors (such as Raspberry Pis). Among these projects were You, Me, the Lichen & Spore, an immersive installation using AI to interpret biological data from organisms sensitive to environmental variations, and Environment Art in The Garden, which explores the generation of immersive 3D landscapes for use in animated cinema. Presented as proof-of-concepts during the latest edition of the festival, these works form a first step. \u201cIt was a starting point; we need to continue every year, with the aim of also measuring the concrete ecological impact of these projects,\u201d adds Sarah MacKenzie (a carbon calculator<\/a> for this type of model has been prototyped by Harvard University). <\/p>\n\n

<\/p>\n\n

\"\"
Mutek AI Ecologies Lab<\/figcaption><\/figure>\n\n

<\/p>\n\n

For Charles Bicari, who oversaw the technical development of the projects, the task remains demanding: \u201cThese are complex implementations, and no out-of-the-box solution exists yet. Despite this, we are committed to making a dozen small models accessible through Ossia Score<\/a> (install the software and download the appropriate models).\u201d More broadly, the SAT now stands out in the ecosystem through its clear intention to make this research visible and understandable, notably through pedagogical publications<\/a> dedicated to Tiny Models (also known as Machine Learning on Microcontrollers or TinyML), already documented in scientific literature<\/a> in recent months.<\/p>\n\n

An Experimental Solar-Powered Model<\/strong><\/h2>\n\n

Another notable initiative for its ambition is The Feral,<\/a><\/em> a project set to be launched in France in 2026 following the recent acquisition of 15 hectares of forest in the Limousin region. Here, the aim is not only to master tools, but to rethink the very conditions of our humanity in light of technological transformation. The Feral is grounded in a living ecosystem and fully embraces a contemporary paradox: \u201cThe more our technologies develop, the more the Earth deteriorates. We want to think about this paradox from within, without trying to solve it but instead making it the material of a practice,\u201d says its director, Ida Soulard. Artist Gr\u00e9gory Chatonsky<\/a> is associated with the project, with the goal of designing closed-circuit models powered by solar energy. Unlike the infrastructures of technological giants\u2014designed to handle millions of simultaneous requests\u2014the AI at The Feral will be conceived for a single site, offline, thus lowering consumption. It will not seek the supposed omniscience of a generalist AI but will adopt a limited form of learning, adapted specifically to its environment, exploring alternative approaches such as JEPA <\/a>or video-based learning. Finally, it will be powered by solar energy. The AI will therefore follow cycles of wakefulness and sleep, like a living organism, challenging our dependence on constant availability. \u201cThis approach leads us to a question: how do our uses evolve when technology adopts circadian rhythms? When it sleeps at night and wakes with the sun?\u201d continues Chatonsky. \u201cIt is a radical rethinking of our relationship to digital tools, synchronizing the artificial with the natural, replacing permanent availability with intermittence.\u201d <\/p>\n\n

A Complex Implementation <\/strong><\/h2>\n\n

These alternative models, however, remain difficult to develop. The Choral Data Trust <\/a>Experiment by artists Holly Herndon and Mat Dryhurst is a case in point. Since 2024, Serpentine <\/a>and Future Art Ecosystems<\/a>\u2014an organization aiming to build infrastructure for the development of art and technologies\u2014have supported the project. Its objective: to test new approaches to governance of AI training data, a peripheral topic but one intimately connected to ecological concerns. A publication <\/a>detailing the results of the experiment emphasizes the central role played by cultural institutions in developing alternative models, the need for significant resources, and the necessity of technical expertise. In other words, without political strategy and structural support, it is difficult to imagine alternative models being widely adopted by artists (see our article \u201cThe Real Cost of Generative AI<\/a>\u201d). <\/p>\n\n

Another obstacle lies in the technical implementation of these open-source models. In a previous article (\u201cAI, Proprietary Software or Open Source<\/a>\u201d), we explained that, unlike proprietary software\u2014with its ergonomic interfaces and marketing power\u2014open-source models, despite available resources (tutorials, libraries, active communities), often remain more austere, more technical to configure, and more confidential. Even more so when the models concerned are those recognized for their environmental attributes. \u201cIf you\u2019re not passionate about the topic, it is very difficult to access the right resources,\u201d notes artist L\u00e9a Collet,<\/a> who explores the interconnections between botany, living systems, and technology. \u201cI do a lot of workshops in schools, and it\u2019s often difficult to deploy an open-source AI on students\u2019 devices.\u201d Her project Racines<\/a>, presented at Le BAL, was\u2014out of pragmatism this time\u2014created using proprietary software. <\/p>\n\n

Exploring Other Modes of Production <\/strong><\/h2>\n\n

L\u00e9a Collet\u2019s approach deserves close attention. \u201cI spend a year making a project, finding a collective dynamic,\u201d she says. Her work Digitalis <\/a>emerged from workshops conducted at a middle school in Tourcoing. Over several months, students explored, through introductory workshops, the relationship between nature and technology. This was followed by scenario-writing sessions around a simple, poetic idea: imagining the students transforming into plants. The students then sent their questions to scientists, enriching the project with documentary and research-based content. The final stage involved working with open-source models deployed locally. The students\u2019 faces were filmed and then transformed into flowers. The process thus deconstructs two myths perpetuated by tech giants: the myth of individualized creative practice\u2014challenged here by the participation of an entire group\u2014and the myth of productivity, usually measured in terms of execution speed. In his latest work Deepscape: Longitudinal<\/a>, artist-researcher Hugo Scurto<\/a> worked with a generative AI model based on data gathered through his year-long observation of Cape Morgiou in the Calanques National Park in Marseille. He too embraces a creation process aligned with the rhythm of his subject. \u201cThe collection of sound, visual and personal data lasted an entire year,\u201d he explains. \u201cI had to work around environmental constraints\u2014whether the park was closed due to wind, or overcrowded with tourists.\u201d In total, the dataset took a year to gather, followed by another year to conceptualize and develop a form that felt coherent. <\/p>\n\n

<\/p>\n\n

\"\"
Digitalis, L\u00e9a Collet<\/figcaption><\/figure>\n\n

<\/p>\n\n

These examples of reappropriating the creative process are summarized by artist Justine Emard<\/a>: \u201cAs artists, we don\u2019t have control over everything, but we do have control over the ecosystem of the work.\u201d She illustrates this with her video-game-installation Chim[AI]ra<\/a> (2024). <\/em>In this project, one of the central data points is the activity of the computer\u2019s graphics card: the ecosystem represented in the game evolves according to the layers of rendered imagery. \u201cI wanted to establish correspondences between what physically happens in the world and what unfolds within the game. It is a way of returning to a kind of anatomy of the image\u2014understanding how a video game is composed.\u201d Thus, the virtual universe of the work, initially overflowing with visual effects, gradually becomes more sparse. As the GPU\u2019s consumption decreases\u2014visualized through an integrated graph\u2014the image is reduced to its essential structure, down to wireframes. <\/p>\n\n

<\/p>\n\n

\n