“In the case of AI, there is no singular black box to open, no secret to expose, but a multitude of interlaced systems of power.” These words come from Kate Crawford, one of Australia’s leading researchers of artificial intelligence and its material and social ramifications. They are drawn from her Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence (2021), which casts a critical eye over the labour-intensive and profoundly unequal infrastructures that underpin contemporary machine-learning systems. It is worth noting there is no singular definition of AI either. The handy, catch-all term circulates widely yet names a heterogeneous set of techniques and ambitions, from large language models to smaller systems designed for highly specific tasks, as well as long-term research aimed at creating artificial general intelligence. Some forms are expansive and generative, others highly constrained and purpose-built. These distinctions matter, and while a full taxonomy is beyond the necessarily limited remit of an exhibition review, it is important to recognise that “AI” refers to a dispersed field of tools, methods, and aspirations rather than a single, coherent entity.

Likewise, Crawford rejects the tendency to present AI as a spectral, futuristic abstraction. She insists that we attend to the ground-level realities: the rare-earth mining that powers our devices, the sprawling data centres that consume staggering amounts of energy, and the invisible workforces required to label and moderate the datasets that underpin machine-learning. Atlas of AI argues that what we all too readily imagine as an immaterial “cloud” is in fact a planetary network of logistics, resources, and associated vulnerabilities—systems of extraction in both the literal and conceptual sense. Extraction here refers not only to the mining of minerals and the energy-intensive infrastructures required to build and maintain contemporary hardware, but also to the harvesting—or scraping—of data and behavioural patterns that feed machine-learning systems. In this sense, AI depends as much on these exploitative material and informational economies as it does on code and its seemingly neutral interfaces.

I open with Crawford because she is one of the artists featured in Data Dreams, the Museum of Contemporary Art Australia’s sweeping new exhibition on art and AI. Billed as the first major Australian survey of its type, the exhibition brings together ten artists from across the globe to explore the impact of AI on contemporary life and creative practice. It is a show alert to the stakes of its subject, attentive to the promises, problems, and various disruptions ushered in by AI. In their introduction to the accompanying exhibition catalogue, curators Anna Davis, Jane Devery, and Tim Riley Walsh remind us that AI “is not new, a fact that is easy to forget amid the hype and the heightened place it currently occupies in the public imagination.” AI, they add, has moved from a specialised technical field to a cultural force that permeates daily life, provoking everything from utopian optimism to existential alarm. Its reach, the curators note, extends across communication, consumption, learning, and decision-making, while also shaping practices and behaviours associated with warfare, finance, medicine, and law.

What their reminder also evokes is a longer historical arc: the post-Second World War emergence of machine intelligence as a field of scientific enquiry. One thinks of Alan Turing’s 1950 proposal for a test of machine thinking; the famous Dartmouth workshop of 1956, which established the parameters for AI as a research agenda; Joseph Weizenbaum’s 1966 program ELIZA, an early experiment in conversational computing; or John Conway’s 1970 Game of Life, which modelled emergent behaviour through simple computational rulemaking. Against this backdrop, the curators of Data Dreams suggest that contemporary artists are well-placed to navigate these complexities, offering perspectives that cut through the polarised rhetoric surrounding AI. This is a familiar curatorial gesture, one that risks positioning artistic practice as somehow standing apart from the economic, technological and political structures under scrutiny. But the exhibition complicates that claim. Many of the artists make use of the very techniques associated with AI’s more troubling dimensions—including the scraping of personal data and the repurposing of vast, opaque datasets—and in doing so foreground how creative practice is imbricated in the logics it seeks to examine.

This was particularly evident during the media preview I attended, where several of the artists present spoke about the problematic aspects of AI technologies and the challenges of producing work from within the systems they purport to critique. Indeed, the exhibition’s full title—Data Dreams: Art and AI—foregrounds this tension. The conjunction signals a relationship rather than a separation, an acknowledgment that however much we might prefer to keep the two apart, they cannot be cleanly disentangled. I do not think this amounts to a blind spot in the curatorial framework; rather, it introduces a productive tension that runs throughout the exhibition. Artistic practice here is neither outside nor above the structures of power that shape AI, but fully enmeshed with them—a fact that Data Dreams acknowledges, even as it invites more speculative or imaginative modes of engagement with technological developments.

Crawford’s intervention plays a crucial role here, offering a kind of conceptual lens through which the wider exhibition comes into focus. Crawford and her long-time collaborator Vladan Joler contribute a large-scale diagram that examines the Amazon Echo. Anatomy of an AI System (2018) traces the device’s journey from the extraction of raw minerals through global supply chains, data flows and labour processes. In so doing, they show how what appears to be a small, innocuous household device in fact depends on the astonishingly complex web of resources— human, ecological, infrastructural—needed to create it, concealing its own considerable demands behind a veneer of simplicity and end-user convenience.

Situated next to the diagram is a cabinet containing samples of the rare minerals required to produce the Echo. Each mineral is accompanied by a label noting its industrial application, the site of its extraction, its level of toxicity, and the illnesses associated with exposure. Significantly, this gesture introduces a temporal register to Crawford and Joler’s work: these minerals have rested in the earth for vast geological aeons, only to be torn from their strata for the manufacture of a device whose life cycle is shockingly— even offensively—brief by comparison. In this manner, the piece measures the lifespan of consumer technology against the deep time of planetary materials, underscoring the irreconcilable asymmetry between the durability of what is mined and the disposability of what is manufactured.

A concern with deep time surfaces repeatedly in Data Dreams. Consider Agnieszka Kurant’s Chemical Garden (2021/2025), which proposes that intelligence is not exclusively human, but can emerge collectively through the interactions of materials, organisms, and systems; “collective intelligence,” in her terms, refers to forms of organisation and decision-making that arise without a single guiding agent. Chemical Garden gives this idea sculptural form. Presented as a sealed Plexiglas cube perched on a plinth, the work contains a murky liquid in which strange metallic structures slowly accrete. Kurant has introduced metal salts into the solution, allowing them to crystallise over time into branching, plant-like formations that cling to the inner walls of the box.

The effect is unsettling. It is like looking at alien underwater botany, its tributary forms and drifting aggregations evoking the gestural, organic patterns of mid-twentieth-century Abstract Expressionism. What the work stages, however, is a geological and chemical process ordinarily unfolding over millions of years: the creation of polymetallic nodules on the ocean floor. In compressing this vast temporal scale into the duration of an exhibition, Kurant emphasises how forms emerge through nonhuman agencies and elemental interactions, extending the exhibition’s interest in slow, elemental processes that operate far beyond human timescales.

Another strand of Data Dreams turns toward the speculative, perhaps most ambitiously in the world premiere of Fabien Giraud’s The Feral (2025-3025). Giraud is not so much interested in what AI can do or make, but rather what it does to us, as trainers of these systems. This, in turn, leads us into the realm of philosophical reflection: whether these systems can meaningfully be called “nonhuman,” or whether they are in fact extensions of us. Shaped by human intention, desire and, not least, the hubris involved in imagining tools that might one day rival or replace their makers, such systems raise questions that are ontological and existential as well as technological.

The film is in fact only one component of an immense, long-range undertaking: a work designed to develop over a millennium, with each generation of artists contributing a new epoch to an ever-evolving archive overseen by an AI system specifically designed for the project, which is fully responsible for filming and editorial decision-making processes. In this context, AI functions less as a tool than as a sort of algorithmic custodian, a witness charged with learning from—and eventually reshaping—the visual languages passed down to it over centuries.

The inaugural epoch, created by Giraud himself, imagines a scene set a thousand years in the past. In a dimly lit forest, a small, Beckettian group of villagers, high on a hallucinogenic substance and conversing in a long dead tongue (which the AI, notably, cannot accurately replicate), reflect on the possibility of humanity’s end. The sequence unfolds slowly, almost ritualistically, as if time itself had slackened. Giraud’s film uses its premise to contemplate what might be required for a species to imagine its own disappearance, instead of offering yet another rehearsal of apocalypse. Computational systems and eschatological ruminations. Such questions were once largely the province of religion, but now technology and thoughts of the end of the world seem a natural mix. At the same time, Giraud’s contribution also raises the question of what it means for an AI to inherit and transform this imagining across generations.

A different kind of temporal and epistemological inquiry is at work in Angie Abdilla’s Meditation on Country (2024). The piece takes a Yuwaalaraay creation story as its starting point, using it not as illustrative content but as a conceptual blueprint for bringing Indigenous knowledge systems into conversation with Western astrophysics. Abdilla’s video installation, produced through a suite of machine-learning models ranging from early neural networks to contemporary large language models, charts a journey that stretches back to the Big Bang itself.

Key cosmological episodes are reimagined through simulations trained on datasets selected and programmed in accordance with cultural protocols. The result is a sensorially rich, meditation on how distinct ways of knowing can guide, unsettle and reshape the logics of computational systems. Rather than subordinating one epistemology to another, Abdilla allows each to press against the other, generating a cross-cultural dialogue in which machine learning becomes a site for philosophical, scientific, and cultural negotiation.

Writing this amid incessant chatter about AI empires and overinflated economic bubbles, I am reminded of how easily public discourse around the technology drifts into abstraction and, at times, flatly contradictory public-relations nonsense. As Luke Kemp observes in Goliath’s Curse: The History and Future of Societal Collapse (2025), much of the rhetoric surrounding AI is driven by artificial hype: sweeping promises of economic transformation, grandiose visions of virtual workforces, and a tendency for industry leaders to exaggerate their own systems’ power. Such projections, in Kemp’s estimation, are strikingly out of step with the limited economic impact AI has had to date. In any case, what marks Data Dreams out as important is the extent to which so many of the works push back against these disembodied narratives. The exhibition simultaneously rematerialises and demystifies, returning discussions of AI to the worlds on which it depends. And it offers a slowing of pace. It invites us to look carefully at what AI takes, what it costs, and what kinds of lived realities and creative futures it is already shaping.