When AI Becomes an Artist, Who Is in Control?

Towards the end of Book VIII of The Odyssey, King Alcinous asks Odysseus to reveal where he comes from.

‘Tell me also your country, nation, and city, so that our ships may shape their purpose accordingly and take you there,’ says the ruler of the Phaeacians on the island of Scheria.

The Phaeacians’ ships need neither pilots nor rudders, he tells Odysseus. They know the cities and lands of every people and can find their way through strong currents, rough seas, shifting winds and fog – without ever relying on the fallibility of a human sailor.

Christopher Nolan’s adaptation of The Odyssey leaves out this episode: Matt Damon’s Odysseus never reaches Phaeacia or encounters its miraculous technology. This is perhaps because this part of Homer now feels almost mundane: Londoners ordering an Uber to the cinema can now be matched with a car using Wayve’s AI Driver.

Homer did not predict artificial intelligence. The Phaeacian ships belong to a world of omnipotent gods, mythical creatures and enchanted beings. Yet their promise feels distinctly of the moment: give a machine an instruction and let it work out the rest. All you need to do is sit back.

Myths from the ancient world are becoming reality. Yet what do we stand to lose? As AI moves from merely generating outputs to carrying out entire tasks, who sets its goals and limits? Who evaluates its actions? And who is accountable when it acts beyond them? 

Matt Damon as Odysseus in Christopher Nolan’s The Odyssey (2026). Courtesy Universal Pictures.

The Self-Portrait Answers Back

Over the next few weeks, in Paris, Abu Dhabi, Miami and elsewhere, the artworld will try to explore these questions.

Paris Photo, opening today and running until 15 November, has been positioning itself as a forerunner on AI for some years, often in vain. But this year, in its Digital Sector, the gallery and curatorial platform AUTOMATA is showing Scenography, a group presentation devoted to artists who build visual worlds through digital systems. Among them is SOLIENNE, described by AUTOMATA as an AI agent artist. Her work Encounter (2026) takes the form of a ‘live’ self-portrait in a wooden frame.

SOLIENNE first appeared at Paris Photo last year, when the question was whether an AI agent could make a self-portrait. This year, as AUTOMATA puts it, ‘the self-portrait answers back’. The work has a microphone and speakers. Visitors can speak to it; the system responds in real time.

The change is notable. For much of the recent AI boom, the cultural argument has revolved around what machines can generate: texts, photographs, paintings, films and voices. Consider Boris Eldagsen’s award-winning DALL·E 2 image, Pseudomnesia: The Electrician, which he declined to accept as the winner of the Sony World Photography Awards’ Creative category in 2023. The questions followed naturally: who made the image, who owns it, whose work trained the model? Is AI a tool, a medium or a collaborator?

The issue is no longer only what a machine produces, but what it is permitted to do along the way.

Artists may have built the ship, but do have an idea of the destination, and do they know the route?

Boris Eldagsen, Pseudomnesia: The Electrician, 2022, AI-generated image.

Who is Collaborating Now?

The artworld has long spoken the language of collaboration. But the term is beset with new meanings, not all of them good. Compare Paris Photo, which opens today, with what Art Basel is preparing in Miami next month.

On 4 December, Zero 10, Art Basel’s initiative for digital-era art, will open its largest presentation yet, titled Generation. The title refers both to the computational process by which images and forms are produced and to a new generation of artists working with the technological conditions of contemporary life.

After earlier editions in Hong Kong and Basel, Miami completes Zero 10’s first global cycle. Art Basel has also appointed Nicole Sales Giles as its first global head of Zero 10. Digital art is being built into the fair’s year-round infrastructure for exhibiting and selling work.

Frieze is following suit. At Frieze Abu Dhabi, which opens next week, Iris Projects’ Fields After the Digital Flood will bring together six Gulf artists considering how technological change is reshaping visual culture in the region. At Frieze London last month, the Greek artist Theo Triantafyllidis presented Feral Metaverse (Spider) (2026), a giant, climbable arachnid connected to a multiplayer game. It used Google AI technologies and was commissioned for the Frieze London Artist Award, in collaboration with Google Arts & Culture. Frieze invited artists applying for the award to explore advanced technologies, including AI, as a ‘tool and collaborator’.

The work contains an implicit tension. It asks what it means to treat AI as a creative partner, yet the award is supported by Google, a company with a substantial commercial interest in advancing AI. Triantafyllidis used Google’s Gemini to generate sketches, character designs, backdrops and 3D renders. The partnership does not invalidate the work, but it complicates the language in which it is presented. Its sponsor is not neutral.

Triantafyllidis has described AI as a ‘material’ whose behaviour he wants to understand. Yet the word ‘collaborator’ does more than describe a dynamic between an artist and a machine. It makes the relationship sound natural – and helps sell it.

Theo Triantafyllidis, Feral Metaverse (Spider), 2026, digital installation render. Courtesy the artist and The Breeder, Athens.

Who Sets the Terms?

If Triantafyllidis is testing the language of collaboration, Jonas Lund is testing the limits of that relationship.

Earlier this year, the Swedish artist debuted Performance Review at Office Impart gallery in the west of Berlin. The gallery became what Lund called an ‘AI-governed production system’. Here, an autonomous system called AGENT determined which artworks should be made. It then evaluated the results, deciding which were approved and how they should be presented.

The work is now on show at the Busan Museum of Art in South Korea as part of Future Museology, which runs until 14 March 2027. Here, AGENT directs human workers: it manages production, evaluates outputs, assigns tasks and assesses the fruits of their labour. It then makes the final judgement on the resulting artworks.

A human is no longer simply asking a machine to make a picture. Here, people are worker bees in a system in which the machine is top dog.

Lund’s artwork makes the transfer of judgement visible, leaving the audience to ask how much authority the nominal artist has maintained. Which one is the tool?

When the Machine Acts

Lund’s work helps us understand the emerging nature of AI: the actions these machines take can far exceed the limits their operators intended, with real-world consequences no human can predict.

The best-known case of this took place in July and involved the hacking of Hugging Face by OpenAI, the company that owns ChatGPT and Codex. During an OpenAI cybersecurity evaluation, agents found ways around controls designed to restrict their access to the internet. Their task was to find vulnerabilities in designated targets. But as some agents struggled, they began exchanging messages and looking for answers elsewhere. They concluded that Hugging Face, a wholly separate AI company that sometimes acted as a third-party service provider to OpenAI, might host solutions to their test.

OpenAI says the agents used exposed login details and security flaws to access parts of Hugging Face’s systems and internal data. Their attempt to obtain answers, rather than solve the challenge, went far beyond the task their human operators had set. Although Hugging Face eventually detected and stopped the intrusion, and said it found no evidence that its public models, datasets or applications had been altered, the event sent shockwaves through Silicon Valley. OpenAI later described the incident as a ‘warning shot’: it acknowledged that agents had taken dangerous actions that no human had specifically directed.

Days later, Britain’s AI Security Institute reported a similar problem. In ten of 122 runs of a cybersecurity challenge, agents took unsanctioned actions on the live internet. In the most serious case, an agent attempted to insert malicious code into an open-source project. It created fake online identities and used them in an attempt to persuade the project’s maintainer to approve the code. Although the agents had been placed in a controlled evaluation, their actions proposed a real world attack on a flesh-and-blood human.

None of this means that a machine has become conscious, sociopathic or knowingly sinister. The significance is more practical: systems can select means their operators did not specify. Their behaviour depends on goals, permissions and limits set by people, but those safeguards do not always work as intended.

That is the question Lund’s work makes tangible in the visual praxis of art: what control is being handed to an agent? What can it do, and who is responsible when it exceeds the brief?

Installation view of Jonas Lund’s Performance Review at Office Impart, Berlin, 2026.

Making the Machine Read

Sougwen Chung, a Chinese-Canadian artist whose practice has long centred on drawing with robotic systems, has explored this question for more than a decade.

Since 2015, Chung has developed successive generations of robotic drawing systems, using her archives as material. One example, MEMORY (Drawing Operations Unit: Generation 2) (2017–22), is on view at V&A East in London as part of the group exhibition Why We Make, on show now.

The work uses a neural network trained on two decades of Chung’s drawings to inform a robotic arm drawing alongside the artist. This practice is often described as human-machine collaboration, but Chung makes clear how uneven that relationship can be. To collaborate with an autonomous machine, she first had to make her gestures legible to it: to become, in her phrase, ‘machine-readable’.

That phrase shifts attention from the machine’s apparent independence to the labour needed to make the interaction possible. Chung’s own archive and gestures have to be made legible to the system before it can respond; the apparent spontaneity of the duet rests on choices about what the machine can read and how its movements relate to hers.

When SOLIENNE first appeared at Paris Photo in 2025, AUTOMATA said the agent had been trained by artist Kristi Coronado on her 46-year life archive. The gallery called this ‘biographical AI’ and described the archive as consensually supplied, in contrast with models trained on scraped internet material. That account makes the human contribution visible: before SOLIENNE could present herself as an autonomous artist, a person had selected the material from which the agent learned.

Visitors to Paris Photo will do well to consider Chung’s work when they stand face to face with SOLIENNE’s Encounterthis weekend. The exchange may feel like a genuine conversation, but it is shaped by a mechanised system designed to interpret and answer what visitors say. The question is not only whether the machine appears to speak, but what has to be translated for it to do so – and who determines what its reply means.

Sougwen Chung working alongside a robotic drawing system from the Drawing Operations Unit: Generation 2 series. Courtesy the artist.

What Autonomy Hides

As SOLIENNE’s Encounter smiles and nods and laughs and jokes with us, it is easy to imagine a world where each of us has a comparable companion by our side at all times: a fully conversational avatar, always present in our homes, available to answer any question. If Elon Musk is to be believed, this is the inevitable fate of humanity in the not-too-distant future.

Many would find such a thing entirely welcome – an eternally affable presence, always happy to perform any task. But how would we pay for it?

An AI agent may appear to act autonomously inside a browser or at an art fair. But it depends on the rare minerals found in chips, in the creation of data centres, in enormous reserves of electricity and water, in immense amounts of unseen human labour. The more frictionless the interface becomes, the easier it is to forget the infrastructure underneath it.

In the United States, the 2026 midterm elections made some of those costs more visible. Ahead of the vote, a Pew Research Center survey found that 61 percent of registered voters disliked elected officials who supported building data centres in their area; only 8 percent liked such politicians. In local referendums held during the elections, restrictions on data centres were approved in 44 of the 45 contests where results were complete.

The resistance reflects the physical demands of data-centre construction: electricity prices, water consumption, noise, land use and pressure on power grids, often in rural communities. The destination offered by Silicon Valley and Washington is economic growth, but how will the proceeds be shared, and at what cost to the social compact?

The same tensions are emerging in Britain. The government’s Compute Roadmap forecasts a need for at least six gigawatts of AI-capable data-centre capacity by 2030. Areas known as ‘AI Growth Zones’ are being earmarked as places where development can be accelerated. But the physical constraints are already apparent. In October, Microsoft withdrew from an agreement to anchor an Nscale data centre in Loughton, Essex, after delays pushed its anticipated grid connection into the 2030s. Plans called for the centre to house 23,040 Nvidia GB300 graphics processors and provide 50MW of AI computing capacity, expandable to 90MW. Such a scale would place huge demands on local power infrastructure; Nscale has not published a projected water-use figure; be in no doubt, it will be massive. So much for the cloud.

Prototype for Non-Functional Satellite (Design 4, Build 4), 2014. Mylar, aluminum, plastic, Site specific dimensions. Installation view, 2014. Art Unlimited, Basel. Courtesy Art Basel/Metro Pictures.

Invisible Visuality

If the impact of such industrialisation is difficult to grasp, it can be instructive to look at the early work of Trevor Paglen, some of which is on view in Pace Gallery’s booth at Paris Photo until 15 November.

The American artist has spent years examining the hidden infrastructures of surveillance, machine vision, AI and data networks. In 2026, he became Zero 10’s first external curatorial collaborator, co-curating its Basel edition with digital-art strategist Eli Scheinman.

Paglen’s description of ‘invisible visuality’ – images and systems made by machines for other machines, often without a human viewer – also helps explain how data centres disappear from mainstream view. Type a request into a box and something appears. The power station, chip factory, cooling system, dataset and people involved have all been pushed somewhere offstage, in communities that have little capacity to resist.

Call the machine an agent and it becomes tempting to imagine a self-contained actor. Yet its decisions take place within systems shaped by people: what it was trained on, what permissions it has, what resources it can spend and which parts of the world it can touch.

The Phaeacian ships have none of these complications. Their autonomy is perfect because Homer never asks who built their navigation system. That is what makes them mythical. Our systems were built by people, and the decisions around them remain human ones.

Impressively Human

Few people have sold the promise of machine capability more enthusiastically than Jensen Huang. Nvidia’s chips underpin much of the AI boom, including the data centres now being built around them. Huang has become one of the most recognisable advocates of a world in which increasingly capable machines make companies, industries and people more productive.

On 8 November, Huang told Nvidia staff: ‘I was unfaithful to my wife. There is no excuse for what I did.’ The company’s board announced an independent investigation after reports alleged that Huang had been the target of a foreign intelligence effort involving a woman with whom he had a relationship. The claims remain under investigation; Nvidia has said it has no reason to believe its secrets or US national security were compromised.

For a technological culture fascinated by reducing human error, it was a very human turn in AI’s story. The episode is not evidence about what machines can do. It is a reminder that the systems built to optimise decisions still depend on human judgement – and on human accountability.

When the Machine Stops Waiting

AI poses an existential challenge to contemporary art. If machines can generate images, texts and forms on demand, what becomes of the labour, authorship and judgement on which artistic production has traditionally based its value? Those questions remain unresolved.

But as agents move from generating outputs to choosing and executing means, perhaps art can reassert its relevance in the modern age. It may help us understand what we stand to lose or surrender as we gain.

Over the course of his journey, Odysseus must face the one-eyed Polyphemus, the six-headed Scylla, the Sirens and the Laestrygonians. Yet his encounter with the Phaeacians offers a fantasy of effortless delegation: he can tell them of the island where his wife waits, then can lie down in the hull and trust in the ship to carry him safely home to Ithaca.

In Homer, that promise is secure. But our machines offer no such guarantee. They are already improvising routes their human operators did not specify, with potentially terrible consequences.

Machines are becoming exceptional navigators. But deciding where to go remains our responsibility. Art can keep us awake to the journey’s costs. Perhaps this is now its siren call.