The film-maker Bennett Miller’s exhibition of hauntingly allusive AI-generated images, with Gagosian, in Paris, is the product of his three-year engagement with a “primordial” early version of OpenAI’s Dall-E text-to-image model.

In 2021, the New York-based Miller interviewed chief executive of OpenAI Sam Altman for an for as yet unreleased documentary on artificial intelligence. He was subsequently offered access to Dall-E by Altman, becoming one of the first ten people outside the company to test what has become one of the transformational models in the industry since its public release in mid-2022.

For the documentary, Miller has spoken to thought-leaders, including Altman, the playwright Tom Stoppard, and the novelist Marilynne Robinson, to address big, existential questions around artificial intelligence. Not about the technology itself but about the human condition at a time of profound technological change.

Miller’s film output includes The Cruise (1998), Capote (2005), Moneyball (2011), and Foxcatcher (2014). He was nominated for Academy Awards for Capote and Foxcatcher and won Best Director Award at the Cannes Film Festival for Foxcatcher. After Foxcatcher, his attention turned to a wish, which he had been nurturing since 2011, to address the accelerating evolution of technological powers, and the accompanying changes to daily life that he found to be as profound as any of the changes wrought by the invention of the printing press or of photography.

He saw the project, he tells The Art Newspaper, as a chance to hit pause and “talk to all sorts of people, visionaries who are leading these fields, from bioengineering to AI. But also to speak to people from the humanities.

“The whole idea was not to promote a thesis. It was just to hit pause and ask the one-on-one questions of philosophy because as we are gaining the powers to do anything, the basic dorm-room type questions—like ‘What is a good life?’ and ‘What is a better world?’—seem to be the headline for me”.

“A wondrous and moving new medium”

Miller says that being invited by Altman to work with Dall-E introduced him to a “wondrous and moving” new medium, one that operates “on a different kind of plane”. His exhibition of sepia-tinted pigment prints is selected from over a hundred thousand images generated working with his beta version of Dall-E 2.

He went to great lengths, he says, to have final images that are “meant to be”, that sometimes “stopped time for me and helped me feel something which is very difficult to feel”—a perspective on “being alive in this moment of profound technological acceleration”.

“I’ve been deep in it for three years,” Miller says of his work with generative AI. “I’m still entranced by it. It’s a new medium, and that’s a hard thing to grasp what that actually means. it invites you to work in ways that were hitherto not possible. And necessarily, consequentially, creates new potentialities for expression.”

It is a medium, he says, that suits him “exquisitely well” for “temperament and aesthetics and myriad reasons. I was as happy that there was no instruction manual”.

Miller’s images have a dream-like, ambiguous and Surreal quality. They are closely cropped and selectively focused, like some heady marriage of Henry Fox Talbot, Man Ray and the faded family-album images printed on soft text paper in paperback editions of WG Sebald’s books Rings of Saturn (1995) or Austerlitz (2001). That quality depends in part on the unvarnished, pre-release, version of Dall-E 2 that he worked with.

That allusive but elusive quality exists in the framed images on Gagosian’s walls—is that a prototype space shuttle being wheeled out for launch? Or might it be the carcass of a beached whale? This is because Miller learnt along the way the requirement that he “give space” to his text-to-image model; the need to let the hallucinatory, big-dreaming, magic happen by knowing when to leave it to the AI, when to get out of the way.

Miller has learnt the same lesson that the London-based digital artist David Sheldrick uncovered while creating his AI-powered image and video series Empire (2025). Empire, launched by the digital art gallery Fellowship earlier this year, was created by running pairs of digital images—scanned from vintage colour film slides acquired on Ebay and other sites—through a specially trained version of the text-to-image AI model, Stable Diffusion.

Sheldrick started by choosing which slides to blend together—with very repetitive results, he said at Empire’s gallery launch in London. The magic happened, he said, when he allowed his computers to randomly select slides “combining things in an unbiased way which produces these really fascinating results to collage them together into one image”.

For Miller, when asked about Sheldrick’s analysis, this letting-go is essential when working with text-to-image models. “That’s what the medium is,” Miller says. “That’s the novelty. That’s the genius. That’s the whole point.”

If you “dial it too tight”, he says, so that the image output is “going to be as literal as it can be to your description, then you’re basically strangling the medium”. You would, he says, “be using it to do something that we’ve always done or are already doing”. Giving the AI model space is the “whole thing”, Miller says.

For Miller, generative AI “allows me to do something I’ve always wanted to do. Even if you couldn’t have described it before.” ”It’s so profound,” he says, “the extent of the control you can have,” Miller says, “while, very importantly, maintaining inconsistent and unpredictable outcomes.”

“It is like working on a film”

Miller says that the serendipity and the randomness of allowing the AI model “to do what it does” is not a case of “release the hounds and see what they come back with”. It is, he says, like a director’s role in prompting creativity in the making of a film.

“You want all of your collaborators, or actors, or cinematographer, costume designer, production designer, whoever you’re working with … to surprise you,” he says. “You want their potentialities. In the role of a director, you have to maintain the vision and know what the thing needs to add up to somehow—but you’re prompting people.”

Underpinning everything in Miller’s exhibition of fugitive, thought-provoking images, is his engagement with his unfinished documentary which, he says, is “not meant to change anything. It’s not advocating. It’s more like a slow, sober, genuine inquiry.” In which he asks thought-leaders: “Tell me what you’re doing, why you’re doing it. What’s the vision? What are these technological imperatives that we all feel?”

“Nobody is being asked these most fundamental questions,” Miller says, “there’s almost a stigma about talking about what actually matters to us.”

• Bennett Miller, Gagosian, 4 rue de Ponthieu, 75008 Paris, until 8 March