Refik Anadol's Unsupervised Made MoMA a Living Room in 2022

Anadol and his studio trained a machine learning model on roughly 138,000 works from MoMA's open-access digital collection, then let that model generate a continuous, never-repeating visual stream. The work wasn't looped. It wasn't edited. The imagery you saw at 10am on a Thursday would never appear again in that exact form. That's the part most early press coverage glossed over, focusing instead on how beautiful it was, which, yes, it was. But the real engineering achievement was unpredictability at scale, in a public-facing context, with no curatorial safety net catching mistakes.

From an event production standpoint, that's a significant risk posture. I've worked with generative visuals in brand activations before, smaller budgets, tighter spaces, and the single hardest conversation is always convincing a client to accept that the output won't be controllable. Brands want a mood board. They want to pre-approve the palette. Anadol, working within one of the most institutionally conservative art spaces in the world, got MoMA to say yes to genuine uncertainty. That's worth paying attention to.

The hardware itself was a custom LED installation running at a resolution most commercial installs don't attempt in spaces this public. Anadol's studio has been transparent in interviews about using NVIDIA GPU infrastructure for the training pipeline. The visual rendering drew on diffusion-model techniques that, in 2022, were still being written about in research papers rather than event briefs. Stable Diffusion's foundational paper had only appeared in late 2021. The speed at which Anadol's team moved from that research context to a finished public installation was, in the timeline of this field, very fast.

Here's the thing... institutions don't usually get viral lobby moments. MoMA is not a venue known for spontaneous crowd energy. But throughout the run of Unsupervised, the lobby became a gathering space in a way that the museum's paid exhibitions rarely produce. People would arrive, find a spot on the floor, and stay. Not browsing, not documenting for content, just watching. A colleague of mine, the kind of person who usually sprints through museum lobbies on the way to something specific, texted me a photo from the floor at 6pm on a weekday. No caption. Just the image.

That behavior, the seated contemplative audience in a circulation space, is something event designers spend enormous effort trying to engineer. You adjust sightlines, you lower ambient sound, you fiddle with seating placement and dwell-time metrics. Unsupervised produced it without any of that infrastructure, in a lobby designed primarily for throughput. Event Marketer's benchmark data consistently shows that average dwell time in activation spaces hovers between 90 seconds and three minutes. Anecdotally, from multiple accounts I collected from contacts who visited, dwell time in front of the Anadol piece ran closer to eight to twelve minutes per visitor. That's not a small gap.

Some of this is obviously attributable to the quality of the work itself. But part of it is structural. The piece asked nothing of the viewer. No app to download, no headset to adjust, no instructions to read. You arrived and the room did the work. In a moment when the experiential industry was arguably over-indexing on participation mechanics and gamification, that restraint landed harder than almost anything else that year. We wrote about this shift in our earlier piece on how Unsupervised reshaped the language of experiential events, but seeing the audience data since has only deepened my conviction about it. For more on this, see The Year the Room Became the Message: How Refik Anadol's Unsupervised Reshaped Experiential Events in 2022.

Okay, I have to be honest here, and slightly embarrassing about it. When I first heard the pitch summary for Unsupervised in early 2022, before I'd seen any visuals, my instinct was dismissive. "AI-generated art in a museum lobby" read to me like the kind of tech-art crossover that usually produces something that looks impressive in a press release and underwhelming in person. I'd been to at least three activations that year that promised generative visuals and delivered what I privately called "fancy screensaver energy." I even said that to someone at a production briefing. I was wrong. Loudly wrong, as it turned out, and I've had to recalibrate my skepticism since.

The broader industry was similarly uncertain. Throughout 2022, conversations in the event design community about AI tools were split between genuine excitement and a kind of defensive suspicion. Would generative visuals replace the creative director? Would clients start demanding "just make it AI" as a budget shortcut? I'm not 100% sure those fears were entirely unfounded, actually, because some of that did happen. But Unsupervised demonstrated something more specific: that AI as a medium, rather than AI as a production shortcut, produces an entirely different quality of audience experience. The distinction matters more than people were willing to admit at the time.

The common misconception, which I still hear in pitches, is that generative work means less authorship, less intention, less craft. Anadol's piece demolished that. The training data selection, the model architecture decisions, the physical installation choices, the scale, the placement, every one of those was a curatorial act. The machine didn't design the room. A person did, using a machine as a collaborator. For event curators watching from the sidelines, that reframing was genuinely useful. It gave us language for a conversation we'd been having badly.

By December 2022, Unsupervised had become a reference point in almost every serious brief I received. Not always by name, but by description: "something that feels alive," "responds to the space," "doesn't feel like a screen." The piece had created a new benchmark language for clients who hadn't previously had vocabulary for what they wanted. That's a real shift. Before a work like this lands, clients describe desired atmospheres in terms of other events they've attended. After it, they describe desired conditions. That's a more interesting conversation to be in.

The question it opened up for this field, and one we're still answering, is about the relationship between institutional and commercial experiential spaces. MoMA could take the risk of genuine unpredictability because its reputation doesn't depend on brand consistency in the way a product launch does. A perfume house launching a fragrance through a generative-AI-driven sensory installation faces a different accountability structure. I've worked on projects where the creative ambition clearly came from seeing work like Anadol's, and where the final execution got pulled back at the client approval stage because "what if it generates something off-brand." That tension is real, and I don't think it's fully resolved yet.

What Unsupervised did for the field's ambition, though, is harder to overstate. Alongside the kind of institutional scale that Superblue Miami had been testing since 2021, and the later spatial intelligence that teamLab's Borderless would push further in 2024, Anadol's MoMA piece established that art institutions were now setting the pace for experiential design, not following it. That reversal of influence matters to anyone building events today. The most interesting work was happening in museum lobbies, not at brand activations. Some of us found that humbling. Some of us found it liberating. I felt both, sometimes in the same afternoon.

A Pew Research report from November 2022 noted that younger audiences were increasingly seeking experiences that felt "unscripted and genuine" over polished, produced ones. Unsupervised was both technically sophisticated and genuinely unscripted in its moment-to-moment output. That combination is rare. It's also, I'd argue, the only combination worth chasing right now. For more on this, see The Room That Remembered You: How teamLab's 2024 Borderless Reopening Reshaped Experiential Events.

Sources

• MoMA Open Access Collection
• Stable Diffusion Research Paper, arXiv 2021
• Event Marketer Benchmarks
• Pew Research: How Teens Navigate Screen Time, November 2022