Want to see the future of AI art? Go grab some scissors
Danish artist Andreas Refsgaard has been combining generative AI with handcrafted prototypes to create unique glimpses of what’s ahead—a future that could one day make artists like him obsolete.
What if instead of asking AI to generate a picture, you built a cardboard model of what you wanted it to depict?
That’s exactly what Refsgaard has been exploring with his Future Mirror project, which invites students to use recycled materials to build prototypes of artifacts from the future. The Future Mirror web app then feeds these handmade objects into Stable Diffusion, producing unexpected visions of worlds we have yet to invent.
“I thought it was kind of magical to have kids build something very scrappy [with] toilet paper rolls and papier-mâché,” Refsgaard says. “Have them imagine wild stories, which they are good at, and then actually see that thing come to life.”
Since launching workshops with school classes in 2023, Refsgaard has seen kids create prototypes for sustainable buildings, autonomous food delivery vehicles, and teddy bears designed for all-ages space travel.
“One strength of this approach is that language has many barriers. When I have an idea, or I have something in mind, words can [often] not describe it,” Refsgaard explains. Building a model with your own hands allows for more detail, which then influences the final AI-generated image. “That’s really, really powerful,” he says.
A store for AI-generated books
Refsgaard has spent about a decade experimenting with machine learning and AI as an artist, often exploring the blurry boundaries between artist and algorithm, authorship and agency. An early example was Booksby.ai, an online bookstore selling paperbacks written entirely by AI.
Built in 2018, Booksby.ai came before today’s powerful large language models. Instead, it used a much weaker text generator that could create realistic-sounding sentences but struggled with coherent plots. “In terms of generative AI, this is medieval,” Refsgaard says, noting that the books were barely legible.
“I don’t think anyone has read an entire book from start to finish,” he says, adding, “They weren’t very interesting books.” Even so, the store sold about 300 copies, mostly through Amazon. Some books even received five-star reviews—one recommending a book because “it’s knodung,” a piece of gibberish generated by AI.
Refsgaard enjoys these moments when AI stumbles, though he knows they are becoming more rare. “Let’s laugh about how bad it is right now, but also think about the future consequences. Because it will be good at some point,” he says.
That moment is coming quickly. Refsgaard says he wouldn’t launch something like Booksby.ai today. “Why would I put AI-generated books onto Amazon [now]? It’s swamped by AI-generated books already.”
The fine line between art and a tech demo
Another recent project brought reality a little too close to Refsgaard’s art. As part of a series of experiments with image-to-text models, he built an online meme generator called MemeCam. The web app lets users snap a photo of an everyday object, then uses AI to turn it into a meme.
MemeCam was meant as a playful exploration but it quickly went viral. “It became extremely popular, my biggest hit I guess,” Refsgaard says, noting that his success came at a price. “I lost quite a lot of money on it.”
Although he’s not personally a big fan of MemeCam’s humor, calling it “a bit middle of the road,” Refsgaard still appreciates its impact. “It’s sort of like you’re indie, and you make a record that gets airplay on big stations,” he says. “You’re okay with it.”
The project also made him reflect on what it means to create art in today’s fast-evolving AI world. Sometimes even he isn’t sure how to define his own work. “Are they interesting art, or are they tech demos?”
When he started working with AI a decade ago, projects relied on basic tools and a lot of custom programming. Today, AI code generators handle much of that work, producing polished but sometimes less-interesting results. “I’m not going to make something new and unique in Midjourney,” Refsgaard says. “I don’t care how beautiful an image can become. But I’m interested in generating that image from cardboard.”
He adds that making AI art today is “way easier technically, but it’s more tricky conceptually.”
When AI art replaces the AI artist
Despite these shifts, Refsgaard still sees a place for artists like him, who approach AI playfully while also exploring its cracks and limitations. These experiments, he argues, help foster a more informed critique of AI. “If you just reject it completely, then the criticism you have of it is typically not very nuanced.”
As for his own work, Refsgaard plans to keep exploring the shifting line between artist and algorithm as AI continues to evolve. “I try to write myself out of my own art,” he says, noting that he’s also looking to experiment with agentic AI to see whether such agents can create entirely original works—even if that means replacing himself in the process.
“I try to write myself out of my own art,” he says again. “I don’t mind not being the artist.”