Can we algorithmize emotion?
In laboratories and auction houses alike, a quiet threshold has been crossed: machines trained on the accumulated visual heritage of humanity can now produce paintings that human eyes cannot reliably distinguish from those made by human hands. A study from the University of Colorado Boulder, testing some 200 observers, found that algorithmic art deceives as often as it convinces — and yet the deeper question it raises is not about perception, but about the soul of creation itself. If a machine can mirror the surface of human expression with such fidelity, what remains uniquely ours, and what does it mean to make something that moves another person?
- Roughly 200 participants failed more often than not to tell AI-generated paintings apart from human ones, signaling that the visual boundary between human and machine creativity has effectively collapsed.
- A Paris-based collective's AI portrait sold at Christie's for $432,500 in 2018 — nearly 43 times its estimate — proving that algorithmic art has already entered the highest tiers of the art market.
- A revealing bias persists: people instinctively assume abstract works come from machines and representational ones from humans, even though abstraction has historically been the most emotionally charged form of human expression.
- Knowing a work was made by a machine consistently lowered how people valued it, exposing a gap between what our eyes accept and what our hearts are willing to credit.
- Researchers are now asking whether emotion itself can be algorithmized — whether machines can one day not just replicate the look of art, but produce the feeling that makes art matter.
A team at the University of Colorado Boulder posed a deceptively simple question: can people tell the difference between a painting made by a human and one made by a machine? Their answer, published in Empirical Studies of the Arts, was unsettling — no, not reliably. When roughly 200 volunteers were shown paintings online, they failed more often than not to identify which came from human hands.
The algorithms behind this illusion were trained on thousands of real works, absorbing patterns of color, composition, and brushwork until they could reproduce them with uncanny precision. The stakes are not merely academic. In 2018, Christie's sold an AI-generated portrait — a blurred, portly gentleman rendered by a generative adversarial network trained on 15,000 historical portraits — for $432,500, after a five-way bidding war dwarfed its $10,000 estimate.
Yet the study uncovered a telling quirk in human perception: participants consistently assumed abstract works were machine-made and representational ones were human — a bias that ignores the fact that Picasso and Rothko poured profound emotion into abstraction. The machines, of course, felt nothing. They simply learned patterns and reproduced them.
Lead researcher Harsha Gangadharbatla identified creativity as "the last frontier where humans still have an advantage," but stopped short of calling it permanent. His real question was whether machines could eventually produce work that genuinely moves people. "Can we algorithmize emotion?" he asked — and left the answer open.
Meanwhile, tools like PaintBot are pushing further still, studying the precise brushwork of Vermeer and Van Gogh stroke by stroke, then generating entirely new works in those styles from photographs. What remains unresolved is whether any of this constitutes art in the human sense, or whether it is something more like a sophisticated mirror — reflecting our own creativity back at us, faithfully, but without understanding what it reflects.
A team of researchers at the University of Colorado Boulder set out to answer a deceptively simple question: Can you tell the difference between a painting made by a human and one made by a machine? The answer, according to their study published in Empirical Studies of the Arts, is no—at least not reliably. When roughly 200 people were shown images online and asked to identify which paintings came from human hands and which from algorithms, they failed more often than not.
The machines in question were trained on thousands of real paintings, learning to recognize patterns in color, composition, brushwork, and style. The more images an algorithm analyzed—whether expressionist abstracts or photorealistic landscapes—the more convincingly it could replicate them, down to the finest details of how paint sits on canvas. This is not a parlor trick. AI-generated artwork is already selling for staggering sums. In 2018, Christie's auctioned off a portrait titled "Edmond de Belamy," created by a Paris-based collective using a generative adversarial network trained on 15,000 portraits from the 14th and 15th centuries. The painting, depicting a slightly blurred, portly gentleman in a dark coat and white collar, sold for $432,500—far exceeding its initial estimate of $10,000 after a five-way bidding war.
Yet the study revealed something curious about human perception. When volunteers were shown abstract and representational works, they tended to assume the abstract pieces came from machines and the representational ones from humans. This bias persisted even though some of the greatest abstract art in history—the work of Picasso, Rothko—emerged from profound human emotion. The algorithms, meanwhile, had no such emotional investment. They simply learned patterns and reproduced them.
Harsha Gangadharbatla, the study's lead author, framed the deeper question this research raises. "Creativity and the creative process is the last frontier where humans still have an advantage," he told the Times. But he was careful not to declare victory permanent. The real frontier, he suggested, is whether machines can eventually produce work that moves people—that triggers the emotional response we associate with art. "If computers are producing something humans implicitly consider human, like the quintessence of what it means to be human, then what does it mean to be us? The big question is: can we algorithmize emotion?"
The study also found that knowing a work's origin shaped how people valued it. Paintings attributed to machines were consistently rated lower than identical pieces attributed to humans—a gap that persists even as AI art commands six-figure prices at auction. Other research teams are pushing the technology further. PaintBot, developed by researchers at the University of Maryland, ByteDance AI Lab, and Adobe Research, doesn't just replicate existing styles; it studies the brushwork of masters like Vermeer and Van Gogh, learning the precise way each artist applied paint to canvas—the color, density, and size of each stroke. Once trained, it can generate entirely new works in that learned style based on photographs fed to it.
What remains unclear is whether any of this constitutes art in the way humans understand it, or whether it is something else entirely: a sophisticated mirror held up to human creativity, reflecting back what we have already made. The machines can fool our eyes. Whether they can touch something deeper—whether they can create rather than merely recombine—remains the question that keeps researchers like Gangadharbatla awake.
Citações Notáveis
Creativity and the creative process is the last frontier where humans still have an advantage. The big question is: can we algorithmize emotion?— Harsha Gangadharbatla, University of Colorado Boulder
A Conversa do Hearth Outra perspectiva sobre a história
So people literally couldn't tell the difference? They saw a painting and had no idea if it came from a person or a computer?
Not consistently, no. About 200 people in the study were shown images and asked to identify the source. They guessed wrong more often than they guessed right. The algorithms had been trained on thousands of real paintings, so they'd learned the patterns—the way colors work together, how brushstrokes sit on canvas.
But surely there's something that gives it away? Some tell-tale sign that it's machine-made?
That's the interesting part. People did have a bias, but it wasn't based on actual quality. They assumed abstract art came from machines and realistic art came from humans. Even though Picasso and Rothko—some of the greatest abstract painters ever—were pouring human emotion into every canvas.
So the bias says more about us than about the art.
Exactly. We have this assumption about what machines should produce, and we see what we expect to see. But the researcher, Gangadharbatla, pointed out something else: we also value art differently once we know its origin. Same painting, same image—but if you tell people a computer made it, they rate it lower.
Even though AI paintings are selling for hundreds of thousands of dollars?
Yes. There's a real disconnect. The market is willing to pay enormous sums, but individual people still seem to believe human-made art is inherently more valuable. The real question Gangadharbatla kept coming back to was whether machines could ever make something that actually moves people emotionally—not just something that looks right, but something that feels true.
And can they?
Not yet. That's the frontier he talked about. The machines can replicate technique perfectly. But emotion? That's still the thing we don't know how to algorithmize.