Why Real Paintings Can Look AI-Generated

Take a look at the two moonlit scenes below. One was painted in 1839. The other was generated by artificial intelligence (AI) in 2026. Which one would you choose as the work of a human artist?

Before looking for the answer, consider what guided your choice. The light? The details? Perhaps one looks a little too polished, while the other feels more natural. How confident are you that those qualities tell you who made it?

I find this question interesting because it asks us to examine something we might otherwise take for granted: what human creativity is supposed to look like.

When the Original Looks Artificial

In a preliminary, unpublished study, I asked 202 U.S. adults to judge images drawn from five 19th-century paintings and five AI-generated versions of the same scenes. Each participant saw five images, presented individually in random order, and identified each as human-made, AI-generated, or something they were unsure about.

Among the 920 answers that named a source, participants were correct about 45 percent of the time. For these images, that was below the 50 percent expected from a coin flip.

But the direction of the mistakes was more interesting. Participants labeled the original paintings as AI-generated 51 percent of the time. They gave that label to the actual AI images only 42 percent of the time. The original looked more artificial than its AI counterpart in four of the five pairs.

And the answer to the opening question? Image A is Johan Christian Dahl’s View of Dresden by Moonlight, painted in 1839. Image B is AI-generated. Among definite answers, 69 percent labeled Dahl’s painting AI, compared with 51 percent for the generated version.

This reversal has a parallel in poetry. Brian Porter and Edouard Machery found that nonexpert readers were more likely to attribute AI-generated poems to humans than poems actually written by established poets (2024). Our expectations about human expression can apparently lead us away from the humans.

Too Perfect to Be Human?

Why might this happen? Participants’ explanations offered a clue. Nearly a quarter mentioned a polished or “too perfect” appearance as a sign of AI. That seems like a reasonable suspicion until we consider how much effort painters have devoted to achieving precisely those qualities.

Consider the still life below. The original is Image B, painted by Severin Roesen in 1852. Its elaborate arrangement and luminous fruit might fit someone’s idea of what an image generator produces. Yet here those qualities belong to a human painter.

One possible explanation is what psychologists Amos Tversky and Daniel Kahneman called the representativeness heuristic: judging something by how closely it resembles our idea of a category (1974). If our mental picture of AI art is smooth, detailed, and beautifully lit, a painting with those features may seem machine-made. This is an interpretation of the findings; the study did not experimentally establish that polish caused the mistakes.

Experience offered little reassurance. Regular users of AI image tools were less likely to answer “not sure,” but showed no detectable accuracy advantage. Being comfortable making a judgment and being correct about it are separate matters.

Of course, five painting pairs cannot tell us how people judge all art. Participants viewed digital reproductions, and each made only five judgments. The results do not establish that expertise is useless or that nobody can detect AI. They do suggest that familiar visual clues can mislead us in this particular setting.

That matters because authorship can change appreciation. In experiments by Lucas Bellaiche and colleagues, artworks received more favorable evaluations when labeled human-created rather than AI-created, even though all the images were generated by AI (2023). The story we attach to an image can become part of what we see in it.

What gives me pause is the possibility that we might dismiss human skill precisely because it looks accomplished. Before treating an image as obviously artificial, I would want to know more about where it came from.

Dahl painted that moonlit scene in 1839. The suspicion that it looks like AI belongs entirely to us.

References