Made by AI, Directed by David Stein

Search
Digital Artists

People Who Love a Piece of AI Art Are 22 Points More Likely to Think a Human Made It

A study of 11,123 respondents finds that liking a picture and believing it's human-made move together — even for people who say they dislike AI art on principle.

3 min read

A working paper posted to arXiv on October 28 by Zihan Chen, Di Zhu, and Lei Nico Zheng digs into a question that’s been debated mostly by vibes until now: when people love a piece of art, does that change what they believe about who made it? The answer, across more than 11,000 respondents, is yes — and the effect doesn’t depend on whether the work was actually made by a human or a machine.

The setup

The authors use a public “AI Art Turing Test”: 11,123 people each looked at 50 pictures — 26 of them AI-generated, the rest human-made — and judged whether each was made by a human, then named their favorite. 9,152 respondents who picked a favorite make up the core analysis sample. Because the comparison is within each respondent and each picture, the study isn’t just measuring that AI art looks different — it’s isolating what happens to the same picture’s perceived authorship depending on whether it happened to be someone’s favorite.

The headline number

Pictures named as a favorite were 22.1 percentage points more likely to be judged human-made than the same picture when it wasn’t anyone’s favorite, holding respondent and picture fixed. That pattern shows up for both AI-generated and human-made favorites, and it holds across all 50 pictures in the set — not just a handful of convincing outliers.

The part that’s genuinely surprising

You might expect this effect to shrink or disappear among people who are skeptical of AI art — surely someone who dislikes the whole category would be more vigilant about spotting it. The opposite happened. Respondents who were most negative about AI art picked AI-generated favorites less often (47% vs. 60% among the most positive respondents) — but when they did pick one, they were more likely to believe it was human-made. Across every opinion group in the study, somewhere between a third and two fifths of respondents chose an AI-generated favorite they genuinely believed a human had made.

Part of the explanation, per the authors, is social: other respondents also tended to judge those same pictures as human-made, so the belief wasn’t purely individual — it tracked a shared, if mistaken, consensus about which images “look like” they came from a person.

It’s not just images

The authors reanalyzed a separately published AI-poetry experiment, where 230 raters judged each poem. The same pattern turned up there too — liking and perceived human authorship moved together — even though the AI-generated poems in that study received higher average ratings overall than the human-written ones.

Why this matters beyond a lab curiosity

The practical implication cuts against a common assumption in AI-art debates: that disclosure and detection are mostly a technical problem — build a better classifier, label things correctly, and perception will follow the facts. This paper suggests perception runs the other way at least some of the time: liking something functions as evidence of human authorship in people’s heads, and belief about authorship can in turn shape how much they enjoy it. If aesthetic response and perceived provenance are entangled like that, no amount of accurate labeling fully solves the problem it’s meant to solve, because the labels are competing with an existing, instinctive heuristic (“I love this, so a person must have made it”) rather than filling a blank slate.

Related Reading