ut the same picture in front of two groups of people. Tell one group a person made it and the other that a program did. In study after study the second group likes it a little less, finds it less moving, and would pay less for it. This page is about that gap: how large it is, what produces it, when it disappears, and what it does and does not tell us about art. It sits between anthropomorphism (we give minds to machines) and algorithm aversion (we hold machines to a harsher standard), and it borrows from both.

the question

An artwork is judged twice: once as an object and once as the trace of an act. The second judgment needs a story about who acted. Philosophers have argued for a long time about how much the story should matter. Psychology asks a narrower thing: how much it does matter, to ordinary viewers, and through which beliefs. The narrow question has become urgent because text-to-image programs now make pictures that most people cannot tell from human work by looking, so the label is often the only difference left.

the classic findings

The groundwork predates image generators. Kruger and colleagues (2004) told people how long a poem or painting had taken to make, holding the work itself constant, and found that more reported effort raised judgments of quality and value: effort works as a heuristic for quality when quality is hard to see. Newman & Bloom (2012) asked why an original is worth more than a perfect copy and found two strands: the original is the record of a particular creative performance, and it carries a kind of contact with the artist that people treat almost as contagion. Neither strand is a fact about the surface of the work.

The labelling experiments followed. Ragot, Martin & Cojean (2020) showed 565 participants paintings that were either made by people or generated by a network, with labels; work labelled human was rated higher on liking, beauty, novelty and meaning. Bellaiche and colleagues (2023) tightened the design by attaching "human" and "AI" labels at random to the same images, so the pictures could not explain the difference. The "human" label raised liking, beauty, profundity and judged worth, and the effect ran partly through how much story and effort viewers read into the work. Millet and colleagues (2023) ran four studies with 1,708 participants: work believed to be machine-made was judged less creative and drew less awe, and the bias was larger in people who hold anthropocentric views of creativity.

One small irony. Over the same years people lost the ability to tell. Gangadharbatla (2022) found viewers could not reliably sort machine-made from human-made images, and that they guessed by genre, crediting representational pictures to people and abstract ones to machines. Porter & Machery (2024) found that non-specialist readers judged poems written by a language model to be human more often than poems by well-known poets, and rated them more favourably on rhythm, beauty and most other qualities; yet in a second study, with a new sample, readers rated any poem lower when told it was machine-made. The authors read this as readers taking their own liking for a poem as evidence that a person wrote it. The penalty lives in the label, not in the page.

what held up, and what did not

The direction of the label effect is consistent across labs, media and countries; its size is modest and varies. What has not held up is the idea that it is fixed. Hong & Curran (2019) found that knowing a work was machine-made did not, by itself, lower ratings across the board; what lowered them was the viewer's prior belief that a machine cannot make art. Chamberlain and colleagues (2018) found a bias against computer-generated art when viewers judged finished images, but it reversed when they watched a robot artist at work: seeing the making supplied the missing story. Both results say the penalty is about beliefs and narratives, which can change, rather than about some permanent property of the images.

the older version of the argument

Every new way of making a picture has met this argument. Baudelaire (1859) wanted photography to be the arts' "very humble servant" and feared its trespass on the imaginary; the full passage is in the archive. A generation earlier, the public argued about whether an automaton could play chess, and Poe (1836) settled it to his own satisfaction:

“It is quite certain that the operations of the Automaton are regulated by mind, and by nothing else.”
— Edgar Allan Poe, "Maelzel's Chess-Player", Southern Literary Messenger, 1836 (Wikisource)

He was right about the chess player, which hid a man. The modern case is the inverse: there is no one hidden in the cabinet, and the question has become whether that matters. Compare the ELIZA effect, where people supplied a mind that was not there, with the label effect, where they withdraw value from a picture because no mind is said to be there.

the human side

The viewers in these studies are not the only humans involved. The image programs were trained on very large collections of pictures gathered from the open web, most made by people who were never asked. Zhou & Lee (2024) followed more than four million artworks posted by over 50,000 users of an art-sharing site and found that adopting text-to-image tools raised output by about a quarter and favourites per view by about half; the most novel ideas got more novel while the average subject got less so, and visual novelty fell. Artists have also built tools back: Glaze (Shan et al., 2023) and Nightshade (Shan et al., 2024) perturb images so that models trained on them learn the wrong style or concept. The legal side, with dates, is on the timeline.

live debates

Is the label effect a bias at all? If the value of a work includes the act that made it, as Newman & Bloom suggest ordinary viewers believe, then rating a labelled machine picture lower may be a correct response to real information, not an error. The experiments show that the label moves judgments; they cannot show that it should not. Does the effect shrink with exposure? Some of the gap looks like unfamiliarity and some like principle; the studies above mostly sample people before generators were in daily use.

methodological critiques

Most studies use online samples, short exposures and images chosen by the researchers, often from one generator, rated on scales built for paintings. Labels are sometimes confounded with the images (the "AI" set really was machine-made), which is why the random-label design of Bellaiche and colleagues matters. Few studies measure behaviour that costs anything, such as buying, and almost none follow viewers over time. The poetry result of Porter & Machery compares a model's imitations with a sample of famous poems read by non-specialists, which says more about how lay readers read than about poetry.

still open

Whether the penalty is about authorship, effort, intention or scarcity, and whether those can be pulled apart; what happens to judgments of human work once anything could have been generated; and how disclosure rules and provenance records will change the picture when the label is no longer a matter of trust. [the house's view] We think the gap is mostly a measure of how much a picture's story matters to people, and that this is worth knowing whatever one thinks of the machines.

in the tour: a wing planned for the tour, called "Who made this?", will have a room called “The Labels Arrive Late” ···

Reading list

1. Kruger et al. (2004). The effort heuristic. Journal of Experimental Social Psychology. doi:10.1016/S0022-1031(03)00065-9
2. Newman & Bloom (2012). Art and authenticity: The importance of originals in judgments of value. Journal of Experimental Psychology: General. doi:10.1037/a0026035
3. Ragot, Martin & Cojean (2020). AI-generated vs. human artworks. A perception bias towards artificial intelligence?. Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3334480.3382892
4. Bellaiche et al. (2023). Humans versus AI: whether and why we prefer human-created compared to AI-created artwork. Cognitive Research: Principles and Implications. doi:10.1186/s41235-023-00499-6
5. Millet et al. (2023). Defending humankind: Anthropocentric bias in the appreciation of AI art. Computers in Human Behavior. doi:10.1016/j.chb.2023.107707
6. Gangadharbatla (2022). The Role of AI Attribution Knowledge in the Evaluation of Artwork. Empirical Studies of the Arts. doi:10.1177/0276237421994697
7. Porter & Machery (2024). AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably. Sci. Rep.. doi:10.1038/s41598-024-76900-1
8. Hong & Curran (2019). Artificial intelligence, artists, and art: attitudes toward artwork produced by humans vs. artificial intelligence. ACM Transactions on Multimedia Computing, Communications, and Applications. doi:10.1145/3326337
9. Chamberlain et al. (2018). Putting the art in artificial: Aesthetic responses to computer-generated art. Psychology of Aesthetics, Creativity, and the Arts. doi:10.1037/aca0000136
10. Baudelaire (1859). Salon de 1859, part II, "Le public moderne et la photographie". Collected in Curiosités esthétiques, Paris: Michel Lévy frères, 1868; transcribed from that edition at Wikisource. fr.wikisource.org/wiki/Curiosit%C3%A9s_esth%C3%A9tiques/Salon_de_1859
11. Poe (1836). Maelzel's Chess-Player. Southern Literary Messenger, April 1836. Transcribed at Wikisource. en.wikisource.org/wiki/Maelzel%27s_Chess-Player
12. Zhou & Lee (2024). Generative artificial intelligence, human creativity, and art. PNAS Nexus. doi:10.1093/pnasnexus/pgae052
13. Shan et al. (2023). Glaze: protecting artists from style mimicry by text-to-image models. USENIX Security 2023; arXiv:2302.04222. arxiv.org/abs/2302.04222
14. Shan et al. (2024). Nightshade: Prompt-Specific Poisoning Attacks on Text-to-Image Generative Models. 2024 IEEE Symposium on Security and Privacy (SP). doi:10.1109/SP54263.2024.00207

see also

	essay: Who made this? — the house on its own pictures
	annotated: machine-made art — every paper on this page, with notes
	archive: Freud, 1919 — the uncanny, and the doll that might be alive