Beauty Is Human: When People Love AI-Generated Art, They Think a Human Made It

1 Introduction

• The paper distinguishes belief-driven evaluation from the beauty-is-human inference, in which aesthetic response serves as evidence of a human maker.It also considers shared perceptions of a work and pressure to give consistent answers as possible links between liking and authorship judgments.Source: §1 · p. 2
• The study analyzes 542,208 judgments from 11,123 people across 50 pictures and reanalyzes a poetry experiment with ratings recorded before authorship judgments.The image test also collected respondents’ favourites, detection strategies, art familiarity, and opinions of AI art.Source: §1 · p. 2
• 22.1 percentage points: a picture was more likely to be judged human-made when it was the respondent’s favourite, holding respondent and picture fixed.The association was positive for all 50 pictures, similar for AI-made and human-made favourites, and present under every reported detection strategy.Source: §1 · p. 2
• In the poetry reanalysis, each point of a poem’s rating predicted an 11.5-percentage-point increase in judging its source as human, for AI- and human-written poems alike.AI-written poems were rated higher, while the gap between ratings for poems believed human- or AI-written was similar in size to the effect of being told the author.Source: §1 · p. 2
• People who disliked AI art chose fewer AI-made favourites but more often believed those favourites were human-made.Across opinion groups, 33.5% to 40.8% named an AI-made favourite they believed a human had made; chosen AI pictures also tended to be judged human by others.Source: §1 · p. 2
• The findings characterize picture-level patterns, relate stated opinions to choices and beliefs, and set out experiments that could distinguish competing explanations.The authors connect these findings to AI disclosure, detection advice, and research that asks people both to rate a work and judge its source.Source: §1 · p. 2

2.1 Telling AI-made work from human work

• People classify AI-made work near chance in several media, while reported detection cues can be predictably misleading.Prior studies describe errors involving intuitive cues in self-presentations and faces, and associations between representational styles and human authorship in art.Source: §2.1 · p. 3

2.2 The human premium

• Explicit AI labels lower evaluations of paintings, music, and news headlines, but label studies do not show how people form authorship beliefs without labels.Proposed reasons for the human premium include valuing effort, authentic connection, experience, and agency attributed to a human maker.Source: §2.2 · p. 3

2.3 Why liking and believed authorship might go together

• Four proposed links are belief-driven evaluation, aesthetic response as evidence of human authorship, shared perceptions that raise liking and human attribution, and response consistency.The beauty-is-human inference predicts an association even without AI aversion; observational data can check predictions but cannot fully separate these nonexclusive processes.Source: §2.3 · pp. 3–4

2.4 Prior evidence on the association

• Earlier studies found liking and believed human authorship linked across images, music, and poetry, but had limited participant and work counts.The paper identifies open questions about variation across works, AI-made works, attitudes, and detection strategies; person and work fixed effects address a stricter adjustment than random intercepts.Source: §2.4 · p. 4

2.5 Opinions, strategies, and familiarity

• AI-art opinions, reported detection strategies, and familiarity may relate differently to favourite choices and beliefs about authorship.The section frames whether negative opinions correspond to fewer AI-made favourites or more human attribution, and whether error-focused strategies or art experience weaken the association.Source: §2.5 · p. 5

2.6 Research questions

• The five research questions cover the association and its consistency, whether it extends to AI-made works, differences across people, poetry with ratings first, and reasons for source judgments.Source: §2.6 · p. 5

3.2 Study 2: ratings recorded before the source judgment

• Study 2 reanalyzes 230 participants who rated ten poems before judging authorship in the no-author-information condition.Each poem was rated on fourteen qualities; ratings of goodness were the primary liking measure, while labelled conditions provided a benchmark for authorship-label effects.Source: §3.2 · p. 6

3.3 Explanations

• Explanations were keyword-coded for merit, intention, liking, technical cues, polish, and lack of merit; this exploratory descriptive analysis was excluded from the RQ1–RQ4 tests.Categories could overlap, and the coding measure may miss paraphrases or misread explanations.Source: §3.3 · p. 6

3.4 Estimation

• Study 1 estimates a linear probability model with respondent and picture fixed effects, so the favourite coefficient is an association in percentage points net of stable respondent and picture differences.Standard errors are clustered by respondent; a logit specification with picture indicators and a respondent leave-one-out measure gives the same answer.Source: §3.4 · pp. 6–7
• Robustness analyses account for uncertainty across the 50 test pictures and compare favourites with other pictures in co-judgment clusters to address shared appearance.The study also interacts favourite status with true AI origin to assess differences by source; the clusters capture shared appearance, not an independent measure of style.Source: §3.4 · p. 7
• Signal-detection summaries compare favourites with the same pictures judged by non-fans, using sensitivity d′ and response criterion c.These are aggregate descriptions across respondents, not estimates of within-person change.Source: §3.4 · p. 7
• For Study 2, participant and poem fixed effects estimate the association between a participant’s authorship judgment and their rating of a poem.The paper compares that association with the difference between assigned author-label conditions only to compare scale; these are different estimands.Source: §3.4 · p. 7

4 Results

• Respondents judged 60.6% of pictures correctly on average; AI-made pictures were judged human 42% of the time and human-made pictures 63% of the time.Median accuracy was 60%.Source: §4 · p. 7

4.1 Favourites are judged human more often (RQ1)

• Holding respondent and picture fixed, respondents judged their own favourite picture human-made 76.6% of the time versus 54.6% for the same pictures judged by non-choosers, a 22.1 pp favourite association.Source: §4.1 · p. 7
• Picture-specific estimates were positive for all 50 pictures, ranging from 5.3 to 38.1 pp, and remained substantial after comparisons addressing shared style, recognition, and favourite overlap.Co-judgment-cluster comparisons yielded 19.0–20.0 pp; among respondents who had seen none of the pictures, the estimate was 22.6 pp.Source: §4.1 · pp. 7–8, Figure 1 · p. 8

4.2 The association appears for AI-made and human-made works alike (RQ2)

• The association was similar for AI-made and human-made favourites—21.8 pp and 22.4 pp—and favourites drew more “human” judgments regardless of source.AI-made pictures were 55.6% of favourites versus 52.2% of other pictures, so favourite status was not positively diagnostic of human authorship in this set.Source: §4.2 · pp. 8–9

4.3 Opinions of AI art, choices, and beliefs (RQ3)

• Among the most negative versus most positive respondents, 47% versus 60% chose an AI-made favourite, yet 80% versus 56% believed an AI-made favourite was human-made.Different picture choices and same-picture belief differences both contributed descriptively; across opinion groups, 33.5%–40.8% named an AI-made favourite they believed human-made.Source: §4.3 · pp. 9–10, Figure 2 · p. 10, Table 4 · p. 10

4.4 Reported strategies and familiarity (RQ3)

• The favourite association appeared under every reported detection strategy, ranging from 17.1 to 25.1 pp, while human-art expertise did not weaken it clearly.It was 7.4 pp among 63 respondents who called AI art their job or greatest passion, 14.7 pp smaller than among all other respondents.Source: §4.4 · p. 11, Table 5 · p. 11, Figure 3 · p. 12

4.5 Poems rated before the source judgment (RQ4)

• In the poetry reanalysis, each additional quality-rating point went with an 11.5 pp higher probability of a “human” judgment, although AI poems received higher ratings.Top-rated poems were 23.6 pp more likely to be judged human than the same participant’s other poems; the top-rated contrast was 18.5 pp for AI-written and 32.7 pp for human-written poems.Source: §4.5 · pp. 11–12, Table 6 · p. 13

4.6 Reasons in explanations (RQ5, exploratory)

• Explanations matching merit keywords usually accompanied “human” judgments, whereas technical-cue and polish explanations more often accompanied “AI” judgments.Merit keywords matched 1.7% of picture explanations and 8.5% of poem explanations; fans’ explanations matched merit keywords in 7.7% of cases versus 1.6% for other respondents explaining the same pictures.Source: §4.6 · pp. 12–13, Table 7 · p. 14

5.1 What the association adds to prior evidence

• The favourite association was positive for all 50 pictures and similar for AI-made and human-made works, extending prior evidence with work-by-work estimates and a poetry reanalysis.The authors note that the association ranged from 5.3 to 38.1 percentage points and that the poetry pattern held when ratings came first.Source: §5.1 · p. 14, §5 · p. 14

5.2 Explaining the association

• Belief-driven evaluation and the beauty-is-human inference are both consistent with the findings, but neither is excluded; separating them requires experiments that change experience or belief while holding the work fixed.The authors also discuss response consistency and unmeasured person-specific perceptions of particular pictures as possible contributors.Source: §5.2 · pp. 14–15

5.3 Opinions, choices, and beliefs

• People who disliked AI art chose fewer AI-made favourites, but when they chose one they were more likely to believe it human-made, leaving 33.5%–40.8% across opinion groups with such a favourite.The authors caution that these post-task comparisons do not identify the direction of influence.Source: §5.3 · p. 15

5.4 Practical implications

• Without labels, people tended to credit the works they liked to people, so the authors propose testing whether reliable disclosure changes choices, sharing, or willingness to pay.The data include no purchases or payments, making market effects a hypothesis for future studies.Source: §5.4 · p. 15
• Reported use of technical-error detection strategies did not weaken the favourite association, but the test removed obvious AI defects and did not measure training or exclusive reliance on a strategy.Whether advice about liking improves judgments remains a question for randomized comparison.Source: §5.4 · p. 15

5.5 Limitations and future research

• Study 1 used voluntary readers of one blog and pictures curated to be difficult, while Study 2 covered ten poems, so the estimates do not establish prevalence for other people or works.The authors call for representative samples of people and works encountered online.Source: §5.5 · pp. 15–16
• Neither study manipulated liking, and ratings recorded before source judgments do not establish that ratings formed before source beliefs; causal direction therefore remains unresolved.The proposed next study would randomize an evaluative experience while holding the work fixed and use an active control and manipulation checks.Source: §5.4 · p. 15, §5.5 · p. 16
• Opinions were measured after the test, several measures were single self-reports, and the data contain no costly choices, limiting conclusions about opinions, expertise, and behavior beyond ratings.The authors suggest improved measures and studies of purchases, sharing, or creator credit before and after disclosure.Source: §5.5 · p. 16

6 Conclusion

• When people cannot tell who made a work, their favourites are the works they judge human-made, including AI-made pictures they may believe were made by people.The association held for every picture and across opinion groups; roughly a third to two fifths in each group favoured an AI-made picture they took to be human.Source: §6 · p. 16

• Included respondents named a favourite and judged that picture.Source: Table 8 · p. 19
• Excluded respondents are shown as a separate group in the favourite-analysis sample flow.Source: Table 8 · p. 19
• Reported shares use all respondents in the corresponding included or excluded group as the denominator.Source: Table 8 · p. 19

• Both the raw gap and adjusted favourite-association estimate are positive for all 50 pictures.The adjusted estimates include respondent and picture fixed effects; standard errors are clustered by respondent.Source: B The pictures of Study 1 · p. 19, Table 9 · p. 19
• Co-judgment clusters group pictures that the same respondents tend to call human, beyond their general tendency to do so.The four-cluster solution contains groups of 18, 17, 11, and 4 pictures; six- and eight-cluster estimates appear in Table 3.Source: B The pictures of Study 1 · p. 19
• The table lists each picture’s true source, the share judged human, and the number of respondents who named it as their favourite.Source: Table 9 · p. 19, B The pictures of Study 1 · p. 19

• Each row in Table 10 is a separate linear model predicting the human-authorship judgment from a standardized quality rating.Source: Table 10 · p. 20
• The models include participant and poem fixed effects, with standard errors clustered by participant.Source: Table 10 · p. 20
• The rated qualities are correlated measures of a single evaluation.Source: Table 10 · p. 20

• Explanations were lower-cased and matched on whole words using keyword categories that are not mutually exclusive.A merit term preceded by a listed negation within three words counted as lack of merit; other non-negated merit terms could still count.Source: D The keyword measure of reasons · p. 21, Table 11 · p. 21
• Before matching technical cues, the coding removed phrases where cue words did not refer to the work.Examples include “hand-painted,” “by hand,” and “human perspective.”Source: D The keyword measure of reasons · p. 21
• The measure misses paraphrases and can misread explanations, with errors that need not be random with respect to judgment.For example, “AI has no emotions” could be coded as lack of merit even when used to judge a poem human.Source: D The keyword measure of reasons · p. 21
• Table 12 provides randomly drawn verbatim examples from explanation categories, including misreadings.The examples were selected from explanations 15–110 characters long within each category.Source: Table 12 · p. 21, D The keyword measure of reasons · pp. 21–22