cs.HCOct 28, 2026
SaveBeauty Is Human: When People Love AI-Generated Art, They Think a Human Made It
Organizations: Stevens Institute of Technology, Hoboken, NJ, USA · University of Massachusetts Boston, Boston, MA, USA
Abstract
Generative AI now produces images and poems that people cannot reliably distinguish from human-made work, so people often judge for themselves who made a work. We examine how these perceived-authorship judgments relate to aesthetic preference. In a public online AI Art Turing Test, 11,123 people judged whether each of 50 pictures (26 AI-generated) was made by a human and were asked for their favourite. Among the 9,152 respondents who judged their favourite, a picture was 22.1 percentage points more likely to be judged human-made when it was the respondent's favourite, holding respondent and picture fixed. The association was similar for AI-generated and human-made favourites, was positive for all 50 pictures, and appeared under every reported detection strategy. Respondents most negative about AI art chose fewer AI-generated favourites (47% against 60% among the most positive) but more often believed them human-made, partly because others also judged those pictures human. As a result, in every opinion group a third to two fifths of respondents named an AI-generated favourite they believed human-made. A reanalysis of a published experiment on AI-generated poetry, in which 230 people rated each poem before judging its authorship, found the same pattern, although the AI poems were rated higher. When provenance is hidden, liking and perceived human authorship travel together, consistent with aesthetic response serving as evidence of a human maker and with beliefs about the maker shaping enjoyment. The findings qualify research on bias against AI art, which typically uses explicit labels, and inform AI disclosure policy.
Figures & tables
| Study | Medium; participants; works | Liking measure | Source judgment | Order | Adjustment for people and works | Moderators |
|---|---|---|---|---|---|---|
| Chamberlain et al. [22] , Study 1 | Images; 65; 60 (30 computer-generated) | Attractiveness, 1–7 | Binary | Randomized between participants (blocks) | Mean ratings by own categorization, within participants | Order; art education |
| Shank et al. [13] , Study 1 | Music; 146 and 149; 20 human-composed excerpts per genre | Liking, 7 points | Binary | Source judgment, then liking | Random intercepts for participants and excerpts | Genre |
| Grassini and Koivisto [24] | Images; 201; 40 (20 AI-made) | Liking, 0–100 | Continuous, 0–100 | Liking listed before origin | Random intercepts for participants and images | Personality, creative identity, technology attitudes |
| Porter and Machery [5] , Study 2 | Poems; 230 of 696 guessed the source; 10 (5 AI-generated) | 14 qualities, 7 points | Binary | Ratings, then source judgment | Random intercepts for participants and poems (factor scores) | Framing arms |
| This paper, Study 1 | Images; 9,152 of 11,123; 50 (26 AI-made) | One favourite | Binary | Favourite after all judgments | Respondent and picture fixed effects; picture-specific estimates | Opinion of AI art, strategy, familiarity |
| This paper, Study 2 | Reanalysis of Porter and Machery [5] , Study 2 | Quality, 1–7 | Binary | Ratings, then source judgment | Participant and poem fixed effects | Source of poem; label benchmark |
Table 1: The closest prior studies of liking and believed authorship, and this study.
| Study 1: AI Art Turing Test | Study 2: poetry experiment | |
|---|---|---|
| Source | Public response file of an online test run by a blog [ 1 , 39 ] | Open data of Porter and Machery [5] , Study 2 [ 40 ] |
| Works | 50 pictures: 24 human-made, 26 AI-made | 10 poems: 5 human-written, 5 AI-generated |
| Respondents | 11,123; analysis sample of 9,152 who named and judged a favourite | 696; analysis sample of 230 in the condition that judged the source |
| Judgments | 542,208 source judgments (453,865 in the analysis sample) | 2,300 source judgments |
| Liking | One favourite picture, named after all judgments | Fourteen 7-point ratings per poem, given before each judgment |
| Explanations | 49,268 free-text explanations for six pictures | 5,238 free-text explanations from Porter and Machery ’s Study 1 |
Table 2: Data sources.
Figure 1: Favourites are judged human more often. (A) Share of judgments “made by a human” for each respondent’s favourite picture and for the same pictures judged by respondents who did not choose them (Study 1, pictures weighted by their number of fans), and for each respondent’s top-rated poem or poems and their other poems (Study 2), by true source. Bars show 95% intervals from 1,000 bootstrap resamples of respondents (participants for poems). These are raw shares; Table 3 and Table 6 give the fixed-effects estimates. (B) Each dot is one of the 49 pictures with at least 30 fans: the share of its fans (vertical) and of all other respondents (horizontal) who judged it human. Dot size is proportional to the number of fans; the dashed line marks equal shares.
| Specification | Estimate, pp (SE) | Respondents |
| Baseline: respondent and picture fixed effects | 22.1 (0.4) | 9,152 |
| Same, standard error clustered by respondent and picture | 22.1 (1.3) | 9,152 |
| Favourite was AI-made | 21.8 (0.6) | 5,092 |
| Favourite was human-made | 22.4 (0.5) | 4,060 |
| Respondent co-judgment-cluster fixed effects (4 clusters) | 20.0 (0.4) | 9,152 |
| Respondent co-judgment-cluster fixed effects (6 clusters) | 19.7 (0.4) | 9,152 |
Table 3: The favourite association in Study 1 across specifications. Linear probability models of judging a picture human-made, with respondent and picture fixed effects unless stated otherwise; standard errors clustered by respondent unless stated otherwise. Co-judgment clusters group pictures that the same people tend to call human. a The logit model uses picture indicators and the respondent’s leave-one-out share of “human” judgments in place of respondent fixed effects. b Each estimate drops one picture and the respondents whose favourite it was, so the number of respondents varies.
| Opinion of AI art on an artistic level | |||||
|---|---|---|---|---|---|
| 1 (loathe) | 2 | 3 | 4 | 5 | |
| Respondents | 1,106 | 1,891 | 2,181 | 2,620 | 1,270 |
| Favourite was AI-made | 47% | 53% | 55% | 60% | 60% |
| Favourite believed human-made | 88% | 84% | 78% | 72% | 64% |
| AI-made favourite believed human-made | 80% | 78% | 69% | 64% | 56% |
| Same pictures judged human by others | 54% | 52% | 47% | 43% | 38% |
Table 4: Opinion of AI art, choice of favourite, and belief about the favourite (Study 1). Respondents who named and judged a favourite. The row on other respondents gives the share of all other respondents who judged the chosen AI-made favourites human. Accuracy is computed for the same respondents. The favourite association is estimated within each opinion group as in Equation ( 1 ).
Figure 2: Choice and belief by opinion of AI art (Study 1). (A) By opinion of AI art, the share of respondents whose favourite was AI-made, whose favourite they believed human-made, and whose favourite was AI-made and believed human-made. (B) The favourite association within each opinion group, with 95% confidence intervals.
| Respondents | Accuracy | Favourite association, pp (SE) | |
| Main strategy (self-reported) | |||
| Looked for minor errors, e.g. hands | 3,703 | 61.9% | 23.4 (0.6) |
| Looked for textural tells | 2,054 | 62.3% | 20.2 (0.8) |
| Guessed themes humans would paint | 1,396 | 58.9% | 19.6 (1.1) |
| Judged good/creative art as human | 1,209 | 59.3% | 23.8 (1.0) |
| Judged technically impressive art as AI | 477 | 58.5% | 17.1 (1.8) |
Table 5: The favourite association by reported strategy and familiarity (Study 1). Accuracy is the mean share of judged pictures answered correctly. The favourite association is estimated within each group as in Equation ( 1 ); differences between groups reported in the text come from a pooled model with group-specific picture effects.
Figure 3: The favourite association by reported strategy and familiarity (Study 1), with 95% confidence intervals. The dashed line marks the association among all respondents.
| Quantity | Estimate (SE) | Unit |
|---|---|---|
| Told-nothing condition: judged “written by a human” | ||
| Each point on “How good is this poem?” (1–7) | 11.5 (0.9) | pp |
| AI-written poems | 11.7 (1.4) | pp |
| Human-written poems | 11.4 (1.1) | pp |
| Human minus AI | 0.3 (1.5) | pp |
| Top-rated poem(s) vs. the respondent’s other poems | 23.6 (2.8) | pp |
Table 6: Liking and judged authorship of poems (Study 2). Linear models with participant and poem fixed effects in the told-nothing condition (230 participants, 2,300 judgments); standard errors clustered by participant. a The label effect compares the two framing conditions (466 participants) with poem fixed effects; the human label named a poet, the AI label “an AI program”.
Figure 4: (A) Share of poems judged “written by a human” by the rating given just before the judgment (Study 2, told-nothing condition), by true source, with 95% intervals from 1,000 bootstrap resamples of participants. (B) Exploratory: share of judgments “human” among explanations matched by each keyword category, for pictures (Study 1) and poems (Study 1 of Porter and Machery ). Intention is shown for pictures only; fewer than 30 poem explanations matched it.
| Pictures (Study 1) | Poems (P&M Study 1) | |||||
|---|---|---|---|---|---|---|
| Explanation mentions… | Share | Human | Diff. (95% CI) | Share | Human | Diff. (95% CI) |
| Merit: beauty, soul, emotion, meaning, creativity | 1.7% | 80% | 20 [ 17, 22] | 8.5% | 90% | 40 [ 36, 43] |
| Intention: deliberate, thoughtful | 2.9% | 68% | 11 [ 9, 14] | 0.4% | 78% a | 27 [ 11, 42] |
| Personal liking: I like / love it | 3.3% | 61% | 7 [ 5, 10] | 4.2% | 60% | 7 [ 1, 14] |
| Technical cue: hands, errors, inconsistency | 13.9% | 33% | 19 [ 20, 18] | 3.2% | 26% | 26 [ 33, 18] |
| Polish: too perfect, too clean, too smooth | 1.4% | 28% | 19 [ 22, 15] | 0.7% | 19% | 36 [ 49, 23] |
Table 7: Reasons in free-text explanations (exploratory). Share of explanations matched by each keyword category (categories are not exclusive), share of the corresponding judgments that were “human”, and the difference in that share between matched and unmatched explanations, within the same picture or poem (pp, 95% CI clustered by respondent). Pictures: six pictures of Study 1. Poems: optional explanations in Study 1 of Porter and Machery [5] . a Fewer than 30 explanations.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| Included | Excluded | |
|---|---|---|
| Respondents | 9,152 | 1,971 |
| Pictures judged (of 50) | 49.6 | 44.8 |
| Accuracy | 60.9% | 59.5% |
| Share of “human” answers | 51.7% | 52.0% |
| Opinion of AI art (1–5) | 3.12 | 2.99 |
| Opinion of AI art missing | 1% | 42% |
Table 8: Respondents included in and excluded from the favourite analysis (Study 1). Included respondents named a favourite and judged it. Shares are of all respondents in each group.
| No. | Picture | Source | Judged human | Fav. | Fans: human | Others: human | Gap | Adjusted (SE) | Cl. |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Angel Woman | Human | 46% | 221 | 80% | 45% | 35.2 | 35.2 (2.8) | 4 |
| 2 | Saint In Mountains | Human | 78% | 329 | 89% | 77% | 12.1 | 12.9 (1.8) | 2 |
| 3 | Blue Hair Anime Girl | Human | 51% | 67 | 71% | 51% | 19.9 | 17.4 (6.2) | 4 |
| 4 | Girl In Field | AI | 38% | 42 | 60% | 38% | 22.0 | 21.1 (7.4) | 1 |
| 5 | Double Starship | Human | 56% | 146 | 79% | 57% | 21.9 | 21.9 (3.5) | 4 |
| 6 | Bright Jumble Woman | AI | 52% | 63 | 70% | 51% | 18.7 | 19.9 (5.7) | 3 |
Table 9: The 50 pictures of Study 1. Source is the true origin; “Judged human” is the share of all respondents calling it human; fans are respondents who named it their favourite and also judged it (Fav.: number of respondents naming it). Gap is the raw difference between fans and other respondents; Adjusted is the picture-specific favourite association with respondent and picture fixed effects (SE clustered by respondent). Cl. is the four-cluster co-judgment solution (Appendix B ).
| Rated quality (1–7) | Per SD, pp (SE) |
|---|---|
| good (overall quality) | 18.7 (1.5) |
| meaningful | 18.0 (1.6) |
| beautiful | 17.8 (1.6) |
| conveys a mood or emotion | 17.7 (1.5) |
| imagery | 16.9 (1.5) |
| moving | 16.6 (1.6) |
Table 10: Judging a poem human-written by each rated quality (Study 2, told-nothing condition). Each row is a separate linear model of the “human” judgment on the standardized rating, with participant and poem fixed effects; standard errors clustered by participant. The qualities are correlated measures of one evaluation.
| Category | Terms |
|---|---|
| Merit | beaut*, gorgeous, lovely, stunning, soul, soulful, emotion, emotions, emotional*, emotive, evocative, expressive, meaningful, meaning, profound, creativ*, imaginative, inspired, inspiring, passion*, charm*, personality, alive, lively, vision, mood, whimsical, playful, humanity, human touch, human experience, sincer*, authentic* |
| Intention | intention*, deliberate*, thoughtful*, purposeful* |
| Personal liking | I like / love / enjoy / adore (optionally “really”), like it, love it, like this, love this, favourite, favorite, my fav |
| Technical cue | hand(s), finger(s), error(s), mistake(s), inconsisten*, incoheren*, nonsens*, doesn’t make sense, makes no sense, no sense, artifact(s), artefact(s), glitch*, warp*, distort*, merg*, melt*, perspective, anatom*, proportion*, asymmetr*, physics, illogical, garbled, deformed, malformed |
| Polish | too perfect / polished / clean / smooth / good / neat / detailed / precise / crisp / sharp / symmetric* / pretty / glossy / shiny / refined / well done; sheen; glossy; overly polished / smooth / detailed / perfect |
| Lack of merit | soulless, lifeless, generic, bland, boring, cheesy, tacky, clich*, trite, meaningless, pointless, uninspir*, uninteresting, sterile, vapid, emotionless, doesn’t say anything, says nothing, nothing to say, no point; negated merit terms |
Table 11: Keyword categories used to code explanations. An asterisk marks any ending; categories are not exclusive.
| Category | Source | Judged | Explanation |
|---|---|---|---|
| Merit | images | human | “Use of positive and negative space, feels vaguely creative” |
| Merit | images | human | “i dont feel like ai would be this creative” |
| Merit | images | AI | “first impression was creatively-arranged cut paper, but then i noticed bizarre irregular silhouettes.” |
| Merit | images | AI | “Same structure as there was before, mess behind does have meaning” |
| Lack of merit | images | human | “Every small stroke has a meaning, making a shape or a light. No meaningless or unclear spots” |
| Lack of merit | images | human | “Looks too boring for an AI” |
Table 12: Verbatim explanations, drawn at random among those of 15–110 characters within each category (spelling as written).
More in this topic
Authorship Attribution
Authorship attribution identifies or verifies who wrote a text, code sample, or other creative work from patterns in its style and content.
See all 67 papers