Humans versus AI: whether and why we prefer human-created compared to AI-created artwork — TEI-tagged view

Lucas BellaichepersName, Rohin ShahipersName, Martin Harry TurpinpersName, Anya RagnhildstveitpersName, Shawn SprockettpersName, Nathaniel BarrpersName, Alexander ChristensenpersName, Paul SelipersName (2023date). “Humans versus AI: whether and why we prefer human-created compared to AI-created artwork.” Cognitive Research: Principles and Implicationstitle·j.

Object of study — whether and why people prefer ostensibly human-created artworks over AI-created artworks, independent of the artterm's physical properties

Research question — do people show a preference for human-labelled over AI-labelled artterm, and which aesthetic-judgement criteria explain that preference

Contribution — two large-sample within-subjects experiments using only AI-created paintings with per-participant randomized 'human'/'AI' labelsterm, isolating top-downterm source effects from bottom-upterm visual differences across nine aesthetic criteria

Full text — tagged

The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.

Humans versus AI: whether and why we prefer human-created compared to AI-created artwork Abstractterm With the recent proliferation of advanced artificial intelligence (AI) models capable of mimicking human artworks, AI creations might soon replace products of human creativityterm, although skeptics argue that this outcome is unlikely. One possible reason this may be unlikely is that, independent of the physical properties of artterm, we place great value on the imbuement of the human experienceterm in artterm. An interesting question, then, is whether and why people might prefer human-compared to AI-created artworks. To explore these questions, we manipulated the purported creatorterm of pieces of artterm by randomly assigning a “Human-created” or “AI-created” labelterm to paintings actually created by AI, and then assessed participants’ judgements of the artworks across four rating criteria (Likingterm, Beautyterm, Profundityterm, and Worthterm). Study 1 found increased positive judgements for human- compared to AI-labelled artterm across all criteria. Study 2 aimed to replicate and extend Study 1 with additional ratings (Emotion, Storyterm, Meaningful, Effortterm·var, and Time to create) intended to elucidate why people more-positively appraise Human-labelled artworks. The main findings from Study 1 were replicated, with narrativityterm (Storyterm) and perceived effortterm·var behind artworks (Effortterm·var) moderating the labelterm effects (“Human-created” vs. “AI-created”), but only for the sensory-level judgements (Likingterm, Beautyterm). Positive personal attitudes toward AI moderated labelterm effects for more-communicative judgements (Profundityterm, Worthterm). These studies demonstrate that people tend to be negatively biased against AI-created artworks relative to purportedly human-created artwork, and suggest that knowledge of human engagement in the artistic process contributes positively to appraisals of artterm.

The online version contains supplementary material available at 10.1186/s41235-023-00499-6.

Introduction Artterm is widely considered to be a uniquely human phenomenon. It can encapsulate and communicate our emotions, be used to express our individualistic and communal experiences, and serve as a social commentary on these experiences, all of which are elements that are commonly thought to be human-specific (Chatterjee, 2014 ). Yet, artterm is also a perceptual product that is engaged through the senses, sometimes independently of the experient’s awareness of the inclusion (or lack thereof) of these human elements in artworks they evaluate. Here, an important distinction emerges between artterm as (1) a purely physical stimulus and (2) a deeper communicativeterm medium of the human experienceterm. The importance of this distinction lies in its ability to permit investigations of the separate and joint influences that these two conceptions of artterm may have on people’s appraisals of artterm, which in turn contributes to our understanding of the factors and processes involved in aesthetic appraisals. And, with the recent development of highly advanced artificial intelligence (AI) models that can produce, as purely physical stimuli , high-quality artworks that are indiscernible from human-created artworks (Gangadharbatla, 2022 ; Johnson, 1997 ; Mazzone & Elgammal, 2019 ), the Computer Age in which we live has afforded considerable experimental control over these two conceptions of artterm. Here, across two studies, we explored the question of whether and why humans might tend to prefer ostensibly human-created artworks over AI-created artworks.

Although the process of appraising artworks undoubtedly involves a subjective element, consistent preferential patterns have nevertheless emerged across peoples’ aesthetic judgementsterm of such artworks (e.g., blue is the most preferred color, whereas yellow is most disliked [Bornstein, 1975 ; Komar & Melamid, 1999 ; Palmer et al., 2013 ]; and representationalterm artterm is enjoyed over abstractterm artterm [Heinrichs & Cupchik, 1985 ; Kettlewell et al., 1990 ; Knapp & Wulff, 1963 ; Mastandrea et al., 2011 , 2021 ]). However, the ways in which we judge artterm are often sensitive to factors in the environment, not just intrinsic preferences. One common finding across literatures of judgement and decision-making is that labelsterm play an important role in people’s general evaluations of things. For instance, people prefer wine they are falsely told is more expensive than other wine (Plassmann et al., 2008 ). The influence that a mere labelterm has on our judgements is also made evident by research showing that, when participants are provided Coca-Cola either with or without its labelterm, they report greater enjoyment of the labelled beverage than the non-labelled beverage, despite their otherwise identical compositions (McClure et al., 2004 ). As demonstrated by these studies, our judgements rely heavily on contextual information. And, perhaps unsurprisingly, the role that context plays in our appraisals extends to judgements of artterm as well. For example, when Newman and Bloom ( 2012 ) provided participants artworks labelled either as originals or as identical forgeries, participants preferred the originals over the forgeries: a finding that was taken to indicate that humans are sensitive to an authentic process of creative productionterm as determined simply by a labelterm. In other words, people consider non-sensory aspects of artterm, like context and background, in their judgements and evaluations of artterm (Blank et al., 1984 ; Chatterjee & Vartanian, 2016 ; Winner, 1982 ).

The influential role that labelsterm play in our judgements is also made clear in contemporary lines of research focused on AI. The general finding emerging from this research is that, in many instances, the labelterm of “AI” is taken to be a pejorative. For instance, Liu et al. ( 2022 ) conducted an experiment wherein participants were asked to read a number of emails and were told that some of the emails (but not others) were drafted with the help of an AI-based language model. Results indicated that, when participants were told that AI played a role in drafting the email, their reports of trust in the email writer decreased. This anti-AI biasterm seems especially apparent in domains that people assume to be human-specific (e.g., those concerning affect or creativityterm; Wilson, 2011 ). Research has only just begun to connect the spheres of aesthetics and AI, yet, the discussion of AI-produced creativityterm arose many decades ago. Indeed, in his 1950 Presidential Address to the American Psychological Association (APA), JP Guilford foreshadowed the development of AI in modern society, noting that, if advances in “thinking machines” (i.e., AI) reached their forecasted height, then “the only economic value of brains left would be the creative thinking of which they are capable” (Guilford, 1950 ). While Guildford’s position is certainly provocative, he may have overestimated the extent to which the apparent ‘last bastion’ of human utility— “creativityterm”—is uniquely human; instead, it could be that, with recent advances in highly advanced “thinking machines,” even our creative abilities could lose their economic value when competing with the abilities of AI.

Initial forays into the AI-art interaction seem to suggest that, not only is AI artterm sufficiently sensorily similar to human-created artterm that people fail to accurately discern its true creator (human or AI), but also that people tend to derogate AI-created artterm as compared to human-created artterm (Chamberlain et al., 2018 ; Gangadharbatla, 2022 ; Mazzone & Elgammal, 2019 ). For example, Chamberlain et al. ( 2018 ) found that participants rated human-created artworks as higher in aesthetic value than AI-created artworks. Interestingly, however, this bias was malleable: indeed, when participants were shown videos of robots that were physically producing artterm, participants’ anti-AI artterm judgements were reduced, suggesting that people may consider the level of artistic effortterm·var exerted by the creator as a marker of the quality of the produced artwork. Similar findings emerged in a more-recent study in which participants were shown, and asked to provide judgements of, two abstractterm paintings (Chiarella et al., 2022 ). While both paintings were in fact created by humans, by randomly assigning a labelterm of “human-created” or “AI-created” to each painting, the researchers deceived participants into believing that one of the paintings was created by AI. Results of this study indicated that participants tended to prefer the painting that was labelled “human-created” relative to the painting labelled “AI-created.” Notably, this preference for human over AI artterm is not specific to visual artterm. In fact, several recent studies have reported similar anti-AI findings in music (Shank et al., 2022 ), creative writing (Raj et al., 2023 ), dance (Darda & Cross, 2023 ), poetry (Köbis & Mossink, 2020 ), and non-art texts (Darda et al., 2023 ). This bias in aesthetic judgementterm is further intensified when judging more “human” aspects of artterm, like evoked emotion, suggesting the need for multiple judgement criteria in AI aesthetics research (Raj et al., 2023 ).

Critically, however, other studies have yielded results that are at odds with those from the aforementioned studies, finding little-to-no differences in evaluations of human- and AI-created artworks (Hong & Curran, 2019 ; Israfilzade, 2020 ; Xu et al., 2020 ). Additionally, though not an explicit investigation of aesthetic-judgements, per se, one study found that participants were equally likely to consider hypothetical products created by human and AI artists as “artterm” (Mikalonytė & Kneer, 2022 ). Put differently, people did not tend to perceive AI artists’ hypothetical products as being lesser than human artists’ hypothetical products. Given these mixed results, it is challenging to ascertain whether humans do in fact prefer human-created artworks over AI-created artworks, and, if so, why this is so. Thus, here, we examined this question with the intention of shedding more light on their answers.

The current studies Across two studies, we sought to extend the extant—albeit sparse—literature on attitudes toward artworks created by humans as compared to artworks created by AI platforms. Our studies had three primary aims. First, while a handful of studies have examined the question of whether people differ with respect to their judgements of human-created versus AI-created artworks, at noted above, mixed results have been yielded, leaving the answer to this question unclear. Thus, by employing two large-sample ( Ns = 150 and 151), highly powered, within-subjects studies including relatively large stimulus sets, we sought to provide greater clarity on the answer to this question. We hypothesized that humans will indeed show a preference for human-labelled artterm relative to AI-labelled artterm.

Second, we sought to identify some of the more-complex criteria that people might rely upon when making aesthetic judgementsterm. Many aesthetic studies have used simple probes assessing general “likingterm” and “aesthetic beautyterm” to index artterm judgements (for a review, see Chatterjee & Cardillo, 2022 ). However, most models of aesthetics consider aesthetic judgementsterm to be a multifaceted and/or hierarchical cascade of evaluations, implicating more-complex and elaborative processes in judgements of artterm that seek to understand meaning and communication through the piece (Berlyne, 1960 , 1971 ; Chatterjee & Vartanian, 2016 ; Cupchik & Berlyn, 1979 ; Graf & Landwehr, 2015 ; Leder et al., 2004 ; Silvia, 2005 ). This is supported, in part, by the growing discussion of artterm not only as a bottom-upterm appraisal of only visual features that induces likingterm or judgements of beautyterm, but a device for communication and socio-epistemic value (see Sherman & Morrissey, 2017 ). This distinction considers human expression and, as such, is essential to investigate in aesthetic judgementsterm of artterm created by AI. Thus, in Study 1, in addition to simply indexing Likingterm and Beautyterm, which we consider to be more-passive, surface-level appraisals in line with fluency models of aesthetics (see Leder et al., 2004 ; Graf & Landwehr, 2015 ), we also obtained assessments of how profound participants found each painting to be, and how much money they would (hypothetically) spend on each painting. These latter criteria require higher cognitive elaboration to determine communicativeterm properties of artterm, as discussed in Graf and Landwehr ( 2015 ) and Sherman and Morrissey ( 2017 ). In Study 2, we further developed our probing procedure by, in addition to asking the four questions from Study 1 (i.e., Likingterm, Beautyterm, Profundityterm, Worthterm), asking participants to rate each painting on the following additional communicativeterm attributes of the artterm to understand potential interactions between surface-level and communicativeterm engagement processes: emotionality (how much the artwork evoked an emotion in the viewer; Emotion), narrativityterm (the degree to which the artwork portrayed an imagined narrative in the viewer; Storyterm), meaningfulness (how personally meaningful the artwork was to the viewer; Meaningful), perceived effortterm·var (how much effortterm·var the viewer thought went into the creation of the artwork; Effortterm·var), and perceived time (how much time the viewer thought went into the creation of the artwork; Time). We predicted that the anticipated preference for human-labelled artterm over AI-labelled artterm can be, at least partially, explained by a perceived lack of integration of the human experienceterm in AI-created artworks as measured by these wide-ranging aesthetic criteria.

Third, given that aesthetic appraisals are personal and subjective (Chatterjee, 2014 ; Roseman & Evdokas, 2004 ), also imperative in their investigation is consideration of individual differences that might reliably predict these judgements. In this vein, past work has considered the personality trait of openness to experienceterm and its interactions with artwork and other creative outputs (Kaufman, 2013 ; McCrae, 2007 ; McCrae & Greenberg, 2014 ), with Silvia et al. ( 2015 ) considering it “an essentially aesthetic trait” (p. 376). In addition, openness has been shown to influence judgements of artterm depending on painting typeterm·var, such that people who are more open to experience tend to have an increased appreciation for abstractterm artterm compared to less-open participants (Feist & Brady, 2004 ). Other personality traits explored in aesthetics and creativityterm research include empathy and embodied cognitions (Freedberg & Gallese, 2007 ; Rusu, 2017 ), and beliefs of creative mindsets (Hass et al., 2016 ; Karwowski, 2014 ). To provide greater clarity on the possible influence of individual-differences measures on appraisals of artwork, in Study 1, we assessed age and scores on the cognitive reflection test (CRT), and in Study 2, age, CRT scores, openness to experienceterm, personal attitudes toward AI, empathy, growth mindsets, and fixed mindsets (below, we outline our rationale for inclusion of these individual-differences measures).

Study 1 Method Study 1 was approved by the University of Waterloo Research Ethics Board (31067). All data and stimuli can be found at https://osf.io/cgw8v/ .

Participants One-hundred and fifty participants, each with at least 100 approved human intelligence tasks and an approval rating of above 90%, were recruited through Amazon’s Mechanical Turk. Power analyses were not performed a priori. Instead, we reviewed the most closely related studies we could find, finding that, in Study 1 of Chamberlain et al. ( 2018 ), 65 participants completed the study. To ensure our study was well-powered, we decided to nearly triple this sample size. Participants were told at sign-up that they would be evaluating pieces of artterm and filling out questionnaires meant to probe the way they think. Participants were directed to Qualtrics to complete the study, and then were debriefed following completion of the study and were compensated for their time. One participant was excluded for bad data (i.e., they used the same response throughout the entire study), resulting in 149 participants ( M age = 42.35, SD = 11.59; female = 65).

Materials Thirty AI-created paintings, considered by the authors of this article to be of high quality, were taken from ArtBreeder, which is a machine learning website that produces artterm (see Fig. 1 for sample images). The images were open-source and pre-existing and were not created by the authors. While it might seem most reasonable to obtain artworks from both AI and humans (and not deceive participants), one possible problem with this approach is that there could be error introduced into the selection process. For instance, although one could do their best to ensure that any selected human and AI artworks are comparable, without first conducting a norming study first (across all of our measures), this outcome could not be verified. For this reason, we opted to present a stimulus set that was exclusively AI-created. Fig. 1 Sample images (representationalterm and abstractterm, respectively) taken from Artbreeder. See https://osf.io/cgw8v/ for all stimuli Of the 30 AI-created paintings, half (15) were representationalterm (i.e., reflected an easily recognizable figure or object), and half (15) were abstractterm (i.e., they included partially or completely unrecognizable referents), as rated by the first author. This distinction was made in the present study because past results have repeatedly revealed that people tend to prefer representationalterm over abstractterm paintings (Heinrichs & Cupchik, 1985 ; Kettlewell et al., 1990 ; Knapp & Wulff, 1963 ; Mastandrea et al., 2011 , 2021 ). Here, we not only wanted to attempt to replicate this finding, but also to determine whether evaluations of different painting typesterm·var (representationalterm and abstractterm) differ as a function of whether the purported creatorterm was a human or AI. Research on AI-art perceptions has primarily used abstractterm paintings as stimuli, which could limit our understanding of the true role of painting typeterm·var in aesthetic judgementsterm (Chiarella et al., 2022 ; Israfilzade, 2020 ). However, Chamberlain et al. ( 2018 ) presented varied painting typesterm·var but found no interaction between painting creator and painting typeterm·var on general aesthetic value ratings; with our extension of rating criteria, potentially different processes of judgements that may rely both on painting typeterm·var and creator could be illuminated. Relatedly, Gangadharbatla ( 2022 ) found an increased willingness for people to believe that abstractterm artterm tends to be created by AI (compared to humans), whereas Chamberlain et al. ( 2018 ) found an increased willingness for people to believe that representationalterm artterm tends to be created by humans. Thus, it seems that painting typeterm·var does bear some relationship to the creator of the painting; to date, however, this is unclear.

Importantly, each image was presented in random order and had individual-level randomization of a labelterm of “human-created” or “AI-created.” In other words, different labelsterm were assigned to different images across participants. Thus, on average, participants were presented 15 images with an AI labelterm, and 15 images with a human labelterm, even though all images were in fact AI-created. 1 Participants were asked to rate each image on the following criteria: “How much do you like this image?” (Likingterm), “How beautiful/aesthetically pleasing is this image?” (Beautyterm), “How profound or meaningful is this image?” (Profundityterm), and “How much money would this work be worthterm?” (Worthterm). All criteria were answered on a 1–5 Likert scale [“Not at all”… “Very much”], except for Worthterm, which was answered on a 1–5 Likert scale with possible responses [“None at all”… “Worthterm quite a lot”]. In addition, for each image, participants were asked “Based on the labelterm above, was this image created by a human or an artificial intelligence computer program?” (Label-Check). This question was used as an attention check. Incorrect responses to the Label-Check were excluded from data analysis, resulting in the removal of 113 trials (out of 4470 total trials across the 149 participants). No participants themselves were removed on the basis of this attention check, only trials.

Before rating the images, participants were asked to complete the 7-item cognitive reflection test (CRT; Frederick, 2005 ; Toplak et al., 2014 ). This extended CRT, which has been shown to have high internal consistency ( ⍺ = 0.74), consists of seven numeracy questions and is administered to assess an individual’s cognitive ability to override instinctual responses for a correct answer to the problem (Campitelli & Gerrans, 2014 ). Thus, the CRT considers individual differences in quantitative skills and bias-overriding. Our rationale for including the CRT was to determine whether CRT scores predict different ratings across human- and AI-created artwork: If we do observe an anti-AI biasterm—with more-positive judgements for human artworks than AI artworks—then one possibility is that individuals who have a greater ability to override their intuitions may show less of a bias, given that this bias may be the result of an automatic (intuitive) response to devalue artworks created by AI. Notably, both CRT and age were included as exploratory variables.

Procedure Participants first provided informed consent following a description of the nature of the study. Next, they were given two bot checks (zero participants failed these checks, and all of them therefore moved forward with the study). Participants were then directed to complete demographic information, followed by the CRT. When finished with these tasks, participants received instructions for the remaining portion of the survey, which asked them to rate the 30 AI-created paintings on Likingterm, Beautyterm, Profundityterm, and Worthterm, with randomized labelsterm of AI- or human-created. Following ratings, participants were debriefed on the purpose of the study, once again provided consent for their data to be used for analyses, and were redirected for compensation.

Study 1 results Paired t -tests with Bonferroni corrections show increased ratings for human-labelled over AI-labelled artterm for all four criteria: Likingterm ( t (148) = 2.644, p = 0.036, d = 0.17), Beautyterm ( t (148) = 3.499, p = 0.002, d = 0.22), Profundityterm ( t (148) = 7.725, p < 0.001, d = 0.47), and Worthterm ( t (148) = 10.042, p < 0.001, d = 0.61) (Fig. 2 ). In addition, for each criterion, a difference score between average human-labelled artterm rating and average AI-labelled artterm rating was calculated for each participant and was plotted against the participant’s CRT score. Only a relationship between CRT and Beautyterm difference scores emerged significant, with lower CRT scores associated with higher Beautyterm difference scores (that is, higher average Beautyterm judgements were given for human-labelled than AI-labelled artterm; r = − 0.17, p = 0.042). 2 No such significant relationships emerged elsewhere (Likingterm: p = 0.07, Profundityterm: p = 0.95, Worthterm: p = 0.37). The same difference scores were also plotted against age; again, no significant relationships emerged ( p = 0.29, p = 0.20, p = 0.26, p = 0.59). A significant relationship emerged between age and CRT performance ( r = 0.23, p = 0.004), such that older participants performed better on the CRT. Fig. 2 Distribution of values on artworks associated with AI and Human for each criterion (Likingterm, Beautyterm, Profundityterm, Worthterm). Error bars represent standard error of the mean. Following multiple-comparison corrections, * p < 0.05, ** p < 0.01, *** p < 0.001 Lastly, exploratory linear mixed-effects models using maximum likelihood estimation were performed for each of the four main criteria as the outcome variables: Likingterm, Beautyterm, Profundityterm, and Worthterm. Across models, participants and paintings were random effects, resulting in crossed random effects models. For all models, our manipulated variables of the artterm’s Labelterm (AI or Human) and Painting Typeterm·var (abstractterm or representationalterm) were included, and we tested a potential interaction between Labelterm and Painting Typeterm·var. Across all criteria, a significant main effect of Painting Typeterm·var emerged, such that participants preferred representationalterm over abstractterm paintings regardless of Labelterm ( p ’s < 0.001). For Likingterm, a non-significant main effect of Labelterm ( p = 0.057) was qualified by a significant interaction between Painting Typeterm·var and Labelterm ( p = 0.041). Additionally, for Profundityterm, a main effect of Labelterm ( p < 0.001) was qualified by a significant interaction between Painting Typeterm·var and Labelterm ( p = 0.036). When a painting is representationalterm, artterm is liked more and found more profound when it is created by a human as opposed to AI. No significant interaction between Painting Typeterm·var and Labelterm emerged in Beautyterm ( p = 0.21) or Worthterm ( p = 0.18).

Study 1 discussion Across all four rating criteria (Likingterm, Beautyterm, Profundityterm, and Worthterm), we found an anti-AI biasterm, with participants showing a preference for artterm labelled as “human-created” compared to “AI-created.” This finding is particularly important given the within-subjects design of our study, its relatively large sample size, range of rating criteria, range of painting typesterm·var, and randomization of labelsterm across a large set of paintings.

Study 1 primarily served as a test of whether an anti-AI-art bias exists across a range of aesthetic judgementterm criteria. Given our relatively large sample, we were confident in the findings of this study, which, on the whole, mirrored findings of Chamberlain et al. ( 2018 ). However, extending these results, here we showed that this pattern of judgements for AI-art biases is reflected across appraisal processes outside of general aesthetic preferences. For instance, not only is AI-art liked less, but it is also viewed as less worthy and less profound, which may have interesting implications for the ways in which people will consume AI-art in the future.

In addition, though some studies also used a randomization of labelsterm (as in the present study; e.g., Chiarella et al., 2022 ), our study was unique in its use of only AI-created images, across many more stimuli, in a within-subjects design. Statistically, this ensured higher power and increased confidence in our findings. In other words, participants were given the opportunity to judge both human- and AI-labelled artworks, and this permitted a sounder comparison of participants’ judgements across the two sources. Perhaps most importantly, though, was that all artworks were from AI in actuality. This underscores the fact that participants are generally incapable of noticing what source created what paintings—as supported by Gangadharbatla ( 2022 ) and Chamberlain et al. ( 2018 )—which ultimately reflects the noteworthy quality of AI-art today. Though we did not explicitly ask participants for their confidence in the accuracy of our labelsterm, our decision to exclusively select AI-created stimuli enabled us to dissociate top-downterm from bottom-upterm processing in these aesthetic judgementsterm. That is, actual human versus actual AI images may differ in certain visual, bottom-upterm qualities. 3 Thus, by exclusively selecting AI artwork and randomizing labelsterm on a random subset of paintings, we ensured that any observed effects were necessarily top-downterm and isolated to source manipulation, as opposed to bottom-upterm, or driven by potential inherent differences across human- and AI-created artworks.

Somewhat surprising was the lack of statistically significant associations between our individual-differences measures (CRT and age) with the difference scores of human-AI ratings for each participant, which might be expected to track an AI bias. Although lower cognitive reflection scores were associated with higher Beautyterm preferences of human than AI artterm, no other significant relationship emerged between our individual-differences measures and the human-AI difference scores. Past work with the CRT has found negative relationships between CRT scores and perceived profundityterm of randomly generated statements (Pennycook et al., 2015 ), implying a relationship between cognitive reflection traits and subjective judgements of profundityterm that could extend to aesthetics. However, in Study 1, we found no support for this relationship. Given that the CRT assesses intuition-overriding in mathematical domains, we can conclude that (1) anti-AI biasesterm largely do not rely on quantitative skills, (2) the anti-AI biasterm is not an intuitive response as probed by the mathematical problems in the CRT, and/or (3) we need more-sophisticated statistical analyses and/or power to uncover the true underlying relationships (if any) among these measures. In addition, the lack of statistically significant relationships between age and judgements of artterm was somewhat surprising. Though exploratory, this is at odds with some previous aesthetic studies that have reported age-dependent differences in judgements of artterm. Specifically, Mockros ( 1993 ) reported that general aesthetic judgementsterm were rated higher by novice professional adults than novice undergraduates. More research is warranted on this topic given that we found no such age effect, which perhaps implies that younger participants—with their lives being more heavily dominated by AI than older participants—do not have different views on the creative products of AI than those of older adults.

Study 2 The results from Study 1 provide initial insight into participant judgements of artterm that is believed to be generated either by AI algorithms or by humans. While not reliably predictable by individual differences such as age or CRT scores, a bias against AI artterm emerged. This bias, however, clearly depended on the criterion used to assess the artwork. Certainly, individuals do not just consider measures of Likingterm, Beautyterm, Profundityterm, and Worthterm when assessing artwork, but a host of other engagement processes.

This nuance to a general “anti-AI biasterm” matches with contemporary aesthetics models that argue that aesthetic judgementsterm are rather complex. Indeed, largely pioneered by David Berlyne’s new experimental aesthetics ( 1960 , 1971 ), artterm appreciation has been argued to be a consequence of elaborative appraisals of criteria like novelty, ambiguity, and complexity (which he deemed “collativeterm” properties; Cupchik & Berlyn, 1979 ). Other emerging models have since provided more nuance to Berlyne’s traditional behaviorist approach (see Berlyne, 1975 ; Silvia, 2005 ), including more cognitive principles, and an increased understanding of the processes people utilize to engage with artterm. These models all agree in their proposal that humans engage in a multi-process evaluation of artterm, considering both sensory and non-sensory aspects (e.g., the Aesthetic Triadterm, Chatterjee & Vartanian, 2016 ; the pleasure-interest model, Graf & Landwehr, 2015 ; the information-processing model, Leder et al., 2004 ). Accordingly, many aesthetic properties—some more literal and surface-level, some non-sensory and more elaborative—can be probed during interactions with artterm. For instance, findings from Chamberlain et al. ( 2018 ) show that participants specifically cite that brush-strokes and other surface-level properties of the artistic process influence their evaluations of the final product. Moreover, when viewing anthropomorphized videos of robots painting, participants show greater appreciation for computer-generated artterm than when no anthropomorphized video of the robot is provided. This mirrors findings by Hong et al. ( 2022 ), who found that anthropomorphization of an AI-music algorithm led to higher acceptance of the algorithm as a true “musician,” which in turn led to increased aesthetic appreciation. Collectively, these studies suggest that there may be an implicit role of effortfulness or embodiment behind peoples’ evaluations of artterm. More specifically, it may be that peoples’ preferences for artterm increase as their beliefs about the amount of effortterm·var that went into creating a piece of artterm increase.

As informed by multi-level models as above, including additional judgement criteria could help us to shed further light on the reasons as to why people are making the aesthetic judgementsterm they’re making. Thus, in addition to Likingterm, Beautyterm, Profundityterm, and Worthterm, in Study 2 we also included the following judgement criteria: emotionality (Emotion), perceived narrativityterm (Storyterm), personal meaning (Meaningful), perceived effortterm·var (Effortterm·var), and estimated time to create (Time) as additional criteria for judgement. Importantly, while these now nine criteria may or may not behave in similar ways to one another, some may moderate ratings of others. In this way, we could gain insight on how levels of processing act interpedently or influence one another in this new sphere of aesthetic judgementsterm with artterm created by non-humans. We specifically included these criteria given emotion’s prominent role in aesthetics and creativityterm (Silvia, 2005 ), the use of narratives as engines of consumption in other forms of artterm (a growing literature in music cognition investigates narratives in response to music, e.g., Margulis et al., 2019 ; McAuley et al., 2021 ), the act of deriving meaning from artwork that is often posited in models of aesthetics (e.g., Leder et al., 2004 ; Pelowski et al., 2017 ), and that effortterm·var and time involved in creation of a product is often a heuristic for quality (Kruger et al., 2004 ).

In addition, we extended the battery of individual-differences measures given our mostly null results from the individual-differences measures in Study 1. While we choose to keep both age and CRT scores as potential individual differences, we also included empathy skills (to determine if participants differ in judgement based on the ability to empathize with other agents, including AI, perhaps explaining AI-art judgements), openness to experienceterm (given its role in aesthetic encounters; Kaufman, 2013 ; McCrae, 2007 ; McCrae & Greenberg, 2014 ; Silvia et al., 2015 ), personal attitudes toward AI (which we hypothesized could predict judgements of artwork made by AI), and facets of the creative mindset scale (CMS; Karwowski, 2014 ), which include growth and fixed mindsets. The CMS specifically asks about views of who can produce products of creativityterm, and we thus deemed it important to include in this study to determine if growth mindsets lead to higher AI-art ratings as we hypothesized. In addition, to detect more sensitive relationships between both individual-differences measures and judgement criteria, we aimed to use more-sophisticated statistical modeling with a pre-registered design. In sum, through implementing linear mixed models, we sought to determine whether we would replicate and explain the anti-AI effect with wider criteria of judgements—to parallel common multi-level processing models of aesthetics—and an extension of individual-differences measures.

Method Study 2’s design and analysis plan were pre-registered. Study 2 was approved by the Duke University Campus Institutional Review Board (2022-0534). All data and stimuli, and the pre-registration, can be found at https://osf.io/cgw8v/ .

Participants One-hundred and fifty-one participants were recruited through Prolific. 4 Only native English speakers with at least a 90% approval rate were permitted to enroll in the study. Like Study 1, they were told at sign-up that they would be evaluating pieces of artterm and filling out questionnaires meant to probe the way they think. Participants were debriefed following completion of the study and were compensated for their time. As per our pre-registration, data from two participants were excluded for failing an attention check (age and birth-year needed to correspond appropriately), and one participant was excluded for bad data, resulting in 148 participants ( M age = 36.66, SD = 12.73; female = 74, non-binary = 2).

Materials The same 30 images from Study 1 were presented in random order in Study 2, again with a randomly assigned labelterm of “human-created” or “AI-created.” Ratings were done on the same criteria: Likingterm, Beautyterm, Profundityterm, Worthterm; note that the phrasing for the Profundityterm question in this study changed from “How profound or meaningful is this image?” to “How profound is this artwork?” for simplicity. Participants rated the images on five additional criteria as well: “To what extent does this artwork elicit an emotional response in you?” (Emotion), “To what extent can you imagine a storyterm communicated through this artwork?” (Storyterm), “To what extent do you find this artwork personally meaningful?” (Meaningful), “How much effortterm·var do you believe was involved in making this artwork?” (Effortterm·var), and “How much time do you believe it took to create this piece?” (Time). All questions were on a 1–5 Likert [“None/not at all”… “Very much”] except Time, which was a free-response question that allowed for numeric entry in a “Days,” “Hours,” and “Minutes” field. Participants were told to simply enter “0” if a certain field did not apply. Lastly, for each image, participants were asked the Label-Check question from Study 1 for trial-based exclusions: “Based on the labelterm above this image, was this image created by a human or an artificial intelligence computer program?” Those who did not provide the correct answer had that trial eliminated, resulting in 114 trials (of 4440 total trials across the 148 participants) excluded from analyses. As in Study 1, no participants were removed on the basis of this attention check, only trials. This specific exclusion criterion was not specified in our original pre-registration plan, though, for purposes of consistency, it follows the procedure from Study 1.

Before the ratings, participants filled out various questionnaires. Participants completed the same CRT as in Study 1. In addition, to probe trait-level empathy skills, participants completed the Toronto Empathy Questionnaire (TEQ; Spreng et al., 2009 ). The TEQ has been demonstrated to be a highly internally consistent ( ⍺ = 0.87) and reliable ( r = 0.81) self-report empathy measure. The 16-item questionnaire was completed on a 1–5 Likert [“Never”… “Always”], and responses were reverse-scored when appropriate and summed for a final empathy score.

Participants also completed the General Attitudes toward Artificial Intelligence Scale (GAAIS) to assess positive and negative sentiments of AI in general (Schepman & Rodway, 2020 ). The 20-item GAAIS has been fully validated and found to report a consistent positive subscale ( ⍺ = 0.88) and negative subscale ( ⍺ = 0.82), with strong convergent validity with other AI-related measures (Schepman & Rodway, 2022 ). Scores for each subscore were aggregated through summation, producing a “Positive Attitude” score and a “Negative Attitude” score.

To probe participant beliefs about creativityterm, we administered the Creative Mindset Scale (CMS: Karwowski, 2014 ). The CMS assesses how much an individual believes creativityterm can be developed and trained through time (i.e., “growth” mindset) and how much they believe creativityterm is static and innate (i.e., “fixed” mindset). Participants rated the 10 items of the CMS on a 1–5 Likert Scale [“Definitely not”… “Definitely yes”]. Scores for each mindset were aggregated through summation, producing a “Growth” value and a “Fixed” value.

Participants also completed the 12-item “Openness to Experienceterm” subset of the NEO-Five Factor Personality Inventory. Though the full NEO-FFI can also assess Consciousness, Extraversion, Agreeableness, and Neuroticism, we chose to implement only the Openness subset given its previous specific involvement in aesthetic encounters (Kaufman, 2013 ; McCrae, 2007 ; McCrae & Greenberg, 2014 ; Silvia et al., 2015 ), and to reduce study length. Items were answered on a 1–5 Likert scale [“Strongly disagree”… “Strongly agree”] and were reverse-scored when appropriate and aggregated through summing.

Lastly, in an effortterm·var to guide future research, we had participants answer a series of exploratory, free-response and single-item questions. These data are listed in our pre-registration as “exploratory,” and, for the sake of clarity and brevity, they are not reported here.

Procedure Upon direction to the Qualtrics survey, participants gave informed consent on the nature of the study. Then, they were given two bot checks; no participant failed, and all moved forward with the study. Participants were then directed to complete demographic information, followed by the CRT, TEQ, openness from the NEO-FFI, CMS, GAAIS, and additional single-item questions, in that order. Importantly, participants were specifically asked their age in the demographic questionnaire, and at the end of the study, were asked their year of birth; this served as an attention check. As per our pre-registration, data from two participants with incongruent birth years and ages were excluded from our analyses. Following questionnaires, participants completed the ratings of the images, after which they were debriefed and redirected to Prolific.

Statistical analyses Linear mixed-effects models using maximum likelihood estimation were estimated in R (version 4.2.0; R Core Team, 2022) using the {lme4} package (version 1.1.33; Bates et al., 2015 ). A mixed effect model was estimated for each of the four main criteria from Study 1 as the outcome variables: Likingterm, Beautyterm, Profundityterm, and Worthterm. The linear mixed-effects models diverged from our pre-registration with the use of frequentist rather than Bayesian models. 5 Power analyses were not performed a priori; instead, the sample size was determined using the same procedure reported in our Study 1.

For all models, our manipulated variables of the artterm’s Labelterm (AI or Human) and Painting Typeterm·var (abstractterm or representationalterm) were included. Both Labelterm and Painting Typeterm·var were fixed-effect variables. We did not include the interaction between Labelterm and Painting Typeterm·var in our reports because this term was not significant in any of the models. Across models, participants and paintings were random effects, resulting in crossed random effects models. For participants, we included the random intercept and Labelterm slope to model people’s differences in outcome related to Labelterm. For paintings, we only included the random intercept term to account for differences in the level of the outcome for each painting.

For Model 1, we focused on individual-differences measures to identify characteristics of who were related to our outcome variables. All individual-differences measures were continuous (i.e., sum scores) fixed effect variables. Each individual-difference measure was mean-centered to increase the interpretability of the model (e.g., each intercept represented the mean of each outcome). Interactions between Labelterm and all individual-differences measures were specified in this model.

For Model 2, we added the five additional criteria (Emotion, Storyterm, Meaningful, Effortterm·var, and Time), as well as their respective interactions with Labelterm, as predictors to the model. This model represented aspects of why outcomes might be rated differently. Time was log-transformed due to its logarithmic distribution. All additional criteria were level-1 or within-painting variables. Since Model 2 represents the full model, the equation is presented here: \[\begin{aligned} {\varvec{L}}1:{Y}_{ij}&= {\beta }_{0j}+ {\beta }_{1j}PaintingTyp{e}_{ij}+ {\beta }_{2j}Labe{l}_{ij}+ {\beta }_{3j}Opennes{s}_{ij}+ {\beta }_{4j}PositiveA{I}_{ij}+ {\beta }_{5j}NegativeA{I}_{ij}\\&\quad+{\beta }_{6j}GrowthMindse{t}_{ij}+ {\beta }_{7j}FixedMindse{t}_{ij}+ {\beta }_{8j}CorrectCR{T}_{ij}+ {\beta }_{9j}Ag{e}_{ij}\\&\quad+ {\beta }_{10j}Emotio{n}_{ij}+ {\beta }_{11j}Stor{y}_{ij}+ {\beta }_{12j}Meaningfu{l}_{ij}+ {\beta }_{13j}Effor{t}_{ij}+ {\beta }_{14j}\mathrm{log}Tim{e}_{ij}\\&\quad+{\beta }_{15j}Labe{l}_{ij}\times Opennes{s}_{ij}+{\beta }_{16j}Labe{l}_{ij}\times PositiveA{I}_{ij}+{\beta }_{17j}Labe{l}_{ij}\times NegativeA{I}_{ij}+{\beta }_{18j}Labe{l}_{ij}\times GrowthMindse{t}_{ij}+{\beta }_{19j}Labe{l}_{ij}\times FixedMindse{t}_{ij}\\&\quad+{\beta }_{20j}Labe{l}_{ij}\times CorrectCR{T}_{ij}+ {\beta }_{21j}Labe{l}_{ij}\times Ag{e}_{ij}+ {\beta }_{22j}Labe{l}_{ij}\times Emotio{n}_{ij}+{\beta }_{23j}Labe{l}_{ij}\times Stor{y}_{ij}\\&\quad+ {\beta }_{24j}Labe{l}_{ij}\times Meaningfu{l}_{ij}+ {\beta }_{25j}Labe{l}_{ij}\times Effor{t}_{ij}+ {\beta }_{26j}Labe{l}_{ij}\times \mathrm{log}Tim{e}_{ij}\\&\quad+{Participan{t}_{oj}}_{i}+{Painting}_{1j}+{\epsilon }_{ij}\end{aligned}\] L 1 : Y ij = β 0 j + β 1 j P a i n t i n g T y p e ij + β 2 j L a b e l ij + β 3 j O p e n n e s s ij + β 4 j P o s i t i v e A I ij + β 5 j N e g a t i v e A I ij + β 6 j G r o w t h M i n d s e t ij + β 7 j F i x e d M i n d s e t ij + β 8 j C o r r e c t C R T ij + β 9 j A g e ij + β 10 j E m o t i o n ij + β 11 j S t o r y ij + β 12 j M e a n i n g f u l ij + β 13 j E f f o r t ij + β 14 j log T i m e ij + β 15 j L a b e l ij × O p e n n e s s ij + β 16 j L a b e l ij × P o s i t i v e A I ij + β 17 j L a b e l ij × N e g a t i v e A I ij + β 18 j L a b e l ij × G r o w t h M i n d s e t ij + β 19 j L a b e l ij × F i x e d M i n d s e t ij + β 20 j L a b e l ij × C o r r e c t C R T ij + β 21 j L a b e l ij × A g e ij + β 22 j L a b e l ij × E m o t i o n ij + β 23 j L a b e l ij × S t o r y ij + β 24 j L a b e l ij × M e a n i n g f u l ij + β 25 j L a b e l ij × E f f o r t ij + β 26 j L a b e l ij × log T i m e ij + P a r t i c i p a n t oj i + Painting 1 j + ϵ ij L 2 : P a r t i c i p a n t 0 j = γ 00 + γ 01 L a b e l oj + u 0 j L 2 : P a i n t i n g 1 j = γ 10 + u 1 j For Model 3, we included only the significant predictors from Model 2. For each model, we report model fit using Akaike’s information criterion (AIC) and variance explained using marginal R 2 (fixed effects only), intraclass correlation (ICC; random effects only), and conditional R 2 (all effects combined). For brevity, we focus our report on Model 3 but provide all models’ results in Tables 1 – 4 to be comprehensive.

Study 2 results Likingterm Results for the Likingterm models are reported in Table 1 . There were two interactions involving Labelterm: Storyterm ( b = − 0.09, p < 0.001) and Effortterm·var ( b = 0.12, p < 0.001). For Storyterm, Likingterm was greater when Labelterm was “AI” than when Labelterm was “human” and this difference increased as Storyterm increased (Fig. 3 ). Based on Fig. 3 and the difference between the conditional means, for low-to-moderate values of Effortterm·var (ratings = 1–4; Fig. 3 ), Likingterm was greater when Labelterm was “AI” than when Labelterm was “human.” For high values of Effortterm·var (rating = 5), Likingterm was greater when Labelterm was “human” than when Labelterm was “AI.” There were main effects for Emotion ( b = 0.32, p < 0.001), Meaningful ( b = 0.24, p < 0.001), and Painting Typeterm·var ( b = 0.28, p < 0.001). For Emotion and Meaningful, Likingterm increased as these criteria increased. For Painting Typeterm·var, Likingterm was greater for representationalterm paintings than abstractterm paintings. The fixed effects of the model explained 57.9% of the variance, the random effects explained 32.0% of the variance, and all effects explained 71.2% of the variance. Table 1 Estimates and associated CIs predicting “Likingterm” judgements Likingterm Model 1 Model 2 Model 3 Fixed effects [95% CI] (Intercept) 2.38 [2.17, 2.60] 0.55 [0.38, 0.71] 0.57 [0.41, 0.73] Labelterm (Human = 1) 0.24 [0.15, 0.32] − 0.26 [− 0.45, − 0.07] − 0.28 [− 0.43, − 0.13] Painting Typeterm·var (Representationalterm = 1) 0.75 [0.48, 1.01] 0.28 [0.13, 0.44] 0.28 [0.13, 0.44] Openness 0.01 [− 0.01, 0.03] 0.01 [− 0.00, 0.03] – Positive AI 0.03 [0.01, 0.04] 0.00 [− 0.01, 0.01] – Negative AI 0.01 [− 0.01, 0.03] 0.01 [− 0.00, 0.02] – Growth 0.06 [0.01, 0.11] 0.01 [− 0.03, 0.04] – Fixed 0.06 [0.02, 0.09] 0.02 [− 0.01, 0.04] – Empathy 0.00 [− 0.01, 0.02] − 0.00 [− 0.01, 0.01] – CRT − 0.04 [− 0.09, 0.01] 0.01 [− 0.03, 0.04] – Age − 0.00 [− 0.01, 0.01] − 0.01 [− 0.01, 0.00] – Emotion – 0.34 [0.29, 0.39] 0.32 [0.29, 0.36] Storyterm – 0.23 [0.19, 0.27] 0.23 [0.19, 0.27] Meaningful – 0.20 [0.14, 0.25] 0.24 [0.20, 0.27] Effortterm·var – 0.11 [0.06, 0.15] 0.11 [0.07, 0.15] Time (log) – 0.01 [− 0.00, 0.03] – Labelterm × Openness 0.00 [− 0.01, 0.02] − 0.01 [− 0.02, 0.00] – Labelterm × Positive AI − 0.02 [− 0.03, − 0.00] − 0.01 [− 0.01, 0.00] – Labelterm × Negative AI − 0.01 [− 0.02, 0.01] 0.00 [− 0.01, 0.01] – Labelterm × Growth 0.03 [− 0.01, 0.07] 0.01 [− 0.02, 0.04] – Labelterm × Fixed 0.01 [− 0.02, 0.04] 0.01 [− 0.01, 0.03] – Labelterm × Empathy 0.00 [− 0.01, 0.01] 0.00 [− 0.01, 0.01] – Labelterm × CRT − 0.03 [− 0.06, 0.01] 0.00 [− 0.03, 0.03] – Labelterm × Age − 0.00 [− 0.01, 0.01] − 0.00 [− 0.01, 0.00] – Labelterm × Emotion – − 0.04 [− 0.10, 0.03] – Labelterm × Storyterm – − 0.10 [− 0.15, − 0.04] − 0.09 [− 0.14, − 0.05] Labelterm × Meaningful – 0.06 [− 0.00, 0.13] – Labelterm × Effortterm·var – 0.12 [0.06, 0.18] 0.12 [0.07, 0.17] Labelterm × Time (log) – − 0.01 [− 0.04, 0.01] – Random effects Participant (Intercept) 0.44 0.19 0.21 Labelterm (Slope) 0.15 0.05 0.06 Painting (Intercept) 0.13 0.04 0.04 Residual 0.95 0.53 0.53 Model Marginal 0.16 0.59 0.58 ICC 0.37 0.30 0.32 Conditional 0.47 0.72 0.71 AIC 12,819.13 10,300.48 10,126.86 Note . CI = confidence interval, Marginal = variance explained by fixed effects, ICC = intraclass correlation or variance explained by random effects, Conditional = variance explained by fixed and random effects, AIC = Akaike Information Criterion. Italicized text = p < 0.05, bolded text = p < 0.001 Fig. 3 Interaction plots showing moderations by Effortterm·var and Storyterm on Likingterm and Beautyterm, two sensory-level criteria. Confidence bands represent the 95% confidence interval Beautyterm Results for the Beautyterm models are reported in Table 2 . Similar to Likingterm, there were interactions involving Labelterm: Storyterm ( b = − 0.08, p < 0.001) and Effortterm·var ( b = 0.10, p < 0.001). The effects similarly parallel Likingterm in that Beautyterm was greater when Labelterm was “AI” than when Labelterm was “human” and this difference increased as Storyterm increased and for low-to-moderate values of Effortterm·var (ratings = 1–4; Fig. 3 ), Beautyterm was greater when Labelterm was “AI” than when Labelterm was “human” whereas the reverse was found for high values of Effortterm·var (rating = 5, Fig. 3 ). There was an additional interaction between Labelterm and Openness to Experienceterm ( b = − 0.01, p = 0.003) such that well below average Openness to Experienceterm (− 20) had higher Beautyterm for the human Labelterm than the AI Labelterm but this difference reversed with below average (− 10) Openness to Experienceterm and higher such that Beautyterm was higher for the AI Labelterm than the human Labelterm. Similar to Likingterm, there were main effects for Emotion ( b = 0.23, p < 0.001), Meaningful ( b = 0.23, p < 0.001), and Painting Typeterm·var ( b = 0.32, p < 0.001) such that Beautyterm increased as Emotion and Meaningful increased and Beautyterm was greater for representationalterm paintings than abstractterm paintings. There was an additional main effect of Age ( b = − 0.01, p = 0.007) such that people younger than average (-20) rated Beautyterm higher than people older than average (40). The fixed effects of the model explained 53.1% of the variance, the random effects explained 34.0% of the variance, and all effects explained 69.0% of the variance. Table 2 Estimates and associated CIs predicting “Beautyterm” judgements Beautyterm Model 1 Model 2 Model 3 Fixed effects [95% CI] (Intercept) 2.43 [2.20, 2.67] 0.75 [0.56, 0.95] 0.74 [0.55, 0.93] Labelterm (Human = 1) 0.23 [0.15, 0.31] − 0.37 [− 0.57, − 0.18] − 0.28 [− 0.43, − 0.13] Painting Typeterm·var (Representationalterm = 1) 0.75 [0.46, 1.05] 0.32 [0.11, 0.52] 0.32 [0.11, 0.52] Openness 0.01 [− 0.01, 0.03] 0.02 [0.00, 0.03] 0.01 [− 0.00, 0.02] Positive AI 0.02 [0.01, 0.03] 0.00 [− 0.01, 0.01] – Negative AI 0.01 [− 0.01, 0.03] 0.01 [− 0.00, 0.03] – Growth 0.04 [− 0.01, 0.09] − 0.01 [− 0.05, 0.02] – Fixed 0.05 [0.01, 0.08] 0.01 [− 0.02, 0.04] – Empathy − 0.00 [− 0.02, 0.01] − 0.01 [− 0.02, 0.00] – CRT − 0.04 [− 0.06, 0.04] 0.03 [− 0.00, 0.07] – Age − 0.01 [− 0.02, 0.01] − 0.01 [− 0.02, − 0.00] − 0.01 [− 0.01, − 0.00] Emotion – 0.24 [0.19, 0.29] 0.23 [0.19, 0.26] Storyterm – 0.24 [0.20, 0.28] 0.24 [0.20, 0.28] Meaningful – 0.21 [0.15, 0.26] 0.23 [0.20, 0.27] Effortterm·var – 0.14 [0.10, 0.19] 0.14 [0.10, 0.18] Time (log) – − 0.00 [− 0.02, 0.02] – Labelterm × Openness − 0.01 [− 0.02, 0.01] − 0.01 [− 0.02, 0.00] − 0.01 [− 0.02, − 0.00] Labelterm × Positive AI − 0.01 [− 0.03, − 0.00] − 0.00 [− 0.01, 0.00] – Labelterm × Negative AI − 0.01 [− 0.02, 0.01] − 0.00 [− 0.01, 0.01] – Labelterm × Growth 0.02 [− 0.02, 0.05] 0.00 [− 0.03, 0.03] – Labelterm × Fixed − 0.01 [− 0.03, 0.02] − 0.01 [− 0.03, 0.01] – Labelterm × Empathy 0.00 [− 0.01, 0.01] − 0.00 [− 0.01, 0.01] – Labelterm × CRT − 0.03 [− 0.06, 0.01] 0.00 [− 0.03, 0.03] – Labelterm × Age − 0.00 [− 0.01, 0.01] 0.00 [− 0.00, 0.01] – Labelterm × Emotion – − 0.02 [− 0.09, 0.04] – Labelterm × Storyterm – − 0.19 [− 0.14, − 0.03] − 0.08 [− 0.12, − 0.04] Labelterm × Meaningful – 0.04 [− 0.02, 0.11] – Labelterm × Effortterm·var – 0.09 [0.03, 0.15] 0.10 [0.05, 0.16] Labelterm × Time (log) – 0.01 [− 0.01, 0.04] – Random effects Participant (Intercept) 0.44 0.21 0.24 Labelterm (Slope) 0.13 0.05 0.05 Painting (Intercept) 0.16 0.08 0.08 Residual 0.90 0.55 0.55 Model Marginal 0.13 0.53 0.53 ICC 0.40 0.33 0.34 Conditional 0.48 0.69 0.69 AIC 12,606.95 10,449.15 10,300.65 Note . CI = confidence interval, Marginal = variance explained by fixed effects, ICC = intraclass correlation or variance explained by random effects, Conditional = variance explained by fixed and random effects, AIC = Akaike Information Criterion. Italicized text = p < 0.05, bolded text = p < 0.001 Profundityterm Results for the Profundityterm models are reported in Table 3 . There was one interaction involving Labelterm and Positive AI attitudes ( b = − 0.01, p = 0.016) such that, based on Fig. 4 and the difference between the conditional means, there was slightly higher Profundityterm when Labelterm was “AI” and Positive AI attitudes were well below average (− 20; Fig. 4 ). This difference grew larger as Positive AI attitudes increased to well above average (20). There were main effects for all five additional criteria (Emotion: b = 0.22, p < 0.001; Storyterm: b = 0.19, p < 0.001; Meaningful: b = 0.26, p < 0.001; Effortterm·var: b = 0.19, p < 0.001; Time (log): b = 0.02, p = 0.011) such that Profundityterm increased as each criterion increased. For both growth ( b = 0.04, p = 0.005) and fixed ( b = 0.02, p = 0.017) mindsets, Profundityterm increased as these characteristics went from well below average (− 10 and − 8, respectively) to well above average (6 and 10, respectively). The fixed effects of the model explained 63.5% of the variance, the random effects explained 26.0% of the variance, and all effects explained 73.0% of the variance. Table 3 Estimates and associated CIs predicting “Profundityterm” judgements Profundityterm Model 1 Model 2 Model 3 Fixed effects [95% CI] (Intercept) 2.20 [2.02, 2.38] 0.42 [0.31, 0.53] 0.40 [0.31 0.50] Labelterm (Human = 1) 0.37 [0.29, 0.58] − 0.15 [− 0.31, 0.01] − 0.14 [− 0.20, − 0.07] Painting Typeterm·var (Representationalterm = 1) 0.40 [0.21, 0.58] − 0.04 [− 0.11, 0.02] – Openness − 0.01 [− 0.03, 0.01] − 0.00 [− 0.01, 0.01] – Positive AI 0.02 [0.01, 0.04] 0.01 [− 0.00, 0.01] 0.01 [− 0.00, 0.01] Negative AI 0.00 [− 0.02, 0.02] 0.00 [− 0.01, 0.01] – Growth 0.11 [0.05, 0.16] 0.04 [0.02, 0.07] 0.04 [0.01, 0.06] Fixed 0.07 [0.03, 0.11] 0.02 [0.00, 0.04] 0.02 [0.00, 0.04] Empathy 0.00 [− 0.01, 0.02] − 0.00 [− 0.01, 0.01] – CRT − 0.08 [− 0.13, − 0.02] − 0.03 [− 0.06, 0.00] – Age 0.01 [− 0.00, 0.02] 0.00 [− 0.00, 0.01] – Emotion – 0.22 [0.18, 0.26] 0.22 [0.19, 0.25] Storyterm – 0.18 [0.15, 0.21] 0.19 [0.17, 0.22] Meaningful – 0.26 [0.22, 0.31] 0.26 [0.22, 0.29] Effortterm·var – 0.20 [0.16, 0.24] 0.19 [0.16, 0.22] Time (log) – 0.02 [0.00, 0.03] 0.02 [0.00, 0.03] Labelterm × Openness 0.01 [− 0.01, 0.02] − 0.01 [− 0.01, 0.00] – Labelterm × Positive AI − 0.01 [− 0.03, − 0.01] − 0.01 [− 0.01, − 0.00] − 0.01 [− 0.01, − 0.00] Labelterm × Negative AI − 0.01 [− 0.02, 0.01] 0.00 [− 0.00, 0.01] – Labelterm × Growth − 0.00 [− 0.04, 0.03] − 0.01 [− 0.03, 0.01] – Labelterm × Fixed − 0.01 [− 0.03, 0.02] − 0.00 [− 0.02, 0.01] – Labelterm × Empathy 0.01 [− 0.00, 0.02] 0.00 [− 0.00, 0.01] – Labelterm × CRT − 0.02 [− 0.06, 0.01] 0.00 [− 0.02, 0.02] – Labelterm × Age − 0.00 [− 0.01, 0.00] − 0.00 [− 0.01, 0.00] – Labelterm × Emotion – − 0.00 [− 0.06, 0.06] – Labelterm × Storyterm – 0.03 [− 0.01, 0.08] – Labelterm × Meaningful – − 0.01 [− 0.07, 0.05] – Labelterm × Effortterm·var – − 0.02 [− 0.07, 0.04] – Labelterm × Time (log) – − 0.00 [− 0.02, 0.02] – Random effects Participant (Intercept) 0.53 0.13 0.13 Labelterm (Slope) 0.12 0.02 0.02 Painting (Intercept) 0.06 0.00 0.01 Residual 0.76 0.41 0.41 Model Marginal 0.15 0.64 0.64 ICC 0.44 0.25 0.26 Conditional 0.52 0.73 0.73 AIC 11,905.68 9098.20 8940.49 Note . CI = confidence interval, Marginal = variance explained by fixed effects, ICC = intraclass correlation or variance explained by random effects, Conditional = variance explained by fixed and random effects, AIC = Akaike Information Criterion. Italicized text = p < 0.05, bolded text = p < 0.001 Fig. 4 Interaction plot of Positive AI attitudes on communicativeterm judgements of Profundityterm and Worthterm. Confidence bands represent the 95% confidence interval Worthterm Results for the Worthterm models are reported in Table 4 . Similar to Profundityterm, there was one interaction involving Labelterm and Positive AI attitudes ( b = -0.01, p = 0.014) but it differed such that there was slightly higher Profundityterm when Labelterm was “human” and Positive AI attitudes were well below average (− 20; Fig. 4 ). This difference reversed as Positive AI attitudes went from well below average (− 20) to below average (− 10) and continued to diverge to well above average (20). Also similar to Profundityterm, there were main effects for all five additional criteria (Emotion: b = 0.18; Storyterm: b = 0.12; Meaningful: b = 0.15; Effortterm·var: b = 0.27; Time (log): b = 0.05; all p ’s < 0.001) such that Profundityterm increased as each criterion increased. The fixed effects of the model explained 57.4% of the variance, the random effects explained 32.0% of the variance, and all effects explained 71.3% of the variance. Table 4. Estimates and associated CIs predicting “Worthterm” judgements Worthterm Model 1 Model 2 Model 3 Fixed effects [95% CI] (Intercept) 2.29 [2.12, 2.45] 0.65 [0.53, 0.77] 0.66 [0.55, 0.77] Labelterm (Human = 1) 0.52 [0.43, 0.61] − 0.19 [− 0.37, − 0.02] − 0.21 [− 0.35, − 0.07] Painting Typeterm·var (Representationalterm = 1) 0.35 [0.18, 0.52] − 0.01 [− 0.09, 0.06] – Openness − 0.01 [− 0.03, 0.01] 0.01 [− 0.01, 0.02] – Positive AI 0.02 [0.00, 0.03] 0.00 [− 0.01, 0.01] 0.01 [− 0.00, 0.01] Negative AI 0.01 [− 0.01, 0.03] 0.01 [− 0.00, 0.02] – Growth 0.07 [0.02, 0.12] 0.01 [− 0.02, 0.04] – Fixed 0.05 [0.01, 0.09] 0.01 [− 0.01, 0.03] – Empathy 0.00 [− 0.01, 0.02] − 0.00 [− 0.01, 0.01] – CRT − 0.08 [− 0.13, − 0.03] − 0.03 [− 0.06, 0.00] – Age 0.00 [− 0.01, 0.01] − 0.00 [− 0.01, 0.01] – Emotion – 0.17 [0.13, 0.21] 0.18 [0.15, 0.20] Storyterm – 0.14 [0.11, 0.17] 0.12 [0.10, 0.14] Meaningful – 0.18 [0.14 0.23] 0.15 [0.12, 0.18] Effortterm·var – 0.25 [0.21, 0.29] 0.27 [0.23, 0.31] Time (log) – 0.04 [0.02, 0.06] 0.05 [0.03, 0.06] Labelterm × Openness 0.01 [− 0.01, 0.03] − 0.00 [− 0.01, 0.01] – Labelterm × Positive AI − 0.02 [− 0.03, − 0.00] − 0.01 [− 0.02, − 0.00] − 0.01 [− 0.02, − 0.00] Labelterm × Negative AI − 0.01 [− 0.03, 0.01] − 0.00 [− 0.01, 0.01] – Labelterm × Growth 0.00 [− 0.04, 0.05] 0.00 [− 0.03, 0.02] – Labelterm × Fixed − 0.01 [− 0.04, 0.02] 0.00 [− 0.02, 0.02] – Labelterm × Empathy 0.00 [− 0.01, 0.02] 0.00 [− 0.01, 0.01] – Labelterm × CRT − 0.02 [− 0.06, 0.03] − 0.00 [− 0.03, 0.02] – Labelterm × Age − 0.01 [− 0.01, 0.00] − 0.00 [− 0.01, 0.00] – Labelterm × Emotion – 0.01 [− 0.04, 0.07] – Labelterm × Storyterm – − 0.04 [− 0.09, 0.00] – Labelterm × Meaningful – − 0.06 [− 0.11, 0.00] – Labelterm × Effortterm·var – 0.06 [0.00, 0.11] 0.02 [− 0.02, 0.06] Labelterm × Time (log) – 0.01 [− 0.01, 0.04] – Random effects Participant (Intercept) 0.47 0.16 0.17 Labelterm (Slope) 0.24 0.10 0.10 Painting (Intercept) 0.05 0.01 0.01 Residual 0.61 0.36 0.36 Model Marginal 0.16 0.59 0.57 ICC 0.45 0.31 0.32 Conditional 0.54 0.72 0.71 AIC 11,040.92 8711.31 8557.09 Note . CI = confidence interval, Marginal = variance explained by fixed effects, ICC = intraclass correlation or variance explained by random effects, Conditional = variance explained by fixed and random effects, AIC = Akaike Information Criterion. Italicized text = p < 0.05, bolded text = p < 0.001 Additional analyses We found a significant main effect of Labelterm on the additional five criteria ( p ’s < 0.001), Emotion, Storyterm, Meaningful, Effortterm·var, and Time, such that these criteria were rated higher when the labelterm was “human-created” than “AI-created.” These full models (constructed as the previous Model 1’s with individual-difference measures) are reported in Additional file 1 : Table A.

Study 2 discussion Study 2 replicated and extended the general anti-AI artterm sentiments from Study 1. Using more-sophisticated modeling, we again found that participants tended to prefer artworks that were labelled human-created over those labelled AI-created artterm. Importantly, this pattern of results was observed across all four original criteria, as well as the additional five criteria added in Study 2.

We designed our analyses such that three models predicted scores for each criterion. In Model 1, individual-difference traits were targeted, as they were the only predictor variables inputted besides the Labelterm and Painting Typeterm·var. In Model 2, we added predictor variables that could vary with each trial—rather than remain static, as with the individual-difference metrics—which were the 5 additional criteria: Emotion, Storyterm, Meaningful, Effortterm·var, and Time. Model 2 aimed to identify the contributions that more-communicative criteria could predict over our original four criteria. For instance, how does the ability of an artterm piece to evoke a storyterm relate to the general likingterm of that piece, and does this relationship depend on the labelterm? Due to the abundance of predictor variables, several of which were not significant in Model 2, we created a third and final model consisting only of significant main effects and interactions from Model 2, which allowed for easier interpretation and discussion. This tripartite model approach was not specified in the pre-registration, though the basic model type and structure (linear mixed models) were.

As determined by our significant-only model (Model 3), when participants reported Likingterm some paintings more than others, regardless of Labelterm, we also found an increased ability to extract emotion out of the painting (Emotion), to create a storyterm alongside the painting (Storyterm), to find personal meaning from the painting (Meaningful), and to believe the painting required more effortterm·var to produce (Effortterm·var). Interestingly, interactions between Labelterm and two criteria (Storyterm, Effortterm·var) emerged, but in opposite directions. Specifically, with a greater ability to extract a storyterm from a painting (i.e., higher narrativityterm), participants liked the painting more if that painting was created by AI as compared to if it was created by humans. Narratives, then, may serve as a catalyst for appreciating products on surface levels. However, if participants believed the painting involved more effortterm·var in its creation than other paintings, participants liked paintings made by humans more than those made by AI. This may be due to a lack of understanding, on the participants’ end, of how AI produce pieces, and a need to actually witness the artistic production as investigated in Chamberlain et al. ( 2018 ). Additionally, it may reflect an effort heuristicterm. As suggested by Kruger et al. ( 2004 ), people use effortterm·var as a metric to inform their judgements of the quality of a product (e.g., poems, paintings, suits of armor; Kruger et al., 2004 ). These findings suggest this effort heuristicterm may only exist for human products and not AI products. No individual-difference traits emerged in Likingterm’s Model 3, but its Model 1 originally found that those with both higher growth-mindset scores and higher fixed-mindset scores liked paintings more in general. That fixed and growth mindsets were both positively associated with aesthetic judgementsterm was surprising, yet interesting. This finding partly mirrors results from Karwowski et al. ( 2019 ), who found some people can hold both fixed and growth attitudes toward creativityterm, and further nuances perceptions of creativityterm that were formerly considered along a continuum between fixed or growth poles (Karwowski, 2014 ). The Model 1 also reported a main effect of positive attitudes toward AIterm·var, which was qualified with an interaction with the Labelterm, such that with higher appreciation for AI, participants liked AI-labelled artterm more than human-labelled artterm.

The final model of Beautyterm revealed similar main effects and interactions as the final Likingterm model. There was also an additional main effect of age, such that younger participants found artterm, regardless of its purported creatorterm, to be more beautiful; this may be in line with findings from Tröndle et al. ( 2014 ), who found that younger participants are more willing to ascribe contemporary artistic creations as “artterm.” An interaction between Labelterm and Openness emerged, such that with higher openness to experienceterm, participants found AI-labelled paintings more beautiful as compared to human-labelled paintings, nuancing claims of openness and aesthetic appreciation by McCrae ( 2007 ). No other individual-difference measures emerged in this final model, but in Model 1, we originally had main effects of positive AI attitudes (qualified with an interaction with the Labelterm, similar to the Likingterm Model 1), and a main effect of fixed mindset. Again, that fixed mindsets emerge is an interesting point that merits future investigation.

We also implemented the same models for our communicativeterm criteria. The final models for Profundityterm and Worthterm showed a main effect of Emotion, Storyterm, Meaningful, Effortterm·var, and (nonsignificant in Likingterm and Beautyterm) Time. For both Profundityterm and Worthterm, there was also a significant interaction between the labelterm and positive attitudes toward AIterm·var, such that more positive attitudes toward AIterm·var led to higher perceived Profundityterm and Worthterm when the artterm was created by AI as compared to by a human. The Profundityterm final model, but not Worthterm’s, also had significant effects of growth and fixed mindsets. However, in both Model 1’s of Profundityterm and Worthterm, main effects of positive AI attitudes (qualified by an interaction with Labelterm), growth and fixed mindsets, and CRT scores emerged. The main effect specifically suggests that lower CRT scores (i.e., poorer ability to override intuitive responses) led to higher ratings of Profundityterm and Worthterm regardless of labelterm. This mirrors work by Pennycook et al. ( 2015 ) that investigated profundityterm and judgements of randomly generated statements, suggesting an overall increased ability to find Profundityterm and Worthterm in the environment when having a lowered filter for regulation. Lastly, supplemental analyses (see Additional file 1 ) also showed that the additional five communicativeterm criteria (Emotion, Storyterm, Meaningful, Effortterm·var, Time) behaved similarly in response to an “AI” or “human” labelterm (such that the latter has consistently higher ratings). In other words, a labelterm of “human” acts as a vessel for elaborative and communicativeterm engagement.

Overall, results from Study 2 supported a distinction in judgement patterns between Likingterm/Beautyterm and Profundityterm/Worthterm across models, replicating and extending these patterns from Study 1. Interestingly, it appears that when raters judge sensory-level criteria, they may rely on some properties that require greater engagement (e.g., story-telling and effortterm·var) depending on the labelterm provided. In this regard, it is possible that the sensory-level judgements can be both a consequence of more-passive appraisals based on solely visual features and an interaction with other judgement mechanisms that require active engagement with the piece, in line with dual-process models like that of Graf and Landwehr ( 2015 ). Meanwhile, when judging communicativeterm criteria like Profundityterm and Worthterm, people tend to rate along the lines of other judgements (e.g., time spent on painting), only moderated by personal positive AI attitudes depending on the labelterm. Individual differences, then, may only affect how viewers differently judge AI or human artterm when judging their communicativeterm properties.

General discussion In line with the extant literature on the role of labelsterm in subjective evaluations—both in aesthetics and other domains of preferences (e.g., Kirk et al., 2009 ; McClure et al., 2004 ; Newman & Bloom, 2012 ; Plassmann et al., 2008 ; TurpinpersName et al., 2019 )—our findings demonstrate that labelsterm significantly influence people’s judgements of visual artworks. Specifically, we investigated how a labelterm of “AI-created” on a painting affects the evaluation of the artwork, as compared to a labelterm of “human-created.” In a time of unprecedented growth in AI-art algorithms, websites, and exhibits, it is evermore timely to consider how people perceive, judge, and interact with artwork produced by AI. We found a bias against AI-labelled artterm across all aesthetic judgementterm criteria, but more so for appraisals related to artterm as a deeper communicativeterm medium (Profundityterm, Worthterm). In fact, differences in surface-level judgements between human- and AI-labelled artterm, though influenced by some communicativeterm processes (e.g., Effortterm·var and Storyterm), were nearly non-significant. These findings have crucial implications for artterm, AI, and creativityterm at large.

Artterm created by AI has been met with both success and controversy (see Chatterjee, 2022 ); and, very recently, these creative products have permeated into mainstream culture like museums and artistic industries (Diaz, 2022 ; Kaleagasi, 2017 ; Kinsella, 2018 ; Zulić, 2019 ). Though contention ensues over whether artificial programs are capable of producing “creative” products (in line with the computer science principle known as the Lovelace Objectionterm; Natale & Henrickson, 2022 ), it is nonetheless critical to understand how audiences perceive such artterm given that contemporary psychological literature considers creativityterm to be not an absolute construct, but a consequence of changing interactions between subjective criteria and socio-cultural factors like technology (i.e., “in the eye of the beholder”; Amabile, 1982 ; Cseh & Jeffries, 2019 ; Hennessey et al., 2011 ; Kaufman & Sternberg, 2010 ). Probing these attitudes toward AI-art, our studies demonstrate clear preferences for artwork labelled as human-created as compared to AI-created, despite all artterm being AI-created (which controlled for potential bottom-upterm discrepancies between actual human- versus AI-created paintings, while also leveraging findings that show AI artterm is indiscernible from human artterm; Chamberlain et al., 2018 ; Gangadharbatla, 2022 ). However, the degree of preference depended on the type of judgement, as our results show a marked distinction between judgements of Likingterm/Beautyterm and Profundityterm/Worthterm. In Study 1, for instance, we found increased effect sizes for differences between human and AI labelsterm for the more-communicative criteria of Profundityterm ( d = 0.47) and Worthterm ( d = 0.61) than Likingterm ( d = 0.17) and Beautyterm ( d = 0.22). This distinction is further emphasized by the different interaction mechanisms discovered in Study 2. Ultimately, these results align well with hierarchical and multi-processing models of aesthetic encounters that view sensory versus communicativeterm engagements as different pathways, including Graf and Landwehr’s ( 2015 ) dual-process model of aesthetic likingterm and Chatterjee and Vartanian’s ( 2016 ) Aesthetic Triadterm, which delineates a distinction between sensory- and meaning-level systems of aesthetics. Corroborating these models, our results suggest that people may feel cognitively obstructed when engaging deeply with and deriving meaning from artworks that are labelled as created by AI (or have any other labelterm that is pejorative). Equivalently, a “human” labelterm encourages elaborative engagement (e.g., deriving emotion, effortterm·var, narratives). However, on quick, surface-level evaluations rather than elaborative appraisals, AI artterm may be better appreciated.

Several interpretations can be drawn from these results. Firstly, participants may hold a nuanced view of creativityterm: while humans can achieve creativityterm across cognition and production, AI can produce only surface-level reflections of creative thinking according to raters. Artworks that are created by humans may reflect a profound human experienceterm—and thus be deemed more monetarily worthy—that AI cannot produce. Indeed, absent the human experienceterm, AI can produce only sensorily similar—still comparatively beautiful and liked—pieces of visual artterm. This provides a fascinating perspective of future creativityterm and aesthetics research: one that does not consider creativityterm as an all-or-nothing trait, but a spectrum of abilities that AI may be able to penetrate.

Even though these two lines of aesthetic engagement emerged, Study 2 showed that sensory-level judgements actually seemed to be influenced in interesting ways by additional judgements and processes. Such moderation effects support sensory versus communicativeterm pathways of engagement with artterm, but also reflect how they may intertwine. Indeed, this approach of aesthetics reconciles how aesthetic likingterm has been found to be both a consequence of passive viewing based on solely bottom-upterm processing of visual features and product of deeper and more-elaborative engagement with the piece (see Graf & Landwehr, 2015 ). We found support for this dual-process model in Study 2. Specifically, for Likingterm and Beautyterm (two sensory judgements), we isolated interactions between the narrativityterm of a painting and the labelterm, where higher narrativityterm led to higher Likingterm and Beautyterm when the labelterm of the painting was “AI-created” more than when the labelterm was “human-created.” This seems to imply that narrativityterm is a key element in surface-level evaluations of artwork, especially with non-human artists. This finding also has strong implications for marketing and product branding, as it is reminiscent of the mass production advertising practices in the early twentieth century where consumers were overall hesitant to purchase goods from multinational conglomerates until narratives were paired with the product (Freeman, 2014 ; Pulizzi, 2012 ). In other words, imagined backstories helped to overcome biases against machine-made products. This vessel for deeper engagement (and consumerism) provides a way for AI artterm to, potentially, have higher evaluations outside of simple bottom-upterm appraisals.

Narrativityterm has also played key roles in other modes of artterm. Music has specifically been investigated as a vessel for stories and resulting aesthetic engagements. Such work (e.g., Margulis et al., 2019 ; McAuley et al., 2021 ) finds interesting cross-cultural differences, implicating the role of media in our consumption of artterm and underlying associations (see Wu et al., 2020 for a cross-cultural study that finds higher appreciation for AI products of artterm from Chinese participants than American participants). The fact that narrativityterm scores interact with the labelterm of another modality in our study (paintings) strengthens the idea that stories serve as an engine for general artterm judgement (that is sensitive to context). Future work could further investigate this possibility by experimentally manipulating the level of narrativityterm of artworks, which would in turn allow for the assessment of the influence of such a manipulation on people’s aesthetic encounters and appraisals. In such a case, we would anticipate that increasing the narrativityterm associated with a given piece of artterm would produce more-positive judgement of that artterm relative to artworks with lower levels of narrativityterm. Narratives could also be indirectly stimulated; that is, certain titles and terms may automatically evoke stories which may affect ratings, similar to how pseudo-profound randomly generated titles increase perceived profundityterm for both computer-generated and human-created abstractterm paintings (TurpinpersName et al., 2019 ).

Another interaction that emerged in Study 2 was between the perceived effortterm·var of paintings and the labelterm, though in the opposite direction as the narrativityterm interaction: higher perceived effortterm·var led to higher Likingterm and Beautyterm when the labelterm was “human-created” more than when it was “AI-created.” Kruger et al.’s ( 2004 ) “effort heuristicterm,” then, may only be immediately available for human-made products, and people seek markers of effortterm·var in such products. For instance, Chamberlain et al. ( 2018 ) found that people observe the meticulous brush strokes (i.e., an/isotropy) in artworks and use these brush strokes as a heuristic for aesthetic judgementsterm. In other words, people may show an appreciation for an artist’s effortfulness (and/or meticulousness) as the labor infuses the product with some temporal or monetary value. That this effort heuristicterm doesn’t emerge for AI-labelled artterm could also reflect a general lack of knowledge about how much effortterm·var goes into AI-created artterm (and what “effortterm·var” even reflects in the context of AI algorithms). Interestingly, however, Chamberlain et al. ( 2018 ) were able to increase appreciation for AI-created artterm by showing participants videos of anthropomorphized robots creating paintings, which could have reasonably led participants to believe that the AI engaged in greater levels of effortterm·var in its artistic process (thereby mitigating the anti-AI biasterm). Thus, effort heuristicsterm may be probed even in instances of AI-art in future work.

In regard to relationships between judgements and with painting typeterm·var (abstractterm versus representationalterm), we only found modest significant interactions between the labelterm and painting typeterm·var for Likingterm and Profundityterm in Study 1, and did not include these interactions in Study 2 as none emerged significantly in initial analyses. This ultimately supports findings from Chamberlain et al. ( 2018 ), who found no significant interactions between the labelterm and painting typeterm·var of general aesthetic preferences. In addition, these findings give nuance to the attribution effects investigated by Gangadharbatla ( 2022 ) and Chamberlain et al. ( 2018 ), who found an increased willingness to attribute an AI creator to abstractterm paintings, and a human creator to representationalterm paintings, respectively. It should be emphasized, however, that in the present study, we defined abstractterm paintings as having partially or completely unrecognizable referents, and representationalterm paintings as having completely recognizable referents. Notably, however, more-strict definitions and criteria to dissociate abstractterm and representationalterm artterm could be used in future research. Importantly, by employing stricter criteria for separating abstractterm from representationalterm artterm, future research may, unlike in our study, find significant differences across these two painting typesterm·var. Future work could also explore intermediary categories like partially representationalterm artterm in such investigations.

Surprisingly, we also found very few significant relationships between our individual-differences measures and our various indices of artterm judgement. Only in Study 2 did we see robust effects of an individual trait on judgement patterns. That is, personal positive attitudes toward AIterm·var interacted with the labelterm such that higher appreciation for AI led to higher appreciation for artterm labelled as created by AI as compared to human-labelled artterm; perhaps as we accept and utilize machine learning in a growing technological world, these traits will naturally proliferate among the layperson, leading to growing acceptance of AI artterm. Overall, though, we were surprised by the lack of other clear predicting individual-difference measures in judgements of artterm. Though some small effects for age and openness emerged in models of Beautyterm in Study 2, ultimately, more research is needed to determine what, if any, personal traits predict why one may judge artterm by AI higher than artterm by humans across criteria.

Limitations Notably, there are some limitations to our studies that should be considered, to appropriately qualify our findings, that will inform improved designs in future research on the topic. Firstly, we cannot be certain of how well we deceived participants with some of our labelsterm (i.e., those that were false in cases wherein the labelsterm indicated that the artworks were created by a human), although we can nevertheless minimize these concerns (at least to some extent) by considering the past literature showing participants’ inability to discriminate between AI- and human-created paintings (Chamberlain et al., 2018 ; Gangadharbatla, 2022 ). In any case, moving forward, this concern can be eliminated by explicitly ensuring that participants are not able to accurately determine that all of our stimuli were in fact created by AI. Alternatively, a future study could employ both (actual) AI- and human-created artwork that has been normed on the criteria of note. Secondly, though our judgement criteria provided more nuance to simple “aesthetic pleasure” by asking for an artwork’s Profundityterm or Worthterm, for instance, it is possible the very act of asking participants to (perhaps atypically) consider these attributes of an artwork alters their natural evaluation. Third, another consideration is that the labelterm of “AI” is quite broad and therefore loose in its meaning, and this may have differentially influenced aesthetic judgementsterm across participants depending on their understanding of the term; future work could specifically ask participants what “AI” means to them, and have independent raters (who are knowledgeable about AI) rate participant responses as an individual-difference measure of AI knowledge that may moderate and/or mediate results. Relatedly, some participants may know more about the specifics of machine learning than others (including, for example, the amount of effortterm·var that goes into producing an AI algorithm, and the amount of time it takes an individual to create an artwork via such an algorithm), which could influence their responses to our various measures. Also, some participants may be inclined to consume potentially biased information about AI through news sources, which could influence their responses. A simple teaching intervention on AI or assessment of past experiences with AI could assuage these concerns.

Lastly, though previous studies have investigated the role of artterm expertise in judgements of AI versus human aesthetics (and aesthetics in general), here, we did not index rater expertise. However, findings on the effects of expertise have been somewhat mixed. Chamberlain et al. ( 2018 ) found no moderating effect of expertise on judgement biases on artterm created by AI or humans, and Moffat and Kelly ( 2006 ) found that non-musicians were surprisingly able to discern AI- versus human-created compositions better than musicians. That said, both Moffat and Kelly ( 2006 ) and Darda and Cross ( 2023 ) found that the anti-AI biasterm is actually stronger in the experts, which matches general trends that experts rate harsher (e.g., Lundy & Smith, 2017 ). However, for experts to have a harsher anti-AI biasterm than non-experts also goes against claims that “novices are… more non-aesthetically biased in their aesthetic judgments compared with experts,” which suggests that experts are more resistant to labelterm effects (Lundy & Smith, 2017 , p. 140). Regardless, research is merited to delineate the true role that background experience has on judgements of aesthetics (including AI aesthetics), though given that experts may have a stronger anti-AI biasterm as some studies have supported, our results may, if anything, reflect more-conservative effects given our sample was not limited to experts.

Conclusions Across two studies, we found evidence of a multi-processing approach to aesthetics, which has strong implications for creative thinking and artterm industries. Firstly, people preferred (purportedly) human-created artterm over AI-created artterm. This preference was particularly evident for criteria that communicated deeper meanings of the artterm (e.g., Profundityterm, Worthterm). On more-sensory levels, the difference between human- and AI-labelled artterm was much more modest, though significant differences were nonetheless observed. As such, AI-labelled artterm can still be greatly appreciated (almost as much as human-labelled artterm) when people utilize varying levels of engagement processes. Interestingly, the sensory-versus communicative-level judgement processes were further distinguished by different interaction processes: the former were moderated by rates of story-telling and perceived effortterm·var (though in opposite directions), while the latter were moderated by personal positive attitudes toward AIterm·var. These interactions shed light on when and why individuals may appreciate artterm made by different creators, as posed in our initial research question, and also reflect how different stages of multi-processing models of aesthetics may interact with one another.

In conclusion, people tend to perceive artterm as reflecting a human-specific experience, though creator labelsterm seem to mediate the ability to derive deeper evaluations from artterm. Thus, creative products like artterm may be achieved—according to human raters—by non-human AI models, but only to a limited extent that still protects a valued anthropocentrism.

Supplementary Information Additional file 1: Additional file 1 includes the Model 1 output for each of the Study 2 additional criteria.