Bias against AI art can enhance perceptions of human creativity

Abstract

The contemporary art world is conservatively estimated to be a $65 billion USD market that employs millions of human artists, sellers, and collectors globally. Recent attention paid to AI-made art in prestigious galleries, museums, and popular media has provoked debate around how these statistics will change. Unanswered questions fuel growing anxieties. Are AI-made and human-made art evaluated in the same ways? How will growing exposure to AI-made art impact evaluations of human creativity? Our research uses a psychological lens to explore these questions in the realm of visual art. We find that people devalue art labeled as AI-made across a variety of dimensions, even when they report it is indistinguishable from human-made art, and even when they believe it was produced collaboratively with a human. We also find that comparing images labeled as human-made to images labeled as AI-made increases perceptions of human creativity, an effect that can be leveraged to increase the value of human effort. Results are robust across six experiments (N = 2965) using a range of human-made and AI-made stimuli and incorporating representative samples of the US population. Finally, we highlight conditions that strengthen effects as well as dimensions where AI-devaluation effects are more pronounced.

Full text

1
Vol.:(0123456789)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports
Bias against AI art can enhance
perceptions of human creativity
C. Blaine Horton Jr
*, Michael W. White & Sheena S. Iyengar
The contemporary art world is conservatively estimated to be a $65 billion USD market that employs
millions of human artists, sellers, and collectors globally. Recent attention paid to AI-made art in
prestigious galleries, museums, and popular media has provoked debate around how these statistics
will change. Unanswered questions fuel growing anxieties. Are AI-made and human-made art
evaluated in the same ways? How will growing exposure to AI-made art impact evaluations of human
creativity? Our research uses a psychological lens to explore these questions in the realm of visual
art. We nd that people devalue art labeled as AI-made across a variety of dimensions, even when
they report it is indistinguishable from human-made art, and even when they believe it was produced
collaboratively with a human. We also nd that comparing images labeled as human-made to images
labeled as AI-made increases perceptions of human creativity, an eect that can be leveraged to
increase the value of human eort. Results are robust across six experiments (N = 2965) using a range
of human-made and AI-made stimuli and incorporating representative samples of the US population.
Finally, we highlight conditions that strengthen eects as well as dimensions where AI-devaluation
eects are more pronounced.
Will AI art devalue human creativity?
e contemporary art world is conservatively estimated to be a $65 billion USD market that employs millions of
human artists, sellers, and collectors across the world1. Yet recent attention paid to art made by articial intel-
ligence (AI) in prestigious galleries2,3, museums4, and popular media5 has provoked heated debate around how
these statistics will change in the future6,7. Anxiety around the changing value of human art is fueled by unan-
swered questions: Will art attributed to AI be evaluated in the same way as art attributed to humans? Should art
markets even treat AI-made art as “art”? Does growing exposure to AI-made art impact evaluations of solitary
human creativity or, as is also happening, evaluations of human artists using AI?
Historical examples from other industries provide ample evidence that, on average, automation decreases
the value of human goods and labor8. But there is also reason to believe that the development of AI technologies
capable of automating creativity (e.g., producing visual art largely indistinguishable from human art) should not
similarly impact perceptions of human art and creativity. As the famous American painter James Whistler once
said, "An artist is not paid for his labor but for his vision"9. Past research supports Whistler’s point. Whether
made by experts or lay audiences, evaluations of art oen depend upon both aesthetic and social dimensions
that can be disentangled from the more tangible costs of production and labor10,11. For example, an artist’s use of
color and emotion, the complexity of their subject matter, and the artist’s brand all impact perceptions of value
and creativity assigned to a given piece of art12,13. Subjective factors like these make it dicult to predict whether
growing exposure to AI-made art will negatively impact the value of human creativity in artistic domains. Put
another way, will the value of human art (monetary and aesthetic) increase or decrease when evaluated next to
works (of comparable artistic style and quality) thought to be produced by AI programs? e growing importance
of questions like this are reected in headlines that detail how some human artists have recently begun to take
legal action against AI companies upon discovering AI programs are being used to emulate their unique artistic
styles with startling accuracy14. Furthermore, examples like these fuel larger concerns that the value of human
labor may be changing. Is it true that, as some have suggested, creative jobs are “the last bastion of humanity”15,
or will AI decrease the value of human labor as has happened in so many other industrial revolutions of the
past8? As amore comprehensive review on this topic suggests, scientic inquiries focusing specically on the
intersections of technology and creativity are needed if we are to understand the impact generative AI is begin-
ning to have on the world16.
e current research adopts a psychological perspective to begin answering these questions. Specically, we
examine how artistic source attributions (e.g., human-made or AI-made) inuence evaluative judgments of lay
audiences across six experiments (total N = 2965). To adopt a comprehensive view, we assess an array of dierent
OPEN
Columbia Business School, New York, USA. *email: cbh2132@columbia.edu
Content courtesy of Springer Nature, terms of use apply. Rights reserved

2
Vol:.(1234567890)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
dimensions that include estimates of monetary value and labor as well as more artistic evaluations, such as per-
ceptions of skill and creativity. Our focus on art builds upon past work suggesting humans oen exhibit costly
aversions to decision-making algorithms, preferring instead to rely on humans for a variety of goods and services,
even when algorithms performs better than or on par with human agents17–19. Art presents a unique domain of
exploration in this eld, in part because recent research suggests humanizing autonomous technologies can help
mitigate aversion to algorithms20–23. is nding highlights a philosophical quandary. What is more humanizing
than the ability to produce art? Indeed, many famed artists and scientists across the ages have proclaimed art to
be a fundamentally human pursuit (e.g., James Joyce once said “art is the human disposition”24 and the esteemed
scientist Brian Greene similarly remarked “art [is what] makes us human”25). Iftrue, these philosophical notions
suggest that artistic works attributed to AI should not be subject to the algorithmic aversion observed in past
workbecause the ability to produce art may, in and of itself, be humanizing. On the other hand, it is alsopos-
sible that art attributedto AI will instead begin to change the ways we perceive and evaluate human creativity
going forward. While the breadth of these speculations extends well beyond the connes of any single scientic
examination, our research presents three key ndings that shed new light on this topic.
First, participants in our experiments consistently devalued art labeled as AI-made relative to art labeled as
human-made. is was true even when the art in question was largely indistinguishable from the art of famed
human artists and when we held the art itself constant (i.e., labeling the exact same piece as either “human-
made” or “AI-made”). ese eects were also evidenced regardless of participants’ overall feelings towards AI
or their background experiences (e.g., education or profession) focused on art or technology. ese points are
important, because many of the eects observed in past work on algorithmic aversion can be explained either
by straightforward confounds like simple dierences in content of stimuli orby unmeasured, individual dier-
ences like participant dispositions towards new technologies26. Even accounting for these, our rst key nding
continues to echo many historical examples of automation in other industries, with devaluation eects being
particularly pronounced on evaluations of skill and monetary value. Qualifying this, however, is the fact that
eects were noticeably less pronounced on moreartistic, aesthetic dimensions (e.g., evaluations of complexity or
emotional intensity) and weakened substantively when participants were not asked to directly compare human
and AI-made eorts.
Our second key nding was that art labeled as human-made was seen as more creative as a function of
exposure to art labeled as AI-made. at is, the same piece of art gained creative value when it was labeled as
human-made and compared to other works of art labeled as AI-made relative to when it was labeled as human-
made and compared to other works of art labeled as human-made. is nding is both surprising and important,
because contrary to predictions expressed in popular media outlets, it indicates there is potential for human
artists to benet from comparisons made between their work and the work of AI artists. It suggests novel avenues
of thought, among them the idea that AI programs may also represent tools that can be used to highlight or
accentuate the creative capacity of humans going forward.
Our third and nal key nding is that although art described as collaboratively made (i.e., art created by
human artists working with AI programs) was, on average, perceived to be less valuable than work described
as human-made and perceived to bemore valuable than work described as AI-made, perceptions of the human
artist as the primary creative agent depended largely upon whether the collaboration was being compared to
human or AI references. Put another way, the evaluative bias against AI-labeled artwork persisted even in circum-
stances where the AI functioned only as a human aid, but perceptions ofthe human artist’s contribution within
the collaboration were higherwhen evaluators were rst anchored on the eorts of AI art produced without the
help of humans. is nding is important because it suggests human artists working with AI can benet from
drawingcomparisons between their collaborative output and the outputof AI programs working alone.
Experiment 1
Results and discussion
In Experiment 1, participants (n = 119) were presented with 28 dierent images (observations = 3332) that were
pretested to capture a range of dierent artistic styles (see Fig.1). Each image was randomly labeled such that
every participant evaluated 14 images labeled as AI-made and 14 images labeled as human-made. ese were
evaluated on a battery of artistic dimensions that included how bright, colorful, complex, emotional, skillful,
inspiring, expensive, and likable participants found each image to be (α = 0.97).
Table1 displays our results. Overall, within-subjects comparisons revealed images labeled as AI-made were
evaluated less favorably, with an aggregate eect of t(118) = 5.52, p < 0.001, mdi = − 0.18, 95% CI [− 0.12, − 0.25],
d = 0.51. e direction and signicance of this eect remained unchanged when considering each dimension
separately (Table1) and when using multilevel regression models to control for variance attributable to specic
image-styles or individual preferences (i.e., models controlling for image eects and nesting within participant).
is eect was most pronounced on evaluations of expensiveness (mdi = − 0.45, d = 0.57) and skill (mdi = − 0.30,
d = 0.48)—and occurred despite the majority of participants (> 70%) reporting they would not have been able
to dierentiate between images without the labels provided. at is, even though most participants reported art
labeled as AI-made was largely indistinguishable from art labeled as human-made, they still evaluated art labeled
as AI-made less favorably. ese results suggest a general bias against AI-made art.
Experiment 2
Results and discussion
In Experiment 2, participants (n = 415) evaluated the same 28 images (observations = 11,620) on the same evalua-
tivedimensions used in Experiment 1 alongside an additional “willingness to pay” measure (α = 0.88). Twoother
key changes were made. First, we expanded our research focus by asking participants to report whether they
Content courtesy of Springer Nature, terms of use apply. Rights reserved

3
Vol.:(0123456789)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
believed each image qualied as “art” or not (1 = no, 2 = maybe, 3 = yes). Second, we used a between-subjects
design to provide a more conservative estimate of the bias observed in Experiment 1. Participants were ran-
domly assigned to one of three conditions where either (a) all images were unlabeled so as not to prime any
thoughts about the dierences between human and non-human art, (b) all images were labeled as AI-made, or
(c) participants were told in advance that some images were made by humans and others by AI but not which
ones (i.e., a ‘mystery’ condition that drew attention to ambiguous source information without using any labels).
e rst condition was used as a control condition, under the presumption that the majority of our participants
would assume the unlabeled images were human-made. is presumption was based upon a pre-test of these
images (conducted before our rst experiment), where participants indicated that when unlabeled, a majority
of the images looked human-made (see Fig.S1 in our Supplemental Information). Nevertheless, the lack of an
explicit manipulation check in this condition is one limitation of this design, something addressed in later stud-
ies where more explicit labels are used.
Figure1. Examples of stimuli used in Experiment 1. Note ese images represent a sample from the larger pool
of 28 images used in our rst study.
Table 1. Experiment 1 within-subjectevaluation dierences between human and AI-labeled images. P-values
reect paired t-tests comparing average ratings by participants on images labeled as Human-made vs. images
labeled AI-made. *p < 0.05, **p < 0.01, ***p < 0.001.
Dimension Mean dierence Eect size (Cohen’s d)
Expensive − 0.45*** 0.57
Skillful − 0.30*** 0.48
Complex − 0.18*** 0.36
Colorful − 0.12** 0.30
Liking − 0.12** 0.25
Emotional − 0.12*** 0.27
Bright − 0.08* 0.21
Inspiring − 0.09* 0.22
Overall (α = 0.95) − 0.18*** 0.51
Content courtesy of Springer Nature, terms of use apply. Rights reserved

4
Vol:.(1234567890)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
Kruskal–Wallis tests indicated overall dierences between conditions on aesthetic dimensions (χ2 [2] = 6.2,
p = 0.04) as well as dierences in how likely participants were to classify images as art (χ2 [2] = 63.19, p < 0.001).
Means and standard errors for all dimensions and conditions are presented in TableS1 in our Supplementary
Information and Dunn’s test pairwise comparisons assessing all dimensions and conditions are presented on
TableS2 in our Supplementary Information. Summarizing some of our key ndings here, the overall dierences
in evaluations were largely explained by more specic dierences in participant’s evaluations of how expensive
(χ2 [2] = 57.19, p < 0.001) and skillful (χ2 [2] = 38.73, p < 0.001) images were perceived to be. Specically, images
labeledas AI-made were rated as being signicantly less expensive (Z = − 7.06, p < 0.001) and skillful (Z = − 5.59,
p < 0.001) relative to unlabeled images, and also less expensive relative to mystery images (Z = − 5.95, p < 0.001).
Moreover, participants were less likely to say that AI-labeled images qualied as art compared to both unlabeled
images (Z = − 7.11, p < 0.001) and mystery images (Z = − 6.70, p < 0.001). One caveat, though, is that many of
the participants in the AI-labels condition (87%) still considered the vast majority of images to qualify as art.
Similar to images labeled as AI-made, participants rated mystery images (i.e., images we said could be either
human or AI-made) as signicantly less skillful (Z = − 5.16, p < 0.001) and marginally less expensive (Z = − 1.17,
p = 0.07), but also found them more inspiring (Z = 4.25, p = 0.07) in contrast to unlabeled images and were no
less likely to consider the images to be art (Z = − 0.47, p ≈ 1). We propose these varied outcomes can be attrib-
uted to participants accurately assuming some images in the mystery condition were human-made and others
AI-made. Indeed, the wider, more bimodal distributions observed in this condition serve as a manipulation
check suggesting comparisons between art thought to be human and AI-made does not lead to the categorical
devaluation of all art presented (when the source is ambiguous), but is instead encouraged by presenting clear
targets in the form of labels.
In sum, these ndings help to clarify results from our rst experiment, conrming that the bias against
AI-made art is particularly pronounced on dimensions of artistic skill and monetary value, with participants
reporting that images explicitly labeled as AI-made are less likely to qualify as art. Also noteworthy, is that many
within-subject dierences on other evaluative dimensions that were observed in our rst experiment (e.g., ratings
of complexity or colorfulness) were not signicant in this more conservative, between-subjects design. is may
suggest the strong bias against art labeled as AI-made observed in our rst experiment was less pronounced in
this experiment precisely because we did not force participants to make back-to-back comparisons of images
explicitly labeled as both human and AI-made.
Experiment 3
Results and discussion
In Experiment 3, we examine how evaluations of human creativity are impacted by immediatecomparisons
made between art labeled as human and AI-made. Specically, we randomize the order of image labels to assess
how human artists are evaluated aer exposure to AI-made art. All participants (n = 405) evaluated two images
selected from the stimuli used in our previous experiments. ese two images were selected because theyhad
comparable styles and were rated similarly across artistic dimensions (see Fig.S2 in our Supplementary Informa-
tion). Participants were randomly assigned to either a control condition (a)where both paintings were explicitly
labeled as human-made, or one of three experimental conditions where (b) both paintings were labeled as AI-
made, (c) the rst painting was labeled human-made and the second painting as AI-made, or (d) the rst paint-
ing was labeled as AI-made and the second painting as human-made. Holding the images constant in this way
allowed us to determine the extent to which a single piece of art gains or loses value as a function of comparing
art labeled as human or AI-made. In addition to evaluating aesthetic dimensions, participants estimated the
monetary value of each painting, as well as the skill, talent, and execution shown by the labeled artist (human
or AI). We also told participants these images were all physical paintings currently for sale at a private gallery
and included fake information provided by that gallery (e.g., “Gallery ID: #A2461 Untitled, 2019 Oil on canvas
24 in × 36 in”). is was done for two reasons. First, it helped us to verify that results observed in our previous
experiments could not be explained by assumptions that AI-produced paintings were simply digital images.
Second, it ensured our results are generalizable to real-world markets where physical (and not just digital) pieces
of art are bought and sold.
Aligned Rank Transformed Contrast (ART-C) tests were used to analyze data because they have been shown
to be a particularly well suited parametric test for comparing groups and have demonstrated more power than
ttests, Mann–Whitney, Wilcoxon, and the standard ART ANOVAs without inating Type I error rates27. Our
analysis revealed main eects of label, such that both paintings were evaluated as being worth less money when
labeled as AI-made (t[403] = − 7.04, p < 0.001; t[403] = − 5.75, p < 0.001), with the presumed AI artist also being
seen as less capable on dimensions of skill, talent, and execution (see TablesS3 and S4 in our Supplementary
Information). As with Experiment 2, we did not observe an eect of labels on aesthetic dimensions such as how
colorful or complex the images were (see TableS3 in our Supplementary Information). Again, we interpret this as
suggesting the more pronounced devaluation eects in Experiment 1which encompassed aesthetic dimensions
were driven in part by forcing participants to make back-to-back comparisons of images labeled as both human
and AI-made (i.e., using a within vs. between-subjects design). Finally, rather than detracting from evaluations
of human creativity, comparisons between art labeled as human and art labeled as AI-made positively impacted
evaluations of human eort (see TableS4 in our Supplementary Information). Specically, the second painting
was rated higher on aesthetic dimensions when it was labeled as human-made and followed a painting labeled
as AI-made (t[401] = 3.30, p = 0.001). at is, the exact same painting was evaluated more favorably when labeled
as human-made and compared against art labeled as AI-made than when it was labeled as human-made and
compared against another painting labeled as human-made.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

5
Vol.:(0123456789)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
Experiment 4
Results and discussion
Experiment 4 was preregistered and conducted to conceptually replicate our ndings from Experiment 3. Speci-
cally, are evaluations of human creativity positively impacted by exposure to artwork attributed to AI? Partici-
pants (n = 789) were shown the same two images used in Experiment 3 but in place of aesthetic dimensions they
were asked to evaluate creativity directly. Specically, participants rated each painting on how creative, novel,
likable, and appropriate to be sold in a gallery it was. ese measures were selected based upon past research28,29
to reect the “standard” denition of creativity in psychology which posits that ideas and objects are creative
when they are perceived to be creative, novel, and appropriate to some goal (e.g., appropriate to be sold in gal-
leries or to be enjoyed by audiences29,30). ese items were averaged together to construct an overall measure of
creativity (Image 1 α = 0.78, Image 2 α = 0.82). Like Experiment 3, participants were randomly assigned to either
a control condition (a) where both paintings were explicitly labeled as human-made, or one of two experimental
conditions where (b) the rst painting was labeled human-made and the second painting as AI-made, or (c) the
rst painting was labeled as AI-made and the second painting as human-made. Participants were also asked to
estimate the monetary value and production-time of each painting (i.e., how much time in hours they thought
it took to produce each painting). Finally, for use as control variables in supplementary models, participants
indicated their artistic and technological backgrounds on items like, “I used to (or currently) work in a job that
primarily deals with the visual arts (e.g. designer, gallery manager, art dealer).”
Overall, both paintings in Experiment 4 were rated as less creative, worth less money, and estimated to have
taken less time to produce when labeled as AI-made (see Table2 below and TableS5 in our Supplementary
Information Material). Consistent with the positive eect on aesthetic dimensions observed in Experiment 3,
the second painting was evaluated as more creative when it was labeled as human-made and followed a painting
labeled as AI-made (see Fig.2). Once more, this suggests directly comparing art labeled as human and AI-made
can increase perceptions of human creativity. at is, the exact same painting was judged to be more creative,
novel, likable, and appropriate to be sold in a gallery when it was labeled human-made and compared against
an AI-made painting than when it was labeled as human-made and compared against another human-made
painting. Note, these results did not change substantively in supplementary regression models used to control for
participants’ experience in either artistic or technological industries; models testing for interactions suggested
the boost in perceptions of creativity was slightly more pronounced in participants with more artistic experience
though the interaction term failed to reach signicance (b = 0.23, p = 0.22). is suggests it is unlikely that the
observed eects are primarily driven by anxieties specic to participants whose passions or livelihoods are more
directly impacted by generative AI technologies.
Experiment 5
Results and discussion
In Experiment 5 (preregistered), we test the generalizability of eects observed in Experiments 3 and 4 by recruit-
ing a representative sample of the US population (n = 709) to evaluate a larger set of images. Each participant was
asked to evaluate two images randomly selected from those used in our rst experiment. Labels were randomized
so that participants were assigned to either a control condition (a)where both paintings were labeled as human-
made, or an experimental condition (b)where the rst painting was labeled as AI-made and the second labeled
as human-made. Creativity, monetary value, and production time were all assessed using the same measures from
Experiment 4. As a control variable, participants indicated their opinion towards AI using four items (α = 0.77)
adaptedfrom the General Attitudes Towards Articial Intelligence Scale31.
Consistent with the bias against AI-made art documented in earlier studies, the rst painting presented to
participants was evaluated as less creative, worth less money, and estimated to have taken less time to produce
when labeled as AI-made (see TableS6 in our Supplementary Information). Consistent with results observed in
Experiment 3, the second painting presented was seen as more creative when labeled as human-made if it fol-
loweda painting labeled as AI-made relative to when it followed apainting labeled ashuman-made (see Table3
and Fig.3). e direction and signicance of these results was unchanged in supplementary regression models
using attitudes towards AI as a control variable; models testing for interactions suggested the boost in perceptions
of creativity may be slightly more pronounced for participants with greater anxiety about AI technologies but
the interaction term failed to reach signicance (b = 0.10, p = 0.29). is suggests these eects are not primarily
driven by a general distaste toward AI. at is, regardless of how participants felt about AI, they generally (a)
evaluated AI-made art less favorably and (b) saw human-made art as more creative aer evaluating AI-made art.
Table 2. Experiment 4 between-subjectevaluation means for Image 2. Standard errors are reported in
parentheses above. P-values reect post-hoc pairwise comparisons using ART-C tests comparing each
experimental group to the control condition. *p < 0.05, **p < 0.01, ***p < 0.001.
Human-anchor AI-anchor Human-anchor
Human label (control condition) Human label AI label
Creative 4.62 (0.08) 4.85 (0.08)* 4.24 (0.08)**
Monetary value 148.35 (11.86) 152.99 (19.73) 111.74 (13.30)***
Estimated time to produce 36.02 (1.43) 32.60 (1.46) 18.29 (1.35)***
Content courtesy of Springer Nature, terms of use apply. Rights reserved

6
Vol:.(1234567890)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
Experiment 6
Results and discussion
Experiment 6 (preregistered) used a representative sample of the US population (n = 527) to test whether the
eects observed in our previous experiments extend to art collaboratively produced by human artists and AI.
To clarify, many works made by AI today require humans to provide specic verbal prompts that direct the AI’s
eorts in some way. is might be consideredcollaborative only in a very strict sense, and wasnot what we were
interested in here. at is, choosing a prompt might be thought of as collaborative in the same way that a human
Figure2. Experiment 4 evaluations of creativity by order and condition. Note Colors correspond to conditions.
Grey is used for the control condition("Con") which only contained images labeled as human-made. Blue is
used for the rst experimental condition("E1") which contained an image labeled as human-made rst and an
image labeled as AI-made second. Green is used for the second experimental condition ("E2")which contained
an image labeled as AI-made rst and an image labeled as human-made second. e error bars represent the
standard errors.
Table 3. Experiment 5 between-subjectevaluation means for Image 2. Means and standard error reported
above. P-values reect between-subjects t-tests comparing the le and right columns. e direction and
signicance of these eects remained unchanged if we used OLS regression to additionally control for variation
attributable to specic images or individuals (e.g., anxiety about AI technologies and artistic experience).*p <
0.05, **p< 0.01, ***p < 0.001.
Dimension
Human anchor AI anchor
Mean dierence
Image 2
Human Label
(Control) Image 2
Human Label
Creativity 4.95 (0.06) 5.14 (0.06) t(688.18) = 2.26, p = 0.024
Monetary Value 132.87 (6.44) 167.83 (19.33) t(467.25) = 1.72, p = 0.087
Estimated Time to Produce 30.51 (1.23) 30.94 (1.18) t(696.81) = 0.25, p = 0.81
Content courtesy of Springer Nature, terms of use apply. Rights reserved

7
Vol.:(0123456789)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
patron might commission another human (with greater artistic skill) to produce a specic artwork without neces-
sarily being granted the shared title of “artist” (e.g., the ceiling of the Sistine Chapel and the Mona Lisa are not
typically thought of as collaborations, despite both pieces being commissioned by patrons who dictated some
or all of the themes and content therein32). Here, however, we were interested in the perception of distributed
artistic collaborations—when artists pool their talents and eort to generate a shared product. is distinction
is valuable because one can imagine that many AI technologies are currently being adopted to supplement,
facilitate, or expedite the creativity of employed human artists who already possess many of the requisite skills
needed to produce high quality art on their own. We deemed this exploration important because it is not clear
whether collaborations between artists and AI will be subject to the same positive and negative eects observed
in our previous experiments. at is, will collaborations between human artists and AI artists (which, by their
very existence, might be likely to prime cognitive comparisons between humans and AI) be evaluated more or
less favorably than art attributed to either party working in isolation?
To examine this question, participants in Experiment 6 rated two randomly ordered paintings (see Fig.S3 in
our Supplementary Information) that werenewly generated for this experiment and pretested to be comparable
in terms of creativity (p = 0.77), monetary value (p = 0.54), and estimated production time (p = 0.60). Participants
were randomly assigned to one of two conditions where they were rst shown a painting that was either labeled as
human-made or AI-made. Subsequently, all participants were shown a second painting that was always labeled as
collaboratively made across both conditions. Specically, participants read the prompt: “e following painting
was created by the artist Avery Taylor, collaborating with an articial intelligence program capable of imagining
and painting entirely of its own accord, in January of 2019.” Creativity, monetary value, and estimated production-
time were all measured using the same items from Experiment 5. Finally, to test whether humans are seen as
more or less responsible for creative output when working with an AI, participants estimated the distribution of
labor for the collaborative painting using a 100-point slider ranging from “All AI Eort” to “All Human Eort.”
Consistent with the ndings from our previous experiment, the rst painting was evaluated less favorably on
all dependent variables when labeled as AI-made (see TableS6 in our Supplementary Information). On average,
Figure3. Experiment 5 sequential evaluations of creativity by condition. Note Colors correspond to conditions.
Grey is used for the control condition which only contained images labeled as human-made. Green is used
for the experimental condition which contained an image labeled as AI-made rst and an image labeled as
human-made second. e error bars represent standard errors. An omnibus ART-C test revealed an overall
dierence between conditions for evaluations of image 2 (t[708] = − 3.084, p = 0.002) and post-hoc comparisons
found signicant dierences between evaluations of all images (p < 0.01) in the directions shown above with the
exception of the two images both labeled as human-made in the control condition (p = 0.11). Eects remained
unchanged if we used regression models to additionally control for variation attributable to specic images or
individuals (e.g., anxiety about AI technologies and artistic experience).
Content courtesy of Springer Nature, terms of use apply. Rights reserved

8
Vol:.(1234567890)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
the second painting (always labeled as collaboratively made) was rated less favorably than human-labeled paint-
ing and more favorably than AI-labeled painting (see Table4). Participants also estimated the human artist in
this collaboration was responsible for a greater portion of creative labor (53.94%) when the collaboration was
compared to a painting labeled as AI-made, but a smaller percentage of creative labor (36.76%) when the col-
laboration was compared against a painting labeled as human-made: t(514.1) = 9.09, p < 0.001, mdi = 17.18, 95%
CI [13.01, 21.35], d = 0.71 (se e Fig.4). In sum, these results indicate two important ndings. First, the bias against
AI-made art persists even when art is a collaborative production of humans and AI working together. Second,
estimates regarding human labor in a collaboration with AI depended largely upon if the collaborative piece was
being compared to solitary human or AI eorts. at is, when compared against AI-made art, human artists were
seen as responsible for more ofthe creative labor in a collaborationbut when compared against human-made
art, human artists were seen as responsible for less ofthe creative labor in a collaboration.
Table 4. Experiment 6 within-subjectmean dierence between “collaboration” painting and anchor. P-values
and standard error are from paired t-tests. Signicant eects reported above remained unchanged when
additionally including evaluations of image 1 as control variable in an OLS regression model. *p < 0.05,
**p < 0.01, ***p < 0.001.
Dimension Collaboration vs. human anchor Collaboration vs. AI anchor
Creative value − 0.58 (0.09)*** 0.46 (0.07)***
Monetary value − 126.60 (36.13)*** 30.06 (5.25)***
Estimated time to produce − 16.45 (1.62)*** 10.64 (1.03)***
Figure4. Estimates of workload distribution in a human-AI collaboration. Note e y-axis represents the
portion of work participants estimated the human was responsible for in their collaboration with an AI
technology. e x-axis represents the dierent conditions (i.e., whether participants evaluated an image labeled
as human-made or AI-made rst). e gray dotted line is included to illustrate where estimates of an even 50/50
split would land. e error bars represent standard errors.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

9
Vol.:(0123456789)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
General discussion
Do we evaluate art attributed to humans and art attributed to AI similarly? Our ndings suggest not. Across six
experiments, our results indicate that even when the content of the art is held constant, simply believing the art
was made by AI negatively impacts appraisal. is bias against art labeled as AI-made was particularly evident
on dimensions of monetary value and skill (see Experiments 1 and 2). On the one hand, these ndings are in
line with anxieties expressed in popular media outlets that like other industries aected by automation in the
past, AI-made art is likely to bring down the average monetary value commanded in artistic markets. Indeed,
in some cases, the eects we observe are substantive. For example, AI-made labels in our last experiment led
to a 62% decrease in monetary value and a 77% decrease in estimated production-time relative to perceptions
ofhuman eort. On the other hand, images labeled as AI-made in our studies were oen less liked and less likely
to be considered “art”, and our results suggest many of these devaluation eects can be mitigated by ensuring
audiences do not directly compare human and AI eorts (see Experiments 2 and 3). Of course, one limitation
in many of our experiments is that many of the AI-generated stimuli used were purposefully designed to be
indistinguishable from human-made art. at is, we pretested images to ensure any eects we observed would be
driven by sources attributions rather than dierences in content. In this regard, we can only speculate as to how
art that has been produced by less constrained AI will go on to impact the art world at large. Indeed, a small but
promising body of work suggests AI art already possess its own style33 and has been perceived as more creative
than contemporary human artists in some cases34. However, in light of these ndings and our own, it is dicult
to imagine how AI will not bring down average price of art in global markets, particularly if markets become
saturated with artistic products that are not only cheaper to produce en masse, but less likely to be valued by
consumers. Based on these conjectures, new strategies that allow human artists to maintain the market value
they currently command may be advisable. For instance, human value may be less subject to change if greater
eorts are made to partition art markets, segregating human and AI-made goods more denitively. One simple
proposition is to adopt the designation “synthography” to create greater psychological distance between specic
forms of humanart (like photography or digital art) and comparable works by made by AI35.
Will evaluations of human creativity be impacted by comparisons to AI? Our ndings suggest yes, but perhaps
not always in straightforward ways. For example, we nd that although participants consistently devalued art
if they believed it was made by AI, art labeled as human-made was seen as more creative when it was compared
against art labeled as AI-made than when it was compared against art attributed to other humans. is refutes
the sensationalized sentiment that “art is dead”5 because of the introduction of AI. If anything, our results predict
AI art has the potential to invigorate audiences to see human creativity in a new light if used carefully.
How will collaborations between humans and AI (e.g., when human artists use AI tools) be received in crea-
tive industries? Here, our ndings are nuanced but illustrative. Participants found images labeled as collabora-
tions between a human artist and AI to be more valuable than art labeled as solely AI-made but less valuable
than art labeled as solelyhuman-made. Importantly, though, we found that the human’s status as the primary
creative agent (i.e., the target more responsible for the output) in a collaboration depended upon whether the
collaboration was compared to solitary human or AI eorts. is nding has signicant practical value, because
maintaining one’s status as the primary creative agent implies the ability to capture a greater portion of economic
value. In short, it suggests human artists who use AI-tools mightbe smart to encouragecomparisons between
their own art andart made byAI workingwithout the assistance of human artists.
We are only just beginning to understand the impact AI technologies will have on the value of human
creativity, but it is worth noting that manyheated debates the birth of generative AI has provoked are strik-
ingly reminiscent of initial reactions to the invention of the camera. For example, the famous French painter
Paul Delaroche declared in response to the camera, “From today, painting is dead”36. Similarly, the famous art
critic Charles Baudelaire once said, “it will not be long before it [photography] has supplanted or corrupted art
altogether”37. In some ways, their anxieties were justied. e livelihoods of nineteenth century portraiture art-
ists were threatened by the invention of the camera38. And yet, the camera also gave us new ways to look at the
world, dening a period of innovation that eventually led to the creation andappreciationof new forms of art
that included impressionism, cubism, and digital photography39. Likewise, respected media outlets and artists
alike have recentlyooded the internet with responses to AI like, “More than ever before…I’m concerned for
the future of human creativity”6,40,41. Our ndings allow us to imagine a dierent future than these prevailing
forecasts of doom and gloom, one where the value of human creativity persists. Just as the invention of photog-
raphy ultimately inspired innovations like impressionism42, we suggest that a heightened appreciation for human
creativity can provide a fertile bed for new forms of human creativity and expression.
Methods
We report all conditions, measures, data exclusions, and provide copies of all study materials on our Open
Science Framework page (see Data and Code Availability). Research protocols were approved by the Human
Research Protection Oce and Institutional Review Board of Columbia University. We conrm that all research
was performed in accordance with relevant guidelines and regulations including the Declaration of Helsinki.
Participants gave informed consent to participate in all studies. Note, our rst two experiments were run in 2017
and 2018, before the more recent introduction of AI-art innovations like Midjourney and DALL-E2. In contrast,
our last experiment was run in 2023 just as Midjourney Version 4 and DALL-E2 were beginning to reach a
national audience. We share these dates because we believe it is important to note, historically, that our data was
collected both before and aer more recent (and sensationalized) coverage of AI art in prestigious media outlets
like the New York Times5 and Washington Post 6. at is, our research documents evaluations of art just as the
implications of AI artists are beginning to be realized.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

10
Vol:.(1234567890)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
Sample size determination and randomization
All sample sizes were determined before collecting data, with data collection halting aer analysis began. For
Experiment 1, we hypothesized a small to medium eect size (Cohen’s d = 0.33), which we used to determine our
sample size (i.e., roughly 95% power to detect an eect). We then used eect sizes from earlier experiments to
make sample size determinations for later experiments. Data quality was ensured in several ways. For instance,
we removed duplicate responses (e.g., repeated IP addresses or research IDs), participants who failed attention
checks, and collected entirely new samples for every study.
Data analysis and reporting
All data analysis was conducted in R (v.4.2.2). Eect sizes were calculated as Cohen’s d using the ‘eectsize’
package43. Whenever a ttest in our analysis did not demonstrate equal variance, a Welch’s t-test with corrected
degrees of freedom was used instead. Reported p-values are all two-sided. Finally, all regression models include
dummy-coded conditions that compare outcomes from against the control condition designated for that experi-
mental design.
Experimental samples and procedures
Pilot study
We pre-tested the 28 images used in Experiments 1, 2, and 5 in a pilot study (n = 105). is was done to ensure our
stimuli captured a range of styles and quality. Half of these images were lesser-known paintings from respected
artists (e.g., William Gear, Andy Warhol, and Paul Gauguin) while the other half were AI-generated images
rendered in the same styles of those artists. ese stimuli were chosen and tested to ensure that (a) participants
could not tell the dierence between human and AI-made art and (b) so that participants in Experiments 1
and 2 would be presented with style-matched pairs randomly labeled as “human-made” or “AI-made”. Results
conrmed images represented a range of quality (m = 4.30, sd = 1.21) and that participants generally could not
tell the dierence between images that were or were not AI-made. For example, aer evaluating each image on
aesthetic dimensions, participants were then asked to guess the origin of each image (1 = denitely human-made,
6 = denitely AI-made). Responses were skewed (see Fig.S1 in our Supplementary Information), indicating that
regardless of each image’s actual origin, participants thought the majority of images were human-made (e.g., the
average guesses for images that were, in reality, human-made were comparable to the guesses for images that
were, in reality, made using AI programs, m = 2.60, sd = 1.63 vs. m = 2.65, sd = 1.64, p = 0.59). To ensure our data
accurately represented lay evaluations of stimuli participants were unfamiliar with (i.e., experimental delity)
we used the question “Before taking this survey, had you seen any of these paintings before?” Participants who
responded yes to this question in any experiment were removed before any analysis, though supplementary
analysis including their responses did not change the direction or signicance of eects reported in this article.
Experiment 1
We recruited 143 English speaking US residents from Mturk. Participants were excluded for failing to pass atten-
tion checks or reporting they recognized stimuli used in the study, yielding a nal sample of n = 119 participants
(men = 52%,
mage
=
34
). Participants were paid $2 to complete the survey.
Aer rating three buer images to acclimate participants to the task, all participants rated 14 images labeled
as human-made and 14 images labeled as AI-made. Images were all presented in a random order. To make sure
that dierences in artistic style did not confound any results, labels were randomly assigned within style-matched
pairs (see our Pilot study) such that one image in each style-matched pair was always labeled as AI-made and
the other as human-made. is allowed us to make comparisons between images labeled as human or AI-made
while holding style constant. Participants then rated each painting on a battery of dimensions: how much they
liked it, how skillfully it was painted, how colorful it was, whether they found it inspiring, how bright it was,
how complex it was, how emotionally evocative, and whether they thought it was expensive (1 = Not at all, 7 = A
great deal; α = 0.88). For exploratory purposes we also asked participants about their general anity for art (e.g.,
“Some people seem to need art in their lives more than others; I consider myself that kind of person.”) and their
feelings about technological innovations (e.g., “I tend to dislike new technologies.”). Supplementary analysis
revealed these had no impact on our main ndings.
Experiment 2
We recruited 555 English speaking US residents from Mturk. Participants were excluded for failing to pass atten-
tion checks, comprehension checks, or reporting they recognized any stimuli used in the experiment, yielding
a nal sample of n = 415 participants (men = 51%,
mage
=
36
). Participants were paid $2 to complete the survey.
Aer rating three buer images to acclimate participants to the task, all participants rated 28 images in
random order. Participants were randomly assigned to one of three conditions. In a control condition we made
no mention of AI, nor did we label images as “human-made”. at is, we believed it important to have a control
condition in one study that did not prime any implicit comparisons of humans and non-humans that might
impact evaluations. Instead, images were unlabeled on the presumption that participants would assume the
images to be human-made. is presumption was based on our pre-test, where participants indicated the major-
ity of images looked human-made even aer being informed them that some were made by AI (Fig.S1 on p.1
of our Supplemental Information). us, participants were simply told that we were interested in “how people
perceive each painting on a number of dimensions.” In experimental conditions participants were either told
“each painting was made by an articial intelligence” or that “some of these paintings were made by a human
and others were made by an articial intelligence” but not which ones. Participants then rated each painting on
a battery of dimensions: how much they liked it, how skillfully it was painted, how colorful it was, whether they
Content courtesy of Springer Nature, terms of use apply. Rights reserved

11
Vol.:(0123456789)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
found it inspiring, how bright it was, how complex it was, how emotionally evocative, whether they thought it
was expensive, and how much they’d be willing to pay (1 = Not at all, 7 = A great deal; α = 0.94). For exploratory
purposes we also asked participants about their mood during the study (e.g., “Overall, my mood is:” = − 10 = Very
unpleasant, 10 = Very pleasant) and about their personal tastes in art (e.g., “I feel I have good taste in art.”). Sup-
plementary analysis revealed that these did not dier by condition and had no impact on our main ndings.
Experiment 3
We recruited 541 English speaking US residents from Mturk. Participants were excluded for failing to pass atten-
tion checks, comprehension checks, or reporting they recognized stimuli used in the study, yielding a nal sample
of n = 405 participants (male = 53%,
mage
=
38
). Participants were paid $1 to complete the survey.
To increase the external validity of our ndings, participants were given a cover story that said these images
represented real paintings for sale at a private gallery:
“On the next page, you’ll be shown two images of paintings currently for sale at the Lenham Private Gallery.
We are curious about consumer impressions of these paintings and the blurbs attached to them. Please review
the painting and information provided by the gallery and answer all questions honestly”.
Participants were then randomly assigned to one of four conditions where they rated two images. In a control
condition, both images were labeled as human-made. In one experimental condition both images were labeled as
AI-made. In another experimental condition, the rst image was labeled human-made and the second AI-made.
And in a second experimental condition, the rst image was labeled AI-made and the second human-made.
Image orderand the images themselves were held constantacross conditions. Human and AI-made labels read as
follows, “e following painting was created by Jamie Kendricks, in January of 2019.” or “e following painting
was created by an articial intelligence program, which imagines and paints images entirely of its own accord,
in January of 2019.” Building upon our cover story, bothpaintings were presented with unique ID numbers and
fabricated gallery information (e.g., “Lenham Private Gallery ID: #A2461; Untitled, 2019; Oil on canvas; 24
in × 36 in). Participants rated each image on a battery of dimensions: how muchthey liked each painting, how
skillfully it was painted, how colorful it was, whether they found it inspiring, how bright it was, how complex
it was, how emotionally evocative, whether the creator was talented, and whether they were impressed by the
execution (1 = Not at all, 7 = A great deal;
αimage
1=
0.86, αimage
1=
0.90
).
In addition, direct estimates of monetary value were obtained on a separate page immediately aer partici-
pants evaluated each painting on the dimensions listed above. On this page, participants were informed about
pricing with the prompt: “e average painting in the Lenham Gallery sells for somewhere between $50 and
$220, with most pieces retailing at $150.”ey were then asked, “How much do you personally think the Len-
ham gallery should sell this painting for?” and “Assuming that you wanted this painting and given the gallery’s
prices, how much would you pay to acquire it?” For exploratory purposes, we asked participants about their
owntaste in art (e.g., “Compared to other people, I generally have a better eye for art.” and “I like artwork that
depicts "real things" more than I like artwork that is abstract.”). Supplementary analysis revealed artistic taste
had no impact on our main ndings.
Experiment 4
We recruited 792 English speaking US residents from Prolic. Participants were excluded for failing to pass
attention checks, comprehension checks, or reporting they recognized any stimuli used in the study, yielding a
nal sample of n = 789 participants (male = 49%,
mage
=
38
). Participants were paid $1 to complete the survey.
Our pre-registration can be found here: https:// aspre dicted. org/ DJV_ MN7.
Participants weregiven the same cover story used in Experiment 3. ey were told that we were curious about
their impressions of paintings currently for sale at the Lenham Private Gallery and then randomly assigned to
one of three conditions where they were asked to evaluate two images. In a control condition, both images were
labeled as human-made by using the names of made-up human artists (e.g., “e following painting was created
by [Jamie Kendricks or Taylor Jennings], in January of 2019”). In one experimental condition, the rst image was
labeled human-made and the second AI-made (e.g., “e following painting was created by an articial intel-
ligence program, which imagines and paints images entirely of its own accord, in January of 2019.”). In another
experimental condition, the rst image was labeled AI-made and the second human-made. We used the same
labels and gallery information provided in Experiment 3. e perceived creativity of each image was measured
by askingparticipants how creative, novel, appropriate (to be sold in a gallery), and likable each image was (1
= not at all, 7 = a great deal;
αimage
1=
0.78, αimage
2=
0.82
). As in Experiment 3, participants were then given
information about pricing on a separate page and asked to estimate the monetary value of each painting. Addi-
tionally, participants were asked to estimate labor with the item: “How many hours of active painting time do
you think it took to create the painting above?” Finally, to make sure eects were not confounded byindividual
expertise in domains of art and technology, participants responded to ve items about artistic experience (e.g.,
“I used to [or currently] work in a job that primarily deals with the visual arts [e.g. designer, gallery manager,
art dealer].”
α
=
0.79)
and ve items about technological experience (e.g., “I used to [or currently]work in a job
that primarily deals with computer programming, data science, or engineering.”
α
=
0.72)
. Expertise did not
dier by condition and supplementary analysis revealed it had no impact on our main ndings.
Experiment 5
We recruited 731 English speaking US residents, using Prolic ltersto collect a representative sample of the U.S.
population. Participants were excluded for failing to pass attention checks, comprehension checks, or reporting
Content courtesy of Springer Nature, terms of use apply. Rights reserved

12
Vol:.(1234567890)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
they recognized stimuli used in the study, yielding a nal sample of n = 710 (ma le = 48%,
mage
=
45
). Participants
were paid $1 to complete the survey. Our pre-registration can be found here: https:// aspre dicted. org/ GJ4_ VS4.
Participants responded to the same survey used in Experiment 4 with two dierences. First, images were
randomly selected and ordered from the larger pool of 28 pretested images used in Experiments 1 and 2. Second,
participants were asked at the end of the survey to indicate their own specic attitudes toward AIusing four
items (
α
=
0.77)
) adaptedfrom the General Attitudes Towards Articial Intelligence Scale31(e.g., “I shiver with
discomfort when I think about future uses of Articial Intelligence.” and “I think Articial Intelligence programs
are an exciting new tool for human artists.”;
α
=
0.77)
. Notably, though participants who felt anxious about AI
technology rated AI-labeled artwork less favorably overall, attitudes towards AI did not dier by condition and
supplementary analysis using attitudes towards AI as a control variable had no impact on our main ndings.
Experiment 6
We recruited 698 English speaking US residents, using Proliclters to collect a representative sample of the
U.S. population. Participants were excluded for failing to pass attention and comprehension checks yielding a
nal sample of n = 527 participants (male = 45%, 49). Participants were paid $1 to complete the survey. Our pre-
registration can be found here: https:// aspre dicted. org/ DXP_ K4M.
Participants were given the same cover story and prompt used at the beginning of Experiments 4 and 5 (i.e.,
rating “paintings currently for sale at the Lenham Private Gallery”) before being presented with two images in
random order. ese images were not drawn from our previous studies but were instead created specically
for this experimentusing the AItool, Midjourney. ese stimuli were generated to address a limitation in our
previous studies, mainly that the images used were constrained to the styles of specic, and historically well-
known artists. at is, we did not allow the AI to really be ‘itself’. Seeking to reduce this concern to a degree,
our new stimuli were generated by giving the AI toolMidjourny more ambiguous prompts, asking for both “a
creative painting” and “a classical painting”. We chose one image (see our supplementary materials) from the four
automaticallygenerated for each prompt at random. We then conducted a pretest on these images to ensure that
any experimental eects couldn’t be attributed to intentional content variations that mightdominate participant
assessments of creativity, monetary value, or estimated production time.
Participants were randomly assigned to one of two conditions. In a control condition, the rst image was
labeled with the tag: “e following painting was created by Jamie Kendricks, in January of 2019.” In our experi-
mental condition, the rst image was labeled with the tag: “e following painting was created by an articial
intelligence program, which imagines and paints images entirely of its own accord, in January of 2019.” For
participants in both conditions, the second image was always labeled with the tag: “e following painting was
created by the artist Avery Taylor, collaborating with an articial intelligence program capable of imagining and
painting images entirely of its own accord, in January of 2019.” Participants rated both images using the same
items from Experiment 5 as well as an additional question, specic to the second image, that asked “how much
work do you think was done by the AI vs the human?” using a sliding scale (0 = All AI Eort, 100 = All Human
Eort).
Data and code availability
All data, survey materials, and code are available on the Open Science Framework at https:// osf. io/ xs8bv/? view_
only= 96a2b 4c29a 5b4ad f8cca 18025 e1881 1c.
Received: 26 May 2023; Accepted: 17 October 2023
References
1. McAndrew, C. e Art Basel and UBS Survey of Global Collecting 2022. Retrieved 15 Mar 2023, from https:// www. ubs. com/ global/
en/ our- rm/ art/ colle cting/ art- market- survey. html (2022).
2. Graham, T. Art made by AI is selling for thousands – is it any good? BBC. Retrieved 9 May 2023, from https:// www. bbc. com/ cultu
re/ artic le/ 20181 210- art- made- by- ai- is- selli ng- for- thous ands- is- it- any- good (2018).
3. Sutton, B. Articial Intelligence Artwork by Mario Klingemann Sells for £40,000 at Sotheby’s. Artsy. Retrieved 12 Jan 2023, from
https:// www. artsy. net/ artic le/ artsy- edito rial- artwo rk- creat ed- ai- sold- 40- 000- sothe bys- faili ng- gener a t e- fervor- prope lled- ai- work-
sell- 40- times- estim ate- year (2019).
4. Anadol, R. Unsupervised—Machine Hallucinations—MoMA. e Museum of Modern Art, New York, November 19, 2022–April
15, 2023. Retrieved 8 May 2023, from https:// www. moma. org/ calen dar/ exhib itions/ 5535 (2022).
5. Roose, K. An A.I.-generated picture won an art prize. Artists aren’t happy. New York Times. Retrieved 10 Jan 2023, from https://
www. nytim es. com/ 2022/ 09/ 02/ techn ology/ ai- arti cial- intel ligen ce- artis ts. html (2022).
6. Burnett, T. B. & Taplin, J. Opinion | to protect human artistry from AI, new safeguards might be essential. e Washington Post.
Retrieved 20 Mar 2023, from https:// www. washi ngton post. com/ opini ons/ 2023/ 03/ 14/ arti cial- intel ligen ce- threa tens- creat ive- artis
ts/ (2023).
7. Santos, R. Can AI-generated art replace creative humans?. VICE. Retrieved 10 Jan 2023, from https:// www. vice. com/ en/ artic le/
epzkwm/ arti cial- intel ligen ce- art- creat ives- ai (2022).
8. Xu, M., David, J. M. & Kim, S. H. e fourth industrial revolution: Opportunities and challenges. Int. J. Financ. Res. 9(2), 90–95
(2018).
9. Garg, A. Another Word a Day: An All-new Romp rough Some of the Most Unusual and Intriguing Words in English 163 (John
Wiley & Sons, 2005).
10. Spee, B. T. et al. Social reputation inuences on liking and willingness-to-pay for artworks: A multimethod design investigating
choice behavior along with physiological measures and motivational factors. PLoS ONE 17(4), e0266020 (2022).
11. Wohl, H. Bound by Creativity: How Contemporary Art is Created and Judged (University of Chicago Press, 2021).
12. Marin, M. M. & Leder, H. Examining complexity across domains: Relating subjective and objective measures of aective environ-
mental scenes, paintings and music. PLoS ONE 8(8), e72412 (2013).
Content courtesy of Springer Nature, terms of use apply. Rights reserved

13
Vol.:(0123456789)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
13. Moulard, J. G., Rice, D. H., Garrity, C. P. & Mangus, S. M. Artist authenticity: How artists’ passion and commitment shape consum-
ers’ perceptions and behavioral intentions across genders. Psychol. Mark. 31(8), 576–590 (2014).
14. Ta, A. What happens when AI creates art that looks like your work?. KCRW. https:// www. kcrw. com/ news/ shows/ press- play- with-
madel eine- brand/ ukrai ne- arti cial- intel- sciss or- siste rs/ ai- artis ts (2023).
15. Salter, P. David Glick on the business creativity, Alexander McQueen and Venture Capital. Forbes. https:// www. forbes. com/ sites/
phili psalt er/ 2022/ 04/ 11/ david- glick- on- the- busin ess- creat ivity- lee- mcque en- and- ventu re- capit al/? sh= 7b524 fe573 62 (2022).
16. Epstein, Z. et al. Art and the science of generative AI. Science 380(6650), 1110–1111 (2023).
17. Castello, N., Bos, M. & Lehmann, D. Task-dependent algorithm aversion. J. Market. Res. 56(5), 809–825 (2019).
18. Dietvorst, B. J., Simmons, J. P. & Massey, C. Algorithm aversion: People erroneously avoid algorithms aer seeing them err. J. Exp.
Psychol. General 144(1), 114 (2015).
19. Fuchs, C., Schreier, M. & Van Osselaer, S. M. e handmade eect: What’s love got to do with it?. J. Market. 79(2), 98–110 (2015).
20. Schweitzer, F., Belk, R., Jordan, W. & Ortner, M. Servant, friend or master? e relationships users build with voice-controlled
smart devices. J. Market. Manag. 35(7–8), 693–715 (2019).
21. Ahn, J., Kim, J. & Sung, Y. AI-powered recommendations: e roles of perceived similarity and psychological distance on persua-
sion. Int. J. Advert. 40(8), 1366–1384 (2021).
22. Jago, A. S. Algorithms and authenticity. Acad. Manag. Discov. 5(1), 38–56 (2019).
23. Epstein, Z., Levine, S., Rand, D. G. & Rahwan, I. Who gets credit for AI-generated art?. Iscience 23(9), 101515 (2020).
24. Joyce, J. Occasional, Critical, and Political Writing (Oxford University Press, 2002).
25. Barkhorn, E. A scientist’s quest to make us care about the cosmos. e Atlantic. https:// www. theat lantic. com/ enter tainm ent/ archi
ve/ 2010/ 06/a- scien tists- quest- to- make- us- care- about- the- cosmos/ 57567/ (2010).
26. Morewedge, C. K. Preference for human, not algorithm aversion. Trends Cogn. Sci. (2022).
27. Elkin, L. A., Kay, M., Higgins, J. J. & Wobbrock, J. O. An aligned rank transform procedure for multifactor contrast tests. In e
34th Annual ACM Symposium on User Interface Soware and Technology 754–768 (2021).
28. Amabile, T. M. & Hennessey, B. A. Consensual assessment. Encycl. Creat. 1, 347–359 (1999).
29. Runco, M. A. & Jaeger, G. J. e standard denition of creativity. Creat. Res. J. 24(1), 92–96 (2012).
30. Sternberg, R. J. Creating a vision of creativity: e rst 25 years. Psychol. Aesthet. Creat. Arts 1, 2 (2006).
31. Schepman, A. & Rodway, P. Initial validation of the general attitudes towards Articial Intelligence Scale. Comput. Hum. Behav.
Rep. 1, 100014 (2020).
32. Edgington, K. e History of Art on Commission. Singulart Magazine. https:// www. singu lart. com/ en/ blog/ 2020/ 05/ 27/ the- histo
ry- of- art- on- commi ssion/ (2023).
33. Manovich, L. & Arielli, E. Articial aesthetics: A critical guide to AI. Media Des. (2021).
34. Elgammal, A., Liu, B., Elhoseiny, M. & Mazzone, M. Can: Creative adversarial networks, generating" art" by learning about styles
and deviating from style norms. arXiv preprint arXiv: 1706. 07068 (2017).
35. Reinhuber, E. (2022, February). Synthography–An invitation to reconsider the rapidly changing toolkit of digital image creation
as a new genre beyond photography. In ArtsIT, Interactivity and Game Creation: Creative Heritage. New Perspectives from Media
Arts and Articial Intelligence. 10th EAI International Conference, ArtsIT 2021, Virtual Event, December 2-3, 2021, Proceedings
321–331(Springer International Publishing, Cham, 2021).
36. Ballinetti, C. “From today painting is dead”: Photography’s revolutionary eect. Art Object. Retrieved 15 Mar 2023, from https://
www. artan dobje ct. com/ news/ today- paint ing- dead- photo graph ys- revol ution ary- eect (2019).
37. Baudelaire, C. (1955). e Salon of 1859. Mirror Art 230–231.
38. McCouat, P. Early inuences of photography on art. J. Art Soc. Retrieved 3 Mar 2023, from h ttps:// www. artin socie ty. com/ pt-1- initi
al- impac ts. html (2018).
39. Martinique, E. How Did Photography Inuence the Impressionists?. Widewalls. Retrieved 3 Mar 2023, from https:// www. widew
alls. ch/ magaz ine/ impre ssion ists- photo graphy- museo- thyss en- borne misza (2019).
40. Marcus, J. Artists decry use of AI art: ‘I’m concerned for the future of human creativity’. Independent. Retrieved 15 Mar 2023, from
https:// www. indep endent. co. uk/ news/ world/ ameri cas/ ai- art- lensa- magic- avatar- b2242 891. html (2022).
41. Sha, S. ‘It’s the opposite of art’: Why illustrators are furious about AI. Guardian. Retrieved 15 Mar 2023, from https:// www. thegu
ardian. com/ artan ddesi gn/ 2023/ jan/ 23/ its- the- oppos ite- of- art- why- illus trato rs- are- furio us- about- ai (2023).
42. Silva, E. How photography pioneered a new understanding of art. Collector. https:// www . theco llect o r. com/ how- photo graphy- trans
formed- art/ (2022).
43. Ben-Shachar, M. S., Lüdecke, D. & Makowski, D. eectsize: Estimation of eect size indices and standardized parameters. J. Open
Source Sow. 5(56), 2815 (2020).
Acknowledgements
We extend a special thanks to Andrew Marks, Professor and Chair of the Department of Physiology and Cellular
Biophysics at Columbia University College of Physicians and Surgeons for providing outside feedback and guid-
ance on early dras of this manuscript; Hod Lipson, the esteemed robotics expert and Professor of Innovation
in the Department of Electrical Engineering at Columbia University for sharing a wealth of information about
his own AI-art journey; and Jake Bernstein, who provided research assistance in the form of various historical
fact-nding missions.
Author contributions
e authors conrm contribution to the paper as follows: study conception and design: C.B.H., M.W.W., S.S.I.;
data collection: C.B.H.; analysis and interpretation of results: C.B.H.; dra manuscript preparation: C.B.H.,
M.W.W., S.S.I.. All authors reviewed the results and approved the nal version of the manuscript.
Competing interests
e authors declare no competing interests.
Additional information
Supplementary Information e online version contains supplementary material available at https:// doi. org/
10. 1038/ s41598- 023- 45202-3.
Correspondence and requests for materials should be addressed to C.B.H.J.
Reprints and permissions information is available at www.nature.com/reprints.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

14
Vol:.(1234567890)
Scientic Reports | (2023) 13:19001 | https://doi.org/10.1038/s41598-023-45202-3
www.nature.com/scientificreports/
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional aliations.
Open Access is article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the
Creative Commons licence, and indicate if changes were made. e images or other third party material in this
article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.
© e Author(s) 2023
Content courtesy of Springer Nature, terms of use apply. Rights reserved

1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com