PNAS Nexus, 2024, 3, 1–8

https://doi.org/10.1093/pnasnexus/pgae052
Advance access publication 5 March 2024
Research Report

Generative artificial intelligence, human creativity, and

art
a, a,b,
Eric Zhou * and Dokyun Lee *
a
Department of Information Systems, Boston University Questrom School of Business, Boston, MA 02215, USA
b

Downloaded from https://academic.oup.com/pnasnexus/article/3/3/pgae052/7618478 by guest on 05 December 2024

Computing & Data Sciences, Boston University, Boston, MA 02215, USA
*To whom correspondence should be addressed: Email: [email protected] (E.Z.); Email: [email protected] (D.L.)
Edited By: Matthew Harding

Abstract
Recent artificial intelligence (AI) tools have demonstrated the ability to produce outputs traditionally considered creative. One such
system is text-to-image generative AI (e.g. Midjourney, Stable Diffusion, DALL-E), which automates humans’ artistic execution to
generate digital artworks. Utilizing a dataset of over 4 million artworks from more than 50,000 unique users, our research shows that
over time, text-to-image AI significantly enhances human creative productivity by 25% and increases the value as measured by the
likelihood of receiving a favorite per view by 50%. While peak artwork Content Novelty, defined as focal subject matter and relations,
increases over time, average Content Novelty declines, suggesting an expanding but inefficient idea space. Additionally, there is a
consistent reduction in both peak and average Visual Novelty, captured by pixel-level stylistic elements. Importantly, AI-assisted
artists who can successfully explore more novel ideas, regardless of their prior originality, may produce artworks that their peers
evaluate more favorably. Lastly, AI adoption decreased value capture (favorites earned) concentration among adopters. The results
suggest that ideation and filtering are likely necessary skills in the text-to-image process, thus giving rise to “generative synesthesia”—
the harmonious blending of human exploration and AI exploitation to discover new creative workflows.

Keywords: generative AI, human–AI collaboration, creative workflow, impact of AI, art

Significance Statement
We investigate the implications of incorporating text-to-image generative artificial intelligence (AI) into the human creative workflow.
We find that generative AI significantly boosts artists’ productivity and leads to more favorable evaluations from their peers. While
average novelty in artwork content and visual elements declines, peak Content Novelty increases, indicating a propensity for idea ex­
ploration. The artists who successfully explore novel ideas and filter model outputs for coherence benefit the most from AI tools,
underscoring the pivotal role of human ideation and artistic filtering in determining an artist’s success with generative AI tools.

Introduction many artistic communities which perceive generative AI as a

Recently, artificial intelligence (AI) has exhibited that it can feas­ threat to substitute the natural human ability to be creative.
ibly produce outputs that society traditionally would judge as cre­ Text-to-image generative AI has emerged as a candidate system
ative. Specifically, generative algorithms have been leveraged to that automates elements of humans’ creative process in produ­
automatically generate creative artifacts like music (1), digital art­ cing high-quality digital artworks. Remarkably, an artwork cre­
works (2, 3), and stories (4). Such generative models allow humans ated by Midjourney bested human artists in an art competition,a
to directly engage in the creative process through text-to-image while another artist refused to accept the top prize in a photo com­
systems (e.g. Midjourney, Stable Diffusion, DALL-E) based on the petition after winning, citing ethical concerns.b Artists have filed
latent diffusion model (5) or by participating in an open dialog lawsuits against the founding companies of some of the most
with transformer-based language models (e.g. ChatGPT, Bard, prominent text-to-image generators, arguing that generative AI
Claude). Generative AI is projected to become more potent to steals from the works upon which the models are trained and in­
automate even more creative tasks traditionally reserved for hu­ fringes on the copyrights of artists.c This has ignited a broader de­
mans and generate significant economic value in the years to bate regarding the originality of AI-generated content and the
come (6). extent to which it may replace human creativity, a faculty that
Many such generative algorithms were released in the past many consider unique to humans. While generative AI has dem­
year, and their diffusion into creative domains has concerned onstrated the capability to automatically create new digital

Competing Interest: The authors declare no competing interest.

Received: January 4, 2024. Accepted: January 23, 2024
© The Author(s) 2024. Published by Oxford University Press on behalf of National Academy of Sciences. This is an Open Access article
distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs licence (https://creativecommons.org/
licenses/by-nc-nd/4.0/), which permits non-commercial reproduction and distribution of the work, in any medium, provided the original
work is not altered or transformed in any way, and that the work is properly cited. For commercial re-use, please contact reprints@oup.­
com for reprints and translation rights for reprints. All other permissions can be obtained through our RightsLink service via the Pe­
rmissions link on the article page on our site—for further information please contact [email protected].
2 | PNAS Nexus, 2024, Vol. 3, No. 3

artifacts, there remains a significant knowledge gap regarding its Creative productivity
impact on productivity in artistic endeavors which lack well- We define creative productivity as the log of the number of arti­
defined objectives, and the long-run implications on human cre­ facts that a user posts in a month. Figure 1a reveals that upon
ativity more broadly. In particular, if humans increasingly rely adoption, artists experience a 50% increase in productivity on
on generative AI for content creation, creative fields may become average, which then doubles in the subsequent month. For the
saturated with generic content, potentially stifling exploration of average user, this translates to approximately 7 additional arti­
new creative frontiers. Given that generative algorithms will re­ facts published in the adoption month and 15 artifacts in the fol­
main a mainstay in creative domains as it continues to mature, lowing month. Beyond the adoption month, user productivity
it is critical to understand how generative AI is affecting creative gradually stabilizes to a level that still exceeds preadoption vol­
production, the evaluation of creative artifacts, and human cre­ ume. By automating the execution stage of the creative process,
ativity more broadly. To this end, our research questions are adopters can experience prolonged productivity gains compared
3-fold: to their nonadopter counterparts.

Downloaded from https://academic.oup.com/pnasnexus/article/3/3/pgae052/7618478 by guest on 05 December 2024

1. How does the adoption of generative AI affect humans’ Creative value
creative production?
If users are becoming more productive, what of the quality of the
2. Is generative AI enabling humans to produce more creative
artifacts they are producing? We next examine how adopters’ ar­
content?
tifacts are evaluated by their peers over time. In the literature, cre­
3. When and for whom does the adoption of generative AI lead
ative Value is intended to measure some aspect of utility,
to more creative and valuable artifacts?
performance and/or attractiveness of an artifact, subject to tem­
poral and cultural contexts (9). Given this subjectivity, we meas­
Our analyses of over 53,000 artists and 5,800 known AI adopters
ure Value as the number of favorites an artwork receives per
on one of the largest art-sharing platforms reveal that creative
view after 2 weeks, reflecting its overall performance and context­
productivity and artwork value, measured as favorites per view,
ual relevance within the community. This metric also hints at the
significantly increased with the adoption of text-to-image
artwork’s broader popularity within the cultural climate, suggest­
systems.
ing a looser definition of Value based on cultural trends.
We then focus our analysis on creative novelty. A simplified
Throughout the paper, the term “Value” will refer to these two
view of human creative novelty with respect to art can be sum­
notions.
marized via two main channels through which humans can inject
Figure 1b reveals an initial nonsignificant upward trend in the
creativity into an artifact: Contents and Visuals. These concepts are
Value of artworks produced by AI adopters. But after 3 months,
rooted in the classical philosophy of symbolism in art which sug­
AI adopters consistently produce artworks judged significantly
gests that the contents of an artwork is related to the meaning or
more valuable than those of nonadopters. This translates to a
subject matter, whereas visuals are simply the physical elements
50% increase in artwork favorability by the sixth month, jumping
used to convey the content (7). In our setting, Contents concern
from the preadoption average of 2% to a steady 3% rate of earning
the focal object(s) and relations depicted in an artifact, whereas
a favorite per view.
Visuals consider the pixel-level stylistic elements of an artifact.
Thus, Content and Visual Novelty are measured as the pairwise
Content Novelty
cosine distance between artifacts in the feature space (see
Figure 1c shows that average Content Novelty decreases over time
Materials and methods for details on feature extraction and how
among adopters, meaning that the focal objects and themes with­
novelty is measured).
in new artworks produced by AI adopters are becoming progres­
Our analyses reveal that over time, adopters’ artworks exhibit
sively more alike over time when compared to control units.
decreasing novelty, both in terms of Concepts and Visual features.
Intuitively, this is equivalent to adopters’ ideas becoming more
However, maximum Content Novelty increases, suggesting an ex­
similar over time. In practice, many publicly available fine-tuned
panding yet inefficient idea space. At the individual level, artists
checkpoints and adapters are refined to enable text-to-image
who harness generative AI while successfully exploring more in­
models to produce specific contents with consistency. Figure 1d,
novative ideas, irrespective of their prior originality, may earn
however, reveals that maximum Content Novelty is increasing
more favorable evaluations from their peers. In addition, the
and marginally statistically significantly within the first several
adoption of generative AI leads to a less concentrated distribution
months after adoption. This suggests two possibilities: either a
of favorites earned among adopters.
subset of adopters are exploring new ideas at the creative frontier
or the adopter population as a whole is driving the exploration and
expansion of the universe of artifacts.
Results
We present results from three analyses. Using an event study Visual Novelty
difference-in-differences approach (8), we first estimate the The result shown in Fig. 1e highlights that average Visual Novelty
causal impact of adopting generative AI on creative productivity, is decreasing over time among adopters when compared to nona­
artwork value measured as favorites per view, and artifact novelty dopters. The same result holds for the maximum Visual Novelty
with respect to Content and Visual features. Then, using a two- seen in Fig. 1f. This suggests that adopters may be gravitating to­
way fixed effects model, we offer correlational evidence regarding ward a preferred visual style, with relatively minor deviations
how humans’ originality prior to adopting generative AI may in­ from it. This tendency could be influenced by the nature of
fluence postadoption gains in artwork value when artists success­ text-to-image workflows, where prompt engineering tends to fol­
fully explore the creative space. Lastly, we show how adoption of low a formulaic approach to generate consistent, high-quality im­
generative AI may lead to a more dispersed distribution of favor­ ages with a specific style. As is the case with contents, publicly
ites across users on the platform. available fine-tuned checkpoints and adapters for these models
 Zhou and Lee | 3

a b

Downloaded from https://academic.oup.com/pnasnexus/article/3/3/pgae052/7618478 by guest on 05 December 2024

c d

e f

Fig. 1. Causal effect of adopting generative AI on a) creative productivity as the log of monthly posts; b) creative value as number of favorites per view; c)
mean Content Novelty; d) maximum Content Novelty; e) mean Visual Novelty; f) maximum Visual Novelty. The error bars represent 95% CI.

may be designed to capture specific visual elements from which there individual-level differences that enable certain artists to
users can sample from to maintain a particular and consistent successfully produce more creative artworks? Specifically, how
visual style. In effect, AI may be pushing artists toward visual does humans’ baseline novelty, in the absence of AI tools, correl­
homogeneity. ate with their ability to successfully explore novel ideas with gen­
erative AI to produce valuable artifacts? To delve into this
heterogeneity, we categorize each user into quartiles based on
Role of human creativity in AI-assisted their average Content and Visual Novelty without AI assistance
value capture to capture each users’ baseline novelty. We then employ a two-
Although aggregate trends suggest novelty of ideas and aesthetic way fixed effects model to examine the interaction between
features is sharply declining over time with generative AI, are adoption, pretreatment novelty quartiles, and posttreatment
4 | PNAS Nexus, 2024, Vol. 3, No. 3

a b

Downloaded from https://academic.oup.com/pnasnexus/article/3/3/pgae052/7618478 by guest on 05 December 2024

Fig. 2. Estimated effect of increases in mean Content and Visual Novelty on Value post-adoption based on a) average Content Novelty quartiles prior to
treatment; b) average Visual Novelty quartiles prior to treatment. Each point shows the estimated effect of postadoption novelty increases given
creativity levels prior to treatment on Value. The error bars represent 95% CI.

adjustments in novelty. Each point in Fig. 2a and b represents the method (11). This method utilizes ideas from cooperative game
estimated impact of increasing mean Content (left) or Visual theory to approximate the predictive signal of covariates, ac­
(right) Novelty on Value based on artists’ prior novelty denoted counting for linear and nonlinear interactions through the
along the horizontal axis. Intuitively, these estimates quantify Markov chain Monte Carlo method. Intuitively, a feature of great­
the degree to which artists can successfully navigate the creative er importance indicates potentially greater impacts on treatment
space based on prior originality in both ideation and visuals to effect heterogeneity among adopters.
earn more favorable evaluations from peers. Refer to SI Figure 3 offers correlational evidence that Content Novelty sig­
Appendix, Section 2B for estimation details. nificantly increases model performance within several months of
Figure 2a presents correlational evidence that users, regardless adoption, whereas Visual Novelty remains marginally impactful
of their proficiency in generating novel ideas, might be able to until the last observation period. This suggests that Content
realize significant gains in Value if they can successfully produce Novelty plays a more significant role in predicting posttreatment
more novel content with generative AI. The lowest quartile of con­ variations in Value gains compared to Visual Novelty. In sum­
tent creators may also experience marginally significant gains. mary, these findings illustrate that content is king in the
However, those same users who benefit from expressing more text-to-image creative paradigm.
novel ideas may also face penalties for producing more divergent
visuals.
Next, Fig. 2b suggests that users who were proficient in creating Platform-level value capture
exceedingly novel visual features before adopting generative AI One question remains: do individual-level differences within
may garner the most Value gains from successfully introducing adopters result in greater concentrations of value among fewer
more novel ideas. While marginally significant, less proficient users at the platform-level? Specifically, are more favorites being
users can also experience weak Value gains. In general, more nov­ captured by fewer users, or is generative AI promoting less
el ideas are linked to improved Value capture. Conversely, users concentrated value capture? To address these questions, we cal­
capable of producing the most novel visual features may face pen­ culate the Gini coefficients with respect to favorites received of
alties for pushing the boundaries of pixel-level aesthetics with never-treated units, not-yet-treated units, and treated units and
generative AI. This finding might be attributed to the contextual conduct permutation tests with 10,000 iterations to evaluate if
nature of Value, implying an “acceptable range” of novelty. adoption of generative AI may lead to a less concentrated distribu­
Artists already skilled at producing highly novel pixel-level fea­ tion of favorites among users. The Gini coefficient is a common
tures may exceed the limit of what can be considered coherent. measure of aggregate inequality where a coefficient of 0 indicates
Despite penalties for pushing visual boundaries, the gains from that all users make up an equal proportion of favorites earned,
exploring creative ideas with AI outweigh the losses from visual and a coefficient of 1 indicates that a single user captures all favor­
divergence. Unique concepts take priority over novel aesthetics, ites. Thus, higher values of the Gini coefficient indicate a greater
as shown by the larger Value gains for artists who were already concentration of favorites captured by fewer users. Figure 4
adept at Visual Novelty before using AI. This suggests users who depicts the differences in cumulative distributions as well as
naturally lean toward visual exploration may benefit more from Gini coefficients of both control groups and the treated group
generative AI tools to explore the idea space. with respect to a state of perfect equality.
Lastly, we estimate Generalized Random Forests (10) config­ First, observe that platform-level favorites are predominantly
ured to optimize the splitting criteria that maximize heterogen­ captured by a small portion of users, reflecting an aggregate con­
eity in Value gains among adopters for each postadoption centration of favorites. Second, this concentration is more pro­
period. With each trained model, we extract feature importance nounced among not-yet-treated units than among never-treated
weights quantified by the SHAP (SHapley Additive exPlanations) units. Third, despite the presence of aggregate concentration,
 Zhou and Lee | 5

Downloaded from https://academic.oup.com/pnasnexus/article/3/3/pgae052/7618478 by guest on 05 December 2024

Fig. 3. SHAP values measuring importance of mean Content and Visual Novelty on Value gains.

Fig. 4. Gini coefficients of treated units vs. never-treated and not-yet-treated units.

favorites captured among AI adopters are more evenly distributed D shows the difference between the treated coefficient and the
compared to both never-treated and not-yet-treated control units. control group coefficients, show that the differences in coeffi­
The results from the permutation tests in Table 1, where column cients are statistically significant between never-treated and
6 | PNAS Nexus, 2024, Vol. 3, No. 3

Table 1. Permutation tests for statistical significance. Novelty diminish. This implies that the universe of creative possi­
bilities is expanding but with some inefficiencies.
Coefficient D P-value
Our results hint that the widespread adoption of generative AI
Never-treated 0.807 −0.0128 0.0673 technologies in creative fields could lead to a long-run equilibrium
Not-yet-treated 0.824 −0.0298 0.0026 where in aggregate, many artifacts converge to the same types of
Treated 0.794
content or visual features. Creative domains may be inundated
The column D denotes the difference in Gini coefficients relative to the treated with generic content as exploration of the creative space dimin­
population. ishes. Without establishing new frontiers for creative exploration,
AI systems trained on outdated knowledge banks run the risk of
perpetuating the generation of generic content at a mass scale
not-yet-treated groups vs. the treated group. This suggests that in a self-reinforcing cycle (17). Before we reach that point, technol­
generative AI may lead to a broader allocation of favorites earned ogy firms and policy makers pioneering the future of generative AI
(value capture from peer feedback), particularly among control

Downloaded from https://academic.oup.com/pnasnexus/article/3/3/pgae052/7618478 by guest on 05 December 2024

must be sensitive to the potential consequences of such technolo­
units who eventually become adopters. gies in creative fields and society more broadly.
Encouragingly, humans assisted by generative AI who can suc­
cessfully explore more novel ideas may be able to push the cre­
Robustness checks and sensitivity analyses ative frontier, produce meaningful content, and be evaluated
To reinforce the validity of our causal estimates, we employ the favorably by their peers. With respect to traditional theories of
generalized synthetic control method (12) (GSCM). GSCM allows creativity, one particularly useful framework for understanding
us to relax the parallel trends assumption by creating synthetic these results is the theory of blind variation and selective reten­
control units that closely match the pretreatment characteristics tion (BVSR) which posits that creativity is a process of generating
of the treated units while also accounting for unobservable factors new ideas (variation) and consequently selecting the most prom­
that may influence treatment outcome. In addition, we conduct ising ones (retention) (18). The blindness feature suggests that
permutation tests to evaluate the robustness of our estimates to variation is not guided by any specific goal but can also involve
potential measurement errors in treatment time identification evaluating outputs against selection criteria in a genetic algo­
and control group contamination. Our results remain consistent rithm framework (19).
even when utilizing GSCM and in the presence of substantial Because we do not directly observe users’ process, this discus­
measurement error. sion is speculative but suggestive that a text-to-image creative
Because adopting generative AI is subject to selection issues, workflow models after a BVSR genetic process. First, humans ma­
one emergent concern is the case where an artist who experiences nipulate and mutate known creative elements in the form of
renewed interest in creating artworks, and thus is more “inspired,” prompt engineering which requires that the human deconstruct
is also more likely to experiment with text-to-image AI tools and an idea into atomic components, primarily in the form of distinct
explore the creative space as they ramp up production. In this words and phrases, to compose abstract ideas or meanings. Then,
way, unobservable characteristics like a renewed interest in creat­ visual realization of an idea is automated by the algorithm, allow­
ing art or “spark of inspiration” might correlate with adoption of AI ing humans to rapidly sample ideas from their creative space and
tools while driving the main effects rather than AI tools them­ simply evaluate the output against selection criteria. The selec­
selves. Thus, we also provide evidence that unobservable charac­ tion criteria varies based on humans’ ability to make sense of
teristics that may correlate with users’ productivity or “interest” model outputs, and curate those that most align with individual
shocks and selection into treatment are not driving the estimated or peer preferences, thus having direct implications on their
effects by performing a series of falsification tests. For a compre­ evaluation by peers. Satisfactory outputs contribute to the genetic
hensive overview of all robustness checks and sensitivity ana­ evolution of future ideas, prompts, and image refinements.
lyses, please refer to SI Appendix, Section 3. Although we can only observe the published artworks, it is
plausible that many more unobserved iterations of ideation,
prompt engineering, filtering, and refinement have occurred.
This is especially likely given the documented increase in creative
Discussion productivity. Thus, it is possible that individuals with less refined
The rapid adoption of generative AI technologies poses exception­ artistic filters are also less discerning when filtering artworks for
al benefits as well as risks. Current research demonstrates that quality which could lead to a flood of less refined content on plat­
humans, when assisted by generative AI, can significantly in­ forms. In contrast, artists who prioritize coherence and quality
crease productivity in coding (13), ideation (14), and written as­ may only publish artworks that are likely to be evaluated
signments (15) while raising concerns regarding potential favorably.
disinformation (16) and stagnation of knowledge creation (17). The results suggest some evidence in this direction, indicating
Our research is focused on how generative AI is impacting that humans who excel at producing novel ideas before adopting
and potentially coevolving with human creative workflows. In generative AI are evaluated most favorably after adoption if they
our setting, human creativity is embodied through prompts them­ successfully explore the idea space, implying that ability to ma­
selves, whereas in written assignments, generative AI is primarily nipulate novel concepts and curate artworks based on coherence
used to source ideas that are subsequently evaluated by humans, are relevant skills when using text-to-image AI. This aligns with
representing a different paradigm shift in the creative process. prior research which suggest that creative individuals are particu­
Within the first few months post-adoption, text-to-image gen­ larly adept at discerning which ideas are most meaningful (20), re­
erative AI can help individuals produce nearly double the volume flecting a refined sensitivity to the artistic coherence of artifacts
of creative artifacts that are also evaluated 50% more favorably by (21). Furthermore, all artists, regardless of their ability to produce
their peers over time. Moreover, we observe that peak Content novel visual features without generative AI, appear to be eval­
Novelty increases over time, while average Content and Visual uated more favorably if they can capably explore more novel
 Zhou and Lee | 7

ideas. This finding hints at the importance of humans’ baseline and maximum distances for each artifact. This month’s artifacts
ideation and filtering abilities as focal expressions of creativity are then added to the baseline set such that all future artworks are
in a text-to-image paradigm. Finally, generative AI appears to pro­ compared to all prior artworks, effectively capturing the time-
mote a more even distribution of platform-level favorites among varying nature of novelty. Continue for all remaining months.
adopters, signaling a potential step toward an increasingly demo­ We apply this approach to all adopters’ artworks and a random
cratized, inclusive creative domain for artists empowered by AI sample of 10,000 control users due to computational feasibility.
tools.
In summary, our findings emphasize that humans’ ideation
proficiency and a refined artistic filter rather than pure mechanic­ Content feature extraction
al skill may become the focal skills required in a future of human– To describe the focal objects and object relationships in an arti­
AI cocreative process as generative AI becomes more mainstream fact, we utilize state-of-the-art multimodal model BLIP-2 (24)
in creative endeavors. This phenomenon in which AI-assisted which takes as input an image and produces a text description
of the content. A key feature of this approach is the availability

Downloaded from https://academic.oup.com/pnasnexus/article/3/3/pgae052/7618478 by guest on 05 December 2024

artistic creation is driven by ideas and filtering is what we term
“generative synesthesia”—the harmonization of human explor­ of controlled text generation hyperparameters that allow us to
ation and AI exploitation to discover new creative workflows. generate more stable descriptions that are systematically similar
This paradigm shift may provide avenues for creatives to focus in structure, having been trained on 129M images and human-
on what ideas they are representing rather than how they re­ annotated data. BLIP-2 can maintain consistent focus and regu­
present it, opening new opportunities for creative exploration. larity while avoiding the noise added by cross-individual
While concerns about automation loom, society must consider a differences.
future where generative AI is not the source of human stagnation, Given the generated descriptions, we then utilize a pretrained
but rather of symphonic collaboration and human enrichment. text embedding model based on BERT (25), which has demon­
strated state-of-the-art performance on semantic similarity
benchmarks while also being highly efficient, to compute high-
Materials and methods dimensional vector representations for each description. Then,
we apply the algorithm described above to measure Content
Identifying AI adopters
Novelty.
Platform-level policy commonly suggests that users disclose their
use of AI assistance in the form of tags associated with their art­
works. Thus, we employ a rule-based classification scheme. As a Visual feature extraction
first-pass, any artwork published before the original DALL-E in To capture visual features of each artifact at the pixel level, we use
January 2021 is automatically labeled as non-AI generated. a more flexible approach via the self-supervised visual represen­
Then, for all artworks published after January 2021, we examine tation learning algorithm DINOv2 (26) which overcomes the limi­
postlevel title and tags provided by the publishing user. We use tations of standard image-text pretraining approaches where
simple keyword matching (AI-generated, Stable Diffusion, visual features may not be explicitly described in text. Because
Midjourney, DALL-E, etc.) for each post to identify for which we are dealing with creative concepts, this approach is particular­
artworks a user employs AI tools. As a second-pass, we track ly suitable to robustly identify object parts in an image and extract
artworks posts published in AI art communities which may low-level pixel features of images while still exhibiting excellent
not include explicit tags denoting AI assistance. We compile all generalization performance. We compute vector representations
of these artworks and simply label them as AI-generated. of each image such that we can apply the algorithm described
Finally, we assign adoption timing based on the first-known above to obtain measures of Visual Novelty.
AI-generated post for each use (SI Appendix, Fig. S2).

Measuring creative novelty Notes

To measure the two types of novelty, we borrow the idea of con­ a
An AI-generated picture won an art prize. Artists are not happy.
ceptual spaces which can be understood as geometric representa­ b
Artist wins photography contest after submitting AI-generated
tions of entities which capture particular attributes of the
image, then forfeits prize.
artifacts along various dimensions (9, 22). This definition natural­ c
The current legal cases against generative AI are just the beginning.
ly aligns with the concept of embeddings, like word2vec (23), which
capture the relative features of objects in a vector space. This con­
cept can be applied to text passages and images such that meas­
uring the distance between these vector representations
Acknowledgments
captures whether an artifact deviates or converges with a refer­ The authors acknowledge the valuable contributions from their
ence object in the space. Business Insights through Text Lab (BITLAB) research assistants
Using embeddings, we apply the following algorithm: take all Animikh Aich, Aditya Bala, Amrutha Karthikeyan, Audrey Mao,
artifacts published before 2022 April 1, as the baseline set of art­ and Esha Vaishnav in helping to prepare the data for analysis.
works. We use this cutoff because nearly all adoption occurs after Furthermore, the authors are grateful for Stefano Puntoni,
May 2022, so all artifacts in future periods are compared to Alex Burnap, Mi Zhou, Gregory Sun, our audiences at the
non-AI-generated works in the baseline period, and it provides Wharton Business & Generative AI Workshop (23/9/8), INFORMS
an adequate number of pretreatment and posttreatment observa­ Workshop on Data Science (23/10/14), INFORMS Annual Meeting
tions (on average 3 and 7, respectively) for the majority of our (23/10/15) and seminar participants at the University of Wisconsin-
causal sample. Then, take all artifacts published in the following Milwaukee (23/9/22), University of Texas-Dallas (23/10/6), and
month and measure the pairwise cosine distance between those MIT Initiative on the Digital Economy (23/11/29) for their insightful
artifacts and the baseline set, recovering the mean, minimum, comments and feedback.
8 | PNAS Nexus, 2024, Vol. 3, No. 3

Supplementary Material 9 Boden MA. 1998. Creativity and artificial intelligence. Artif Intell.
103(1–2):347–356.
Supplementary material is available at PNAS Nexus online.
10 Athey S, Tibshirani J, Wager S. 2019. Generalized random forests.
Ann Statist. 47(2):1148–1178.
Funding 11 Lundberg S, Lee S-I. 2017. A unified approach to interpreting
model predictions. In: Proceedings of the 31st International
The authors declare no funding.
Conference on Neural Information Processing Systems. p.
4768–4777.
Author Contributions 12 Xu Y. 2017. Generalized synthetic control method: causal infer­
ence with interactive fixed effects models. Polit Anal. 25(1):57–76.
D.L. and E.Z. designed the research and wrote the paper. E.Z.
13 Peng S, Kalliamvakou E, Cihon P, Demirer M. 2023. The impact of
analyzed data and performed research with guidance from D.L.
AI on developer productivity: evidence from github copilot,

Downloaded from https://academic.oup.com/pnasnexus/article/3/3/pgae052/7618478 by guest on 05 December 2024

arXiv, arXiv:2302.06590, preprint: not peer reviewed.
Preprints 14 Noy S, Zhang W. 2023. Experimental evidence on the productiv­
A preprint of this article is available at SSRN. ity effects of generative artificial intelligence. Science. 381(6654):
187–192.
15 Dell’Acqua F, et al. 2023. Navigating the jagged technological fron­
Data Availability tier: field experimental evidence of the effects of AI on knowledge
Replication archive with code is available at Open Science worker productivity and quality. Harvard Business School
Framework at https://osf.io/jfzyp/. Data have been anonymized Technology & Operations Mgt. Unit Working Paper, (24-013).
for the privacy of the users. 16 Spitale G, Biller-Andorno N, Germani F. 2023. AI model GPT-3
(dis)informs us better than humans. Sci Adv. 9(26):eadh1850.
17 Burtch G, Lee D, Chen Z. 2023. The consequences of generative AI
References for UGC and online community engagement. Available at SSRN
1 Dong H-W, Hsiao W-Y, Yang L-C, Yang Y-H. 2018. MuseGAN: 4521754.
multi-track sequential generative adversarial networks for sym­ 18 Campbell DT. 1960. Blind variation and selective retentions in
bolic music generation and accompaniment. In: Proceedings of creative thought as in other knowledge processes. Psychol Rev.
the Thirty-Second AAAI Conference on Artificial Intelligence 67(6):380–400.
and Thirtieth Innovative Applications of Artificial Intelligence 19 Simonton DK. 1999. Creativity as blind variation and selective re­
Conference and Eighth AAAI Symposium on Educational tention: is the creative process Darwinian? Psychol Inq. 10(4):309–
Advances in Artificial Intelligence. p. 34–41. 328.
2 Tan WR, Chan CS, Aguirre H, Tanaka K. 2017. ArtGAN: artwork 20 Silvia PJ. 2008. Discernment and creativity: how well can people
synthesis with conditional categorical GANs. In: 2017 IEEE identify their most creative ideas? Psychol Aesthet Creat Arts. 2(3):
International Conference on Image Processing (ICIP). IEEE. p. 139–146.
3760–3764. 21 Ivcevic Z, Mayer JD. 2009. Mapping dimensions of creativity in
3 Elgammal A, Liu B, Elhoseiny M, Mazzone M. 2017. CAN: creative the life-space. Creat Res J. 21(2–3):152–165.
adversarial networks, generating “art” by learning about styles 22 McGregor S, Wiggins G, Purver M. 2014. Computational creativ­
and deviating from style norms, arXiv, arXiv:1706.07068, pre­ ity: a philosophical approach, and an approach to philosophy.
print: not peer reviewed. In: International Conference on Innovative Computing and
4 Brown TB, et al. 2020. Language models are few-shot learners. Cloud Computing. p. 254–262.
Adv Neural Inf Process Syst. 33:1877–1901. 23 Mikolov T, Chen K, Corrado G, Dean J. 2013. Efficient estimation
5 Rombach R, Blattmann A, Lorenz D, Esser P, Ommer B. 2022. of word representations in vector space, arXiv, arXiv:1301.3781,
High-resolution image synthesis with latent diffusion models. preprint: not peer reviewed.
In: Proceedings of the IEEE/CVF conference on Computer Vision 24 Li J, Li D, Savarese S, Hoi S. 2023. Bootstrapping language-image
and Pattern Recognition. p. 10684–10695. pre-training with frozen image encoders and large language
6 Huang S, Grady P, GPT-3. 2022. Generative AI: a creative new models, arXiv, arXiv:2301.12597, preprint: not peer reviewed.
world. Sequoia Capital US/Europe. https://www.sequoiacap. 25 Reimers N, Gurevych I. 2019. Sentence-bert: sentence embed­
com/article/generative-ai-a-creative-new-world/ dings using siamese bert-networks, arXiv, arXiv:1908.10084, pre­
7 Wollheim R. 1970. Nelson Goodman’s languages of art. J Philos. print: not peer reviewed.
67(16):531–539. 26 Oquab M, et al. 2023. DINOv2: learning robust visual features
8 Callaway B, Sant’Anna PHC. 2021. Difference-in-differences with without supervision, arXiv, arXiv:2304.07193, preprint: not peer
multiple time periods. J Econom. 225(2):200–230. reviewed.