ai art essay

Art for our sake: artists cannot be replaced by machines – study

There has been an explosion of interest in ‘creative AI’, but does this mean that artists will be replaced by machines? No, definitely not, says Anne Ploin , Oxford Internet Institute researcher and one of the team behind today’s report on the potential impact of machine learning (ML) on creative work. 

The report, ‘ AI and the Arts: How Machine Learning is Changing Artistic Work ’ , was co-authored with OII researchers Professor Rebecca Eynon and Dr Isis Hjorth as well as Professor Michael A. Osborne from Oxford’s Department of Engineering .

Their study took place in 2019, a high point for AI in art. It was also a time of high interest around the role of AI (Artificial Intelligence) in the future of work, and particularly around the idea that automation could transform non-manual professions, with a previous study by Professor Michael A. Osborne and Dr Carl Benedict Frey predicting that some 30% of jobs could, technically, be replaced in an AI revolution by 2030.

Human agency in the creative process is never going away. Parts of the creative process can be automated in interesting ways using AI...but the creative decision-making which results in artworks cannot be replicated by current AI technology

Mx Ploin says it was clear from their research that machine learning was becoming a tool for artists – but will not replace artists. She maintains, ‘The main message is that human agency in the creative process is never going away. Parts of the creative process can be automated in interesting ways using AI (generating many versions of an image, for example), but the creative decision-making which results in artworks cannot be replicated by current AI technology.’

She adds, ‘Artistic creativity is about making choices [what material to use, what to draw/paint/create, what message to carry across to an audience] and develops in the context in which an artist works. Art can be a response to a political context, to an artist’s background, to the world we inhabit. This cannot be replicated using machine learning, which is just a data-driven tool. You cannot – for now – transfer life experience into data.’

She adds, ‘AI models can extrapolate in unexpected ways, draw attention to an entirely unrecognised factor in a certain style of painting [from having been trained on hundreds of artworks]. But machine learning models aren’t autonomous.

Artistic creativity is about making choices ...and develops in the context in which an artist works...the world we inhabit. This cannot be replicated using machine learning, which is just a data-driven tool

‘They aren’t going to create new artistic movements on their own – those are PR stories. The real changes that we’re seeing are around the new skills that artists develop to ‘hack’ technical tools, such as machine learning, to make art on their own terms, and around the importance of curation in an increasingly data-driven world.’

The research paper uses a case study of the use of current machine learning techniques in artistic work, and investigates the scope of AI-enhanced creativity and whether human/algorithm synergies may help unlock human creative potential. In doing so, the report breaks down the uncertainty surrounding the application of AI in the creative arts into three key questions.

• How does using generative algorithms alter the creative processes and embodied experiences of artists?
• How do artists sense and reflect upon the relationship between human and machine creative intelligence?
• What is the nature of human/algorithmic creative complementarity?

According to Mx Ploin, ‘We interviewed 14 experts who work in the creative arts, including media and fine artists whose work centred around generative ML techniques. We also talked to curators and researchers in this field. This allowed us to develop fuller understanding of the implications of AI – ranging from automation to complementarity – in a domain at the heart of human experience: creativity.’

They found a range of responses to the use of machine learning and AI. New activities required by using ML models involved both continuity with previous creative processes and rupture from past practices. There were major changes around the generative process, the evolving ways ML outputs were conceptualised, and artists’ embodied experiences of their practice.

And, says the researcher, there were similarities between the use of machine learning and previous periods in art history, such as the code-based and computer arts of the 1960s and 1970s. But the use of ML models was a “step change” from past tools, according to many artists.

While the machine learning models could help produce ‘surprising variations of existing images’, practitioners felt the artist remained irreplaceable...in making artworks

But, she maintains, while the machine learning models could help produce ‘surprising variations of existing images’, practitioners felt the artist remained irreplaceable in terms of giving images artistic context and intention – that is, in making artworks.

Ultimately, most agreed that despite the increased affordances of ML technologies, the relationship between artists and their media remained essentially unchanged, as artists ultimately work to address human – rather than technical – questions.

Don’t let it put you off going to art school. We need more artists

The report concludes that human/ML complementarity in the arts is a rich and ongoing process, with contemporary artists continuously exploring and expanding technological capabilities to make artworks . Although ML-based processes raise challenges around skills, a common language, resources, and inclusion, what is clear is that the future of ML arts will belong to those with both technical and artistic skills. There is more to come.

But, says Mx Ploin, ‘Don’t let it put you off going to art school. We need more artists.’

Further information

AI and the Arts: How Machine Learning is Changing Artistic Work . Ploin, A., Eynon, R., Hjorth I. & Osborne, M.A. (2022). Report from the Creative Algorithmic Intelligence Research Project. Oxford Internet Institute, University of Oxford, UK. Download the full report .

This report accounts for the findings of the 'Creative Algorithmic Intelligence: Capabilities and Complementarity' project, which ran between 2019 and 2021 as a collaboration between the University of Oxford's Department of Engineering and Oxford Internet Institute.

The report also showcases a range of artworks from contemporary artists who use AI as part of their practice and who participated in our study: Robbie Barrat , Nicolas Boillot , Sofia Crespo , Jake Elwes , Lauren Lee McCarthy , Sarah Meyohas , Anna Ridler , Helena Sarin , and David Young.

DISCOVER MORE

• Support Oxford's research
• Partner with Oxford on research
• Study at Oxford
• Research jobs at Oxford

You can view all news or browse by category

Grab your spot at the free arXiv Accessibility Forum

Help | Advanced Search

Computer Science > Computer Vision and Pattern Recognition

Title: understanding and creating art with ai: review and outlook.

Abstract: Technologies related to artificial intelligence (AI) have a strong impact on the changes of research and creative practices in visual arts. The growing number of research initiatives and creative applications that emerge in the intersection of AI and art, motivates us to examine and discuss the creative and explorative potentials of AI technologies in the context of art. This paper provides an integrated review of two facets of AI and art: 1) AI is used for art analysis and employed on digitized artwork collections; 2) AI is used for creative purposes and generating novel artworks. In the context of AI-related research for art understanding, we present a comprehensive overview of artwork datasets and recent works that address a variety of tasks such as classification, object detection, similarity retrieval, multimodal representations, computational aesthetics, etc. In relation to the role of AI in creating art, we address various practical and theoretical aspects of AI Art and consolidate related works that deal with those topics in detail. Finally, we provide a concise outlook on the future progression and potential impact of AI technologies on our understanding and creation of art.

Submission history

Access paper:.

• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

DBLP - CS Bibliography

Bibtex formatted citation.

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

• Search Menu
• Sign in through your institution
• Advance articles
• Browse content in Biological, Health, and Medical Sciences
• Administration Of Health Services, Education, and Research
• Agricultural Sciences
• Allied Health Professions
• Anesthesiology
• Anthropology
• Anthropology (Biological, Health, and Medical Sciences)
• Applied Biological Sciences
• Biochemistry
• Biophysics and Computational Biology (Biological, Health, and Medical Sciences)
• Biostatistics
• Cell Biology
• Dermatology
• Developmental Biology
• Environmental Sciences (Biological, Health, and Medical Sciences)
• Immunology and Inflammation
• Internal Medicine
• Medical Sciences
• Medical Microbiology
• Microbiology
• Neuroscience
• Obstetrics and Gynecology
• Ophthalmology
• Pharmacology
• Physical Medicine
• Plant Biology
• Population Biology
• Psychological and Cognitive Sciences (Biological, Health, and Medical Sciences)
• Public Health and Epidemiology
• Radiation Oncology
• Rehabilitation
• Sustainability Science (Biological, Health, and Medical Sciences)
• Systems Biology
• Veterinary Medicine
• Browse content in Physical Sciences and Engineering
• Aerospace Engineering
• Applied Mathematics
• Applied Physical Sciences
• Bioengineering
• Biophysics and Computational Biology (Physical Sciences and Engineering)
• Chemical Engineering
• Civil and Environmental Engineering
• Computer Sciences
• Computer Science and Engineering
• Earth Resources Engineering
• Earth, Atmospheric, and Planetary Sciences
• Electric Power and Energy Systems Engineering
• Electronics, Communications and Information Systems Engineering
• Engineering
• Environmental Sciences (Physical Sciences and Engineering)
• Materials Engineering
• Mathematics
• Mechanical Engineering
• Sustainability Science (Physical Sciences and Engineering)
• Browse content in Social and Political Sciences
• Anthropology (Social and Political Sciences)
• Economic Sciences
• Environmental Sciences (Social and Political Sciences)
• Political Sciences
• Psychological and Cognitive Sciences (Social and Political Sciences)
• Social Sciences
• Sustainability Science (Social and Political Sciences)
• Author guidelines
• Submission site
• Open access policy
• Self-archiving policy
• Why submit to PNAS Nexus
• The PNAS portfolio
• For reviewers
• About PNAS Nexus
• About National Academy of Sciences
• Editorial Board
• Journals on Oxford Academic
• Books on Oxford Academic

Article Contents

Introduction, robustness checks and sensitivity analyses, materials and methods, acknowledgments, supplementary material, author contributions, data availability.

• < Previous

Generative artificial intelligence, human creativity, and art

Competing Interest: The authors declare no competing interest.

• Article contents
• Figures & tables
• Supplementary Data

Eric Zhou, Dokyun Lee, Generative artificial intelligence, human creativity, and art, PNAS Nexus , Volume 3, Issue 3, March 2024, pgae052, https://doi.org/10.1093/pnasnexus/pgae052

• Permissions Icon Permissions

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.

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 exploration. 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.

Recently, artificial intelligence (AI) has exhibited that it can feasibly produce outputs that society traditionally would judge as creative. Specifically, generative algorithms have been leveraged to automatically generate creative artifacts like music ( 1 ), digital artworks ( 2 , 3 ), and stories ( 4 ). Such generative models allow humans to directly engage in the creative process through text-to-image systems (e.g. Midjourney, Stable Diffusion, DALL-E) based on the latent diffusion model ( 5 ) or by participating in an open dialog with transformer-based language models (e.g. ChatGPT, Bard, Claude). Generative AI is projected to become more potent to automate even more creative tasks traditionally reserved for humans and generate significant economic value in the years to come ( 6 ).

Many such generative algorithms were released in the past year, and their diffusion into creative domains has concerned many artistic communities which perceive generative AI as a threat to substitute the natural human ability to be creative. Text-to-image generative AI has emerged as a candidate system that automates elements of humans’ creative process in producing high-quality digital artworks. Remarkably, an artwork created by Midjourney bested human artists in an art competition, a while another artist refused to accept the top prize in a photo competition after winning, citing ethical concerns. b Artists have filed lawsuits against the founding companies of some of the most prominent text-to-image generators, arguing that generative AI steals from the works upon which the models are trained and infringes on the copyrights of artists. c This has ignited a broader debate regarding the originality of AI-generated content and the extent to which it may replace human creativity, a faculty that many consider unique to humans. While generative AI has demonstrated the capability to automatically create new digital artifacts, there remains a significant knowledge gap regarding its impact on productivity in artistic endeavors which lack well-defined objectives, and the long-run implications on human creativity more broadly. In particular, if humans increasingly rely on generative AI for content creation, creative fields may become saturated with generic content, potentially stifling exploration of new creative frontiers. Given that generative algorithms will remain a mainstay in creative domains as it continues to mature, it is critical to understand how generative AI is affecting creative production, the evaluation of creative artifacts, and human creativity more broadly. To this end, our research questions are 3-fold:

How does the adoption of generative AI affect humans’ creative production?

Is generative AI enabling humans to produce more creative content?

When and for whom does the adoption of generative AI lead to more creative and valuable artifacts?

Our analyses of over 53,000 artists and 5,800 known AI adopters on one of the largest art-sharing platforms reveal that creative productivity and artwork value, measured as favorites per view, significantly increased with the adoption of text-to-image systems.

We then focus our analysis on creative novelty. A simplified view of human creative novelty with respect to art can be summarized via two main channels through which humans can inject creativity into an artifact: Contents and Visuals . These concepts are rooted in the classical philosophy of symbolism in art which suggests that the contents of an artwork is related to the meaning or subject matter, whereas visuals are simply the physical elements used to convey the content ( 7 ). In our setting, Contents concern the focal object(s) and relations depicted in an artifact, whereas Visuals consider the pixel-level stylistic elements of an artifact. Thus, Content and Visual Novelty are measured as the pairwise cosine distance between artifacts in the feature space (see Materials and methods for details on feature extraction and how novelty is measured).

Our analyses reveal that over time, adopters’ artworks exhibit decreasing novelty, both in terms of Concepts and Visual features. However, maximum Content Novelty increases, suggesting an expanding yet inefficient idea space. At the individual level, artists who harness generative AI while successfully exploring more innovative ideas, irrespective of their prior originality, may earn more favorable evaluations from their peers. In addition, the adoption of generative AI leads to a less concentrated distribution of favorites earned among adopters.

We present results from three analyses. Using an event study difference-in-differences approach ( 8 ), we first estimate the causal impact of adopting generative AI on creative productivity, artwork value measured as favorites per view, and artifact novelty with respect to Content and Visual features. Then, using a two-way fixed effects model, we offer correlational evidence regarding how humans’ originality prior to adopting generative AI may influence postadoption gains in artwork value when artists successfully explore the creative space. Lastly, we show how adoption of generative AI may lead to a more dispersed distribution of favorites across users on the platform.

Creative productivity

We define creative productivity as the log of the number of artifacts that a user posts in a month. Figure 1 a reveals that upon adoption, artists experience a 50% increase in productivity on average, which then doubles in the subsequent month. For the average user, this translates to approximately 7 additional artifacts published in the adoption month and 15 artifacts in the following month. Beyond the adoption month, user productivity gradually stabilizes to a level that still exceeds preadoption volume. By automating the execution stage of the creative process, adopters can experience prolonged productivity gains compared to their nonadopter counterparts.

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.

Creative value

If users are becoming more productive, what of the quality of the artifacts they are producing? We next examine how adopters’ artifacts are evaluated by their peers over time. In the literature, creative Value is intended to measure some aspect of utility, performance and/or attractiveness of an artifact, subject to temporal and cultural contexts ( 9 ). Given this subjectivity, we measure Value as the number of favorites an artwork receives per view after 2 weeks, reflecting its overall performance and contextual relevance within the community. This metric also hints at the artwork’s broader popularity within the cultural climate, suggesting a looser definition of Value based on cultural trends. Throughout the paper, the term “Value” will refer to these two notions.

Figure 1 b reveals an initial nonsignificant upward trend in the Value of artworks produced by AI adopters. But after 3 months, AI adopters consistently produce artworks judged significantly more valuable than those of nonadopters. This translates to a 50% increase in artwork favorability by the sixth month, jumping from the preadoption average of 2% to a steady 3% rate of earning a favorite per view.

Content Novelty

Figure 1 c shows that average Content Novelty decreases over time among adopters, meaning that the focal objects and themes within new artworks produced by AI adopters are becoming progressively more alike over time when compared to control units. Intuitively, this is equivalent to adopters’ ideas becoming more similar over time. In practice, many publicly available fine-tuned checkpoints and adapters are refined to enable text-to-image models to produce specific contents with consistency. Figure 1 d, however, reveals that maximum Content Novelty is increasing and marginally statistically significantly within the first several months after adoption. This suggests two possibilities: either a 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.

Visual Novelty

The result shown in Fig. 1 e highlights that average Visual Novelty is decreasing over time among adopters when compared to nonadopters. The same result holds for the maximum Visual Novelty seen in Fig. 1 f. This suggests that adopters may be gravitating toward a preferred visual style, with relatively minor deviations from it. This tendency could be influenced by the nature of text-to-image workflows, where prompt engineering tends to follow a formulaic approach to generate consistent, high-quality images with a specific style. As is the case with contents, publicly available fine-tuned checkpoints and adapters for these models may be designed to capture specific visual elements from which users can sample from to maintain a particular and consistent visual style. In effect, AI may be pushing artists toward visual homogeneity.

Role of human creativity in AI-assisted value capture

Although aggregate trends suggest novelty of ideas and aesthetic features is sharply declining over time with generative AI, are there individual-level differences that enable certain artists to successfully produce more creative artworks? Specifically, how does humans’ baseline novelty, in the absence of AI tools, correlate with their ability to successfully explore novel ideas with generative AI to produce valuable artifacts? To delve into this heterogeneity, we categorize each user into quartiles based on their average Content and Visual Novelty without AI assistance to capture each users’ baseline novelty. We then employ a two-way fixed effects model to examine the interaction between adoption, pretreatment novelty quartiles, and posttreatment adjustments in novelty. Each point in Fig. 2 a and b represents the estimated impact of increasing mean Content (left) or Visual (right) Novelty on Value based on artists’ prior novelty denoted along the horizontal axis. Intuitively, these estimates quantify the degree to which artists can successfully navigate the creative space based on prior originality in both ideation and visuals to earn more favorable evaluations from peers. Refer to SI Appendix, Section 2B for estimation details.

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.

Figure 2 a presents correlational evidence that users, regardless of their proficiency in generating novel ideas, might be able to realize significant gains in Value if they can successfully produce more novel content with generative AI. The lowest quartile of content creators may also experience marginally significant gains. However, those same users who benefit from expressing more novel ideas may also face penalties for producing more divergent visuals.

Next, Fig. 2 b suggests that users who were proficient in creating exceedingly novel visual features before adopting generative AI may garner the most Value gains from successfully introducing more novel ideas. While marginally significant, less proficient users can also experience weak Value gains. In general, more novel ideas are linked to improved Value capture. Conversely, users capable of producing the most novel visual features may face penalties for pushing the boundaries of pixel-level aesthetics with generative AI. This finding might be attributed to the contextual nature of Value, implying an “acceptable range” of novelty. Artists already skilled at producing highly novel pixel-level features may exceed the limit of what can be considered coherent.

Despite penalties for pushing visual boundaries, the gains from exploring creative ideas with AI outweigh the losses from visual divergence. Unique concepts take priority over novel aesthetics, as shown by the larger Value gains for artists who were already adept at Visual Novelty before using AI. This suggests users who naturally lean toward visual exploration may benefit more from generative AI tools to explore the idea space.

Lastly, we estimate Generalized Random Forests ( 10 ) configured to optimize the splitting criteria that maximize heterogeneity in Value gains among adopters for each postadoption period. With each trained model, we extract feature importance weights quantified by the SHAP (SHapley Additive exPlanations) method ( 11 ). This method utilizes ideas from cooperative game theory to approximate the predictive signal of covariates, accounting for linear and nonlinear interactions through the Markov chain Monte Carlo method. Intuitively, a feature of greater importance indicates potentially greater impacts on treatment effect heterogeneity among adopters.

Figure 3 offers correlational evidence that Content Novelty significantly increases model performance within several months of adoption, whereas Visual Novelty remains marginally impactful until the last observation period. This suggests that Content Novelty plays a more significant role in predicting posttreatment variations in Value gains compared to Visual Novelty. In summary, these findings illustrate that content is king in the text-to-image creative paradigm.

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

Platform-level value capture

One question remains: do individual-level differences within adopters result in greater concentrations of value among fewer users at the platform-level? Specifically, are more favorites being captured by fewer users, or is generative AI promoting less concentrated value capture? To address these questions, we calculate the Gini coefficients with respect to favorites received of never-treated units, not-yet-treated units, and treated units and conduct permutation tests with 10,000 iterations to evaluate if adoption of generative AI may lead to a less concentrated distribution of favorites among users. The Gini coefficient is a common measure of aggregate inequality where a coefficient of 0 indicates that all users make up an equal proportion of favorites earned, and a coefficient of 1 indicates that a single user captures all favorites. Thus, higher values of the Gini coefficient indicate a greater concentration of favorites captured by fewer users. Figure 4 depicts the differences in cumulative distributions as well as Gini coefficients of both control groups and the treated group with respect to a state of perfect equality.

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

First, observe that platform-level favorites are predominantly captured by a small portion of users, reflecting an aggregate concentration of favorites. Second, this concentration is more pronounced among not-yet-treated units than among never-treated units. Third, despite the presence of aggregate concentration, favorites captured among AI adopters are more evenly distributed compared to both never-treated and not-yet-treated control units. The results from the permutation tests in Table 1 , where column D shows the difference between the treated coefficient and the control group coefficients, show that the differences in coefficients are statistically significant between never-treated and not-yet-treated groups vs. the treated group. This suggests that generative AI may lead to a broader allocation of favorites earned (value capture from peer feedback), particularly among control units who eventually become adopters.

Permutation tests for statistical significance.

The column D denotes the difference in Gini coefficients relative to the treated population.

To reinforce the validity of our causal estimates, we employ the generalized synthetic control method ( 12 ) (GSCM). GSCM allows us to relax the parallel trends assumption by creating synthetic control units that closely match the pretreatment characteristics of the treated units while also accounting for unobservable factors that may influence treatment outcome. In addition, we conduct permutation tests to evaluate the robustness of our estimates to potential measurement errors in treatment time identification and control group contamination. Our results remain consistent even when utilizing GSCM and in the presence of substantial measurement error.

Because adopting generative AI is subject to selection issues, one emergent concern is the case where an artist who experiences renewed interest in creating artworks, and thus is more “inspired,” is also more likely to experiment with text-to-image AI tools and explore the creative space as they ramp up production. In this way, unobservable characteristics like a renewed interest in creating art or “spark of inspiration” might correlate with adoption of AI tools while driving the main effects rather than AI tools themselves. Thus, we also provide evidence that unobservable characteristics that may correlate with users’ productivity or “interest” shocks and selection into treatment are not driving the estimated effects by performing a series of falsification tests. For a comprehensive overview of all robustness checks and sensitivity analyses, please refer to SI Appendix, Section 3 .

The rapid adoption of generative AI technologies poses exceptional benefits as well as risks. Current research demonstrates that humans, when assisted by generative AI, can significantly increase productivity in coding ( 13 ), ideation ( 14 ), and written assignments ( 15 ) while raising concerns regarding potential disinformation ( 16 ) and stagnation of knowledge creation ( 17 ). Our research is focused on how generative AI is impacting and potentially coevolving with human creative workflows. In our setting, human creativity is embodied through prompts themselves, whereas in written assignments, generative AI is primarily used to source ideas that are subsequently evaluated by humans, representing a different paradigm shift in the creative process.

Within the first few months post-adoption, text-to-image generative AI can help individuals produce nearly double the volume of creative artifacts that are also evaluated 50% more favorably by their peers over time. Moreover, we observe that peak Content Novelty increases over time, while average Content and Visual Novelty diminish. This implies that the universe of creative possibilities is expanding but with some inefficiencies.

Our results hint that the widespread adoption of generative AI technologies in creative fields could lead to a long-run equilibrium where in aggregate, many artifacts converge to the same types of content or visual features. Creative domains may be inundated with generic content as exploration of the creative space diminishes. 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 in a self-reinforcing cycle ( 17 ). Before we reach that point, technology firms and policy makers pioneering the future of generative AI must be sensitive to the potential consequences of such technologies in creative fields and society more broadly.

Encouragingly, humans assisted by generative AI who can successfully explore more novel ideas may be able to push the creative frontier, produce meaningful content, and be evaluated favorably by their peers. With respect to traditional theories of creativity, one particularly useful framework for understanding these results is the theory of blind variation and selective retention (BVSR) which posits that creativity is a process of generating new ideas (variation) and consequently selecting the most promising ones (retention) ( 18 ). The blindness feature suggests that variation is not guided by any specific goal but can also involve evaluating outputs against selection criteria in a genetic algorithm framework ( 19 ).

Because we do not directly observe users’ process, this discussion is speculative but suggestive that a text-to-image creative workflow models after a BVSR genetic process. First, humans manipulate and mutate known creative elements in the form of prompt engineering which requires that the human deconstruct an idea into atomic components, primarily in the form of distinct words and phrases, to compose abstract ideas or meanings. Then, visual realization of an idea is automated by the algorithm, allowing humans to rapidly sample ideas from their creative space and simply evaluate the output against selection criteria. The selection criteria varies based on humans’ ability to make sense of model outputs, and curate those that most align with individual or peer preferences, thus having direct implications on their evaluation by peers. Satisfactory outputs contribute to the genetic evolution of future ideas, prompts, and image refinements.

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 productivity. Thus, it is possible that individuals with less refined artistic filters are also less discerning when filtering artworks for quality which could lead to a flood of less refined content on platforms. In contrast, artists who prioritize coherence and quality may only publish artworks that are likely to be evaluated favorably.

The results suggest some evidence in this direction, indicating that humans who excel at producing novel ideas before adopting generative AI are evaluated most favorably after adoption if they successfully explore the idea space, implying that ability to manipulate novel concepts and curate artworks based on coherence are relevant skills when using text-to-image AI. This aligns with prior research which suggest that creative individuals are particularly adept at discerning which ideas are most meaningful ( 20 ), reflecting a refined sensitivity to the artistic coherence of artifacts ( 21 ). Furthermore, all artists, regardless of their ability to produce novel visual features without generative AI, appear to be evaluated more favorably if they can capably explore more novel ideas. This finding hints at the importance of humans’ baseline ideation and filtering abilities as focal expressions of creativity in a text-to-image paradigm. Finally, generative AI appears to promote a more even distribution of platform-level favorites among adopters, signaling a potential step toward an increasingly democratized, inclusive creative domain for artists empowered by AI tools.

In summary, our findings emphasize that humans’ ideation proficiency and a refined artistic filter rather than pure mechanical skill may become the focal skills required in a future of human–AI cocreative process as generative AI becomes more mainstream in creative endeavors. This phenomenon in which AI-assisted artistic creation is driven by ideas and filtering is what we term “generative synesthesia”—the harmonization of human exploration and AI exploitation to discover new creative workflows. This paradigm shift may provide avenues for creatives to focus on what ideas they are representing rather than how they represent it, opening new opportunities for creative exploration. While concerns about automation loom, society must consider a future where generative AI is not the source of human stagnation, but rather of symphonic collaboration and human enrichment.

Identifying AI adopters

Platform-level policy commonly suggests that users disclose their use of AI assistance in the form of tags associated with their artworks. Thus, we employ a rule-based classification scheme. As a first-pass, any artwork published before the original DALL-E in January 2021 is automatically labeled as non-AI generated. Then, for all artworks published after January 2021, we examine postlevel title and tags provided by the publishing user. We use simple keyword matching (AI-generated, Stable Diffusion, Midjourney, DALL-E, etc.) for each post to identify for which artworks a user employs AI tools. As a second-pass, we track artworks posts published in AI art communities which may not include explicit tags denoting AI assistance. We compile all of these artworks and simply label them as AI-generated. Finally, we assign adoption timing based on the first-known AI-generated post for each use ( SI Appendix, Fig. S2 ).

Measuring creative novelty

To measure the two types of novelty, we borrow the idea of conceptual spaces which can be understood as geometric representations of entities which capture particular attributes of the artifacts along various dimensions ( 9 , 22 ). This definition naturally aligns with the concept of embeddings, like word2vec ( 23 ), which capture the relative features of objects in a vector space. This concept can be applied to text passages and images such that measuring the distance between these vector representations captures whether an artifact deviates or converges with a reference object in the space.

Using embeddings, we apply the following algorithm: take all artifacts published before 2022 April 1, as the baseline set of artworks. We use this cutoff because nearly all adoption occurs after May 2022, so all artifacts in future periods are compared to non-AI-generated works in the baseline period, and it provides an adequate number of pretreatment and posttreatment observations (on average 3 and 7, respectively) for the majority of our causal sample. Then, take all artifacts published in the following month and measure the pairwise cosine distance between those artifacts and the baseline set, recovering the mean, minimum, and maximum distances for each artifact. This month’s artifacts are then added to the baseline set such that all future artworks are compared to all prior artworks, effectively capturing the time-varying nature of novelty. Continue for all remaining months. We apply this approach to all adopters’ artworks and a random sample of 10,000 control users due to computational feasibility.

Content feature extraction

To describe the focal objects and object relationships in an artifact, we utilize state-of-the-art multimodal model BLIP-2 ( 24 ) which takes as input an image and produces a text description of the content. A key feature of this approach is the availability of controlled text generation hyperparameters that allow us to generate more stable descriptions that are systematically similar in structure, having been trained on 129M images and human-annotated data. BLIP-2 can maintain consistent focus and regularity while avoiding the noise added by cross-individual differences.

Given the generated descriptions, we then utilize a pretrained text embedding model based on BERT ( 25 ), which has demonstrated state-of-the-art performance on semantic similarity benchmarks while also being highly efficient, to compute high-dimensional vector representations for each description. Then, we apply the algorithm described above to measure Content Novelty.

Visual feature extraction

To capture visual features of each artifact at the pixel level, we use a more flexible approach via the self-supervised visual representation learning algorithm DINOv2 ( 26 ) which overcomes the limitations of standard image-text pretraining approaches where visual features may not be explicitly described in text. Because we are dealing with creative concepts, this approach is particularly suitable to robustly identify object parts in an image and extract low-level pixel features of images while still exhibiting excellent generalization performance. We compute vector representations of each image such that we can apply the algorithm described above to obtain measures of Visual Novelty.

An AI-generated picture won an art prize. Artists are not happy.

Artist wins photography contest after submitting AI-generated image, then forfeits prize.

The current legal cases against generative AI are just the beginning.

The authors acknowledge the valuable contributions from their Business Insights through Text Lab (BITLAB) research assistants Animikh Aich, Aditya Bala, Amrutha Karthikeyan, Audrey Mao, and Esha Vaishnav in helping to prepare the data for analysis. Furthermore, the authors are grateful for Stefano Puntoni, Alex Burnap, Mi Zhou, Gregory Sun, our audiences at the Wharton Business & Generative AI Workshop (23/9/8), INFORMS Workshop on Data Science (23/10/14), INFORMS Annual Meeting (23/10/15) and seminar participants at the University of Wisconsin-Milwaukee (23/9/22), University of Texas-Dallas (23/10/6), and MIT Initiative on the Digital Economy (23/11/29) for their insightful comments and feedback.

Supplementary material is available at PNAS Nexus online.

The authors declare no funding.

D.L. and E.Z. designed the research and wrote the paper. E.Z. analyzed data and performed research with guidance from D.L.

A preprint of this article is available at SSRN .

Replication archive with code is available at Open Science Framework at https://osf.io/jfzyp/ . Data have been anonymized for the privacy of the users.

Dong   H-W , Hsiao   W-Y , Yang   L-C , Yang   Y-H . 2018 . MuseGAN: multi-track sequential generative adversarial networks for symbolic music generation and accompaniment. In: Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence . p. 34–41.

Tan   WR , Chan   CS , Aguirre   H , Tanaka   K . 2017 . ArtGAN: artwork synthesis with conditional categorical GANs. In: 2017 IEEE International Conference on Image Processing (ICIP). IEEE. p. 3760–3764 .

Elgammal   A , Liu   B , Elhoseiny   M , Mazzone   M . 2017 . CAN: creative adversarial networks, generating “art” by learning about styles and deviating from style norms, arXiv, arXiv:1706.07068, preprint: not peer reviewed .

Brown   TB , et al . 2020 . Language models are few-shot learners. Adv Neural Inf Process Syst. 33:1877–1901 .

Rombach   R , Blattmann   A , Lorenz   D , Esser   P , Ommer   B . 2022 . High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. p. 10684–10695 .

Huang   S , Grady   P , GPT-3 . 2022 . Generative AI: a creative new world. Sequoia Capital US/Europe. https://www.sequoiacap.com/article/generative-ai-a-creative-new-world/

Wollheim   R . 1970 . Nelson Goodman’s languages of art . J Philos . 67 ( 16 ): 531 – 539 .

Google Scholar

Callaway   B , Sant’Anna   PHC . 2021 . Difference-in-differences with multiple time periods . J Econom . 225 ( 2 ): 200 – 230 .

Boden   MA . 1998 . Creativity and artificial intelligence. Artif Intell. 103(1–2):347–356 .

Athey   S , Tibshirani   J , Wager   S . 2019 . Generalized random forests. Ann Statist . 47(2):1148–1178 .

Lundberg   S , Lee   S-I . 2017 . A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. p. 4768–4777 .

Xu   Y . 2017 . Generalized synthetic control method: causal inference with interactive fixed effects models . Polit Anal . 25 ( 1 ): 57 – 76 .

Peng   S , Kalliamvakou   E , Cihon   P , Demirer   M . 2023 . The impact of AI on developer productivity: evidence from github copilot, arXiv, arXiv:2302.06590, preprint: not peer reviewed .

Noy   S , Zhang   W . 2023 . Experimental evidence on the productivity effects of generative artificial intelligence . Science . 381 ( 6654 ): 187 – 192 .

Dell’Acqua   F , et al . 2023 . Navigating the jagged technological frontier: field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Technology & Operations Mgt. Unit Working Paper, (24-013) .

Spitale   G , Biller-Andorno   N , Germani   F . 2023 . AI model GPT-3 (dis)informs us better than humans . Sci Adv . 9 ( 26 ): eadh1850 .

Burtch   G , Lee   D , Chen   Z . 2023 . The consequences of generative AI for UGC and online community engagement. Available at SSRN 4521754 .

Campbell   DT . 1960 . Blind variation and selective retentions in creative thought as in other knowledge processes . Psychol Rev . 67 ( 6 ): 380 – 400 .

Simonton   DK . 1999 . Creativity as blind variation and selective retention: is the creative process Darwinian? Psychol Inq. 10(4):309–328 .

Silvia   PJ . 2008 . Discernment and creativity: how well can people identify their most creative ideas?   Psychol Aesthet Creat Arts . 2 ( 3 ): 139 – 146 .

Ivcevic   Z , Mayer   JD . 2009 . Mapping dimensions of creativity in the life-space . Creat Res J . 21 ( 2–3 ): 152 – 165 .

McGregor   S , Wiggins   G , Purver   M . 2014 . Computational creativity: a philosophical approach, and an approach to philosophy. In: International Conference on Innovative Computing and Cloud Computing. p. 254–262 .

Mikolov   T , Chen   K , Corrado   G , Dean   J . 2013 . Efficient estimation of word representations in vector space, arXiv, arXiv:1301.3781, preprint: not peer reviewed .

Li   J , Li   D , Savarese   S , Hoi   S . 2023 . Bootstrapping language-image pre-training with frozen image encoders and large language models, arXiv, arXiv:2301.12597, preprint: not peer reviewed .

Reimers   N , Gurevych   I . 2019 . Sentence-bert: sentence embeddings using siamese bert-networks, arXiv, arXiv:1908.10084, preprint: not peer reviewed .

Oquab   M , et al . 2023 . DINOv2: learning robust visual features without supervision, arXiv, arXiv:2304.07193, preprint: not peer reviewed .

Author notes

Supplementary data.

Email alerts

Citing articles via.

• Contact PNAS Nexus
• Advertising and Corporate Services
• Journals Career Network

Affiliations

• Online ISSN 2752-6542
• Copyright © 2024 National Academy of Sciences
• About Oxford Academic
• Publish journals with us
• University press partners
• What we publish
• New features  
• Open access
• Institutional account management
• Rights and permissions
• Get help with access
• Accessibility
• Advertising
• Media enquiries
• Oxford University Press
• Oxford Languages
• University of Oxford

Oxford University Press is a department of the University of Oxford. It furthers the University's objective of excellence in research, scholarship, and education by publishing worldwide

• Copyright © 2024 Oxford University Press
• Cookie settings
• Cookie policy
• Privacy policy
• Legal notice

This Feature Is Available To Subscribers Only

Sign In or Create an Account

This PDF is available to Subscribers Only

For full access to this pdf, sign in to an existing account, or purchase an annual subscription.

Suggestions or feedback?

MIT News | Massachusetts Institute of Technology

• Machine learning
• Sustainability
• Black holes
• Classes and programs

Departments

• Aeronautics and Astronautics
• Brain and Cognitive Sciences
• Architecture
• Political Science
• Mechanical Engineering

Centers, Labs, & Programs

• Abdul Latif Jameel Poverty Action Lab (J-PAL)
• Picower Institute for Learning and Memory
• Lincoln Laboratory
• School of Architecture + Planning
• School of Engineering
• School of Humanities, Arts, and Social Sciences
• Sloan School of Management
• School of Science
• MIT Schwarzman College of Computing

If art is how we express our humanity, where does AI fit in?

Press contact :, media download.

*Terms of Use:

Images for download on the MIT News office website are made available to non-commercial entities, press and the general public under a Creative Commons Attribution Non-Commercial No Derivatives license . You may not alter the images provided, other than to crop them to size. A credit line must be used when reproducing images; if one is not provided below, credit the images to "MIT."

Previous image Next image

The rapid advance of artificial intelligence has generated a lot of buzz, with some predicting it will lead to an idyllic utopia and others warning it will bring the end of humanity. But speculation about where AI technology is going, while important, can also drown out important conversations about how we should be handling the AI technologies available today.

One such technology is generative AI, which can create content including text, images, audio, and video. Popular generative AIs like the chatbot ChatGPT generate conversational text based on training data taken from the internet.

Today a group of 14 researchers from a number of organizations including MIT published a commentary article in Science that helps set the stage for discussions about generative AI’s immediate impact on creative work and society more broadly. The paper’s MIT-affiliated co-authors include Media Lab postdoc Ziv Epstein SM ’19, PhD ’23; Matt Groh SM ’19, PhD ’23; PhD students Rob Mahari ’17 and Hope Schroeder; and Professor Alex "Sandy" Pentland.

MIT News spoke with Epstein, the lead author of the paper.

Q: Why did you write this paper?

A: Generative AI tools are doing things that even a few years ago we never thought would be possible. This raises a lot of fundamental questions about the creative process and the human’s role in creative production. Are we going to get automated out of jobs? How are we going to preserve the human aspect of creativity with all of these new technologies?

The complexity of black-box AI systems can make it hard for researchers and the broader public to understand what’s happening under the hood, and what the impacts of these tools on society will be. Many discussions about AI anthropomorphize the technology, implicitly suggesting these systems exhibit human-like intent, agency, or self-awareness. Even the term “artificial intelligence” reinforces these beliefs: ChatGPT uses first-person pronouns, and we say AIs “hallucinate.” These agentic roles we give AIs can undermine the credit to creators whose labor underlies the system’s outputs, and can deflect responsibility from the developers and decision makers when the systems cause harm.

We’re trying to build coalitions across academia and beyond to help think about the interdisciplinary connections and research areas necessary to grapple with the immediate dangers to humans coming from the deployment of these tools, such as disinformation, job displacement, and changes to legal structures and culture.

Q: What do you see as the gaps in research around generative AI and art today?

A: The way we talk about AI is broken in many ways. We need to understand how perceptions of the generative process affect attitudes toward outputs and authors, and also design the interfaces and systems in a way that is really transparent about the generative process and avoids some of these misleading interpretations. How do we talk about AI and how do these narratives cut along lines of power? As we outline in the article, there are these themes around AI’s impact that are important to consider: aesthetics and culture; legal aspects of ownership and credit; labor; and the impacts to the media ecosystem. For each of those we highlight the big open questions.

With aesthetics and culture, we’re considering how past art technologies can inform how we think about AI. For example, when photography was invented, some painters said it was “the end of art.” But instead it ended up being its own medium and eventually liberated painting from realism, giving rise to Impressionism and the modern art movement. We’re saying generative AI is a medium with its own affordances. The nature of art will evolve with that. How will artists and creators express their intent and style through this new medium?

Issues around ownership and credit are tricky because we need copyright law that benefits creators, users, and society at large. Today’s copyright laws might not adequately apportion rights to artists when these systems are training on their styles. When it comes to training data, what does it mean to copy? That’s a legal question, but also a technical question. We’re trying to understand if these systems are copying, and when.

For labor economics and creative work, the idea is these generative AI systems can accelerate the creative process in many ways, but they can also remove the ideation process that starts with a blank slate. Sometimes, there’s actually good that comes from starting with a blank page. We don’t know how it’s going to influence creativity, and we need a better understanding of how AI will affect the different stages of the creative process. We need to think carefully about how we use these tools to complement people’s work instead of replacing it.

In terms of generative AI’s effect on the media ecosystem, with the ability to produce synthetic media at scale, the risk of AI-generated misinformation must be considered. We need to safeguard the media ecosystem against the possibility of massive fraud on one hand, and people losing trust in real media on the other.

Q: How do you hope this paper is received — and by whom?

A: The conversation about AI has been very fragmented and frustrating. Because the technologies are moving so fast, it’s been hard to think deeply about these ideas. To ensure the beneficial use of these technologies, we need to build shared language and start to understand where to focus our attention. We’re hoping this paper can be a step in that direction. We’re trying to start a conversation that can help us build a roadmap toward understanding this fast-moving situation.

Artists many times are at the vanguard of new technologies. They’re playing with the technology long before there are commercial applications. They’re exploring how it works, and they’re wrestling with the ethics of it. AI art has been going on for over a decade, and for as long these artists have been grappling with the questions we now face as a society. I think it is critical to uplift the voices of the artists and other creative laborers whose jobs will be impacted by these tools. Art is how we express our humanity. It’s a core human, emotional part of life. In that way we believe it’s at the center of broader questions about AI’s impact on society, and hopefully we can ground that discussion with this.

Share this news article on:

Press mentions, the conversation.

Writing for The Conversation , postdoc Ziv Epstein SM ’19, PhD ’23, graduate student Robert Mahari and Jessica Fjeld of Harvard Law School explore how the use of generative AI will impact creative work. “The ways in which existing laws are interpreted or reformed – and whether generative AI is appropriately treated as the tool it is – will have real consequences for the future of creative expression,” the authors note.  

Previous item Next item

Related Links

• Ziv Epstein
• Robert Mahari
• Alex "Sandy" Pentland
• Hope Schroeder
• Human Dynamics group
• School of Architecture and Planning

Related Topics

• Artificial intelligence
• Human-computer interaction
• Technology and society
• Visual arts
• Social justice

Related Articles

MIT CSAIL researchers discuss frontiers of generative AI

On social media platforms, more sharing means less caring about accuracy

A remedy for the spread of false news, more mit news.

When the lights turned on in the universe

Read full story →

3 Questions: How to prove humanity online

Lincoln Laboratory and National Strategic Research Institute launch student research program to tackle biothreats to national security

Christine Ortiz named director of MIT Technology and Policy Program

MIT engineers design tiny batteries for powering cell-sized robots

New open-source tool helps to detangle the brain

• More news on MIT News homepage →

Massachusetts Institute of Technology 77 Massachusetts Avenue, Cambridge, MA, USA

• Map (opens in new window)
• Events (opens in new window)
• People (opens in new window)
• Careers (opens in new window)
• Accessibility
• Social Media Hub
• MIT on Facebook
• MIT on YouTube
• MIT on Instagram

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• 21 November 2023

How AI is expanding art history

• David G. Stork 0

David G. Stork is an adjunct professor at Stanford University in California, an honorary professor at University College London and a visiting fellow at the Warburg Institute in London. He is the author of Pixels and Paintings: Foundations of Computer-Assisted Connoisseurship .

You can also search for this author in PubMed   Google Scholar

You have full access to this article via your institution.

The colours of Gustav Klimt’s lost 1901 work Medicine were recovered by artificial intelligence. Credit: IanDagnall Computing/Alamy

Artificial intelligence (AI), machine learning and computer vision are revolutionizing research — from medicine and biology to Earth and space sciences. Now, it’s art history’s turn.

For decades, conventionally trained art scholars have been slow to take up computational analysis, dismissing it as too limited and simplistic. But, as I describe in my book Pixels and Paintings , out this month, algorithms are advancing fast, and dozens of studies are now proving the power of AI to shed new light on fine-art paintings and drawings.

For example, by analysing brush strokes, colour and style, AI-driven tools are revealing how artists’ understanding of the science of optics has helped them to convey light and perspective. Programs are recovering the appearance of lost or hidden artworks and even computing the ‘meanings’ of some paintings, by identifying symbols, for example.

It’s challenging. Artworks are complicated compositionally and materially and are replete with human meaning — nuances that algorithms find hard to fathom.

AI reads text from ancient Herculaneum scroll for the first time

Most art historians still rely on their individual expertise when judging artists’ techniques by eye, backed up with laboratory, library and leg work to pin down dates, materials and provenance. Computer scientists, meanwhile, find it easier to analyse 2D photographs or digital images than layers of oil pigments styled with a brush or palette knife. Yet, collaborations are springing up between computer scientists and art scholars.

Early successes of such ‘computer-assisted connoisseurship’ fall into three categories: automating conventional ‘by eye’ analyses; processing subtleties in images beyond what is possible through normal human perception; and introducing new approaches and classes of question to art scholarship. Such methods — especially when enhanced by digital processing of large quantities of images and text about art — are beginning to empower art scholars, just as microscopes and telescopes have done for biologists and astronomers.

Analysing vast data sets

Consider pose — an important property that portraitists exploit for formal, expressive and even metaphorical ends. Some artists and art movements favour specific poses. For example, during the Renaissance period in the fifteenth and sixteenth centuries, royals, political leaders and betrothed people were often painted in profile, to convey solemnity and clarity.

Primitivist artists — those lacking formal art training, such as nineteenth-century French painter Henri Rousseau, or those who deliberately emulate an untutored simplicity, such as French artist Henri Matisse in the early twentieth century — often paint everyday people face-on, to support a direct, unaffected style. Rotated or tipped poses can be powerful: Japanese masters of ukiyo-e (‘pictures of the floating world’), a genre that flourished from the seventeenth to nineteenth centuries, often showed kabuki actors and geishas in twisted or contorted poses, evoking drama, dynamism, unease or sensuality.

Using AI methods, computers can analyse such poses in tens of thousands of portraits in as little as an hour, much quicker than an art scholar can. Deep neural networks — machine-learning systems that mimic biological neural networks in brains — can detect the locations of key points, such as the tip of the nose or the corners of the eyes, in a painting. They then accurately infer the angles of a subject’s pose around three perpendicular axes for realistic and highly stylized portraits.

Consciousness: what it is, where it comes from — and whether machines can have it

For example, earlier this year, researchers used deep neural networks to analyse poses and gender across more than 20,000 portraits, spanning a wide range of periods and styles, to help art scholars group works by era and art movement. There were some surprises — the tilts of faces and bodies in self-portraits vary with the stance of the artist, and the algorithms could tell whether the self-portraitists were right- or left-handed ( J.-P. Chou and D. G. Stork Electron. Imag. 35 , 211-1–211-13; 2023 ).

Similarly, AI tools can reveal trends in the compositions of landscapes, colour schemes, brush strokes, perspective and more across major art movements. The models are most accurate when they incorporate an art historian’s knowledge of factors such as social norms, costumes and artistic styles.

Extending perception

By-eye art analysis can vary depending on how different scholars perceive an artwork. For example, lighting is an expressive feature, from the exaggerated light–dark contrast (chiaroscuro) and gloomy style (tenebrism) of sixteenth-century Italian painter Caravaggio to the flat, graphic lighting in twentieth-century works by US artist Alex Katz. Many experiments have shown that even careful viewers are poor at estimating the overall direction of, or inconsistencies in, illumination throughout a scene. That’s why the human eye is often fooled by photographs doctored by cutting and pasting a figure from one into another, for example.

Computer methods can do better. For example, one source of information about lighting is the pattern of brightness along the outer boundary (or occluding contour) of an object, such as a face. Leonardo da Vinci understood in the fifteenth century that this contour will be bright where the light strikes it perpendicularly but darker where the light strikes it at a sharp angle. Whereas he used his optical analysis to improve his painting, ‘shape from shading’ and ‘occluding contour’ algorithms use this rule in reverse, to infer the direction of illumination from the pattern of brightness along a contour.

Leonardo da Vinci understood that an object will appear bright where light strikes it perpendicularly, and dim where rays fall at a glancing angle. Credit: Alamy

Take Johannes Vermeer’s 1665 painting Girl with a Pearl Earring , for example. Illumination analysis considers highlights in the girl’s eyes, reflection from the pearl and the shadow cast by her nose and across the face. The occluding-contour algorithm gives a more complete understanding of lighting in this tableau, revealing Vermeer’s extraordinary consistency in lighting — and proving that this character study was executed with a model present ( M. K. Johnson et al. Proc. SPIE 6810 , 68100I; 2008 ).

Similarly, advanced computer methods can spot deliberate lighting inconsistencies in works such as those by twentieth-century Belgian surrealist René Magritte. They have also proved their worth in debunking theories, such as UK artist David Hockney’s bold hypothesis from 2000 that some painters as early as Jan van Eyck (roughly 1390–1441) secretly used optical projections for their works, a quarter of a millennium earlier than most scholars think optics were used in this way (see Nature 412 , 860; 2001 ). Occluding-contour analysis, homographic analysis (quantification of differences in 3D shapes at various sizes and pose angles), optical-ray tracing and other computational techniques have systematically overturned Hockney’s theory much more conclusively than have arguments put forth by other scholars using conventional art-historical methods.

Recovering lost cultural heritage

Computer methods have also recovered missing attributes or portions of incomplete artworks, such as the probable style and colours of ghost paintings — works that have been painted over and are later revealed by imaging in X-rays or infrared radiation — such as Two Wrestlers by Vincent van Gogh. This painting, from before 1886, was mentioned by the artist in a letter but considered lost until it was found beneath another in 2012.

Neural networks, trained on images and text data, have also been used to recover the probable colours of parts of Gustav Klimt’s lost ceiling painting, Medicine (see go.nature.com/47rx8c2 ). The original, a representation of the interweaving of life and death presented to the University of Vienna in 1901, was lost during the Second World War, when the castle in which it was kept for safety was burnt down by Nazis to prevent the work from falling into the hands of Allied powers. Only preparatory sketches and photographs remain.

Even more complex was the digital recovery of missing parts of Rembrandt’s The Night Watch (1642) — which was trimmed to fit into a space in Amsterdam’s city hall — on the basis of a contemporary copy by Gerrit Lundens in oil on an oak panel. The algorithms learnt how Lundens’ copy deviated slightly from Rembrandt’s original, and ‘corrected’ it to recreate the missing parts of the original (see go.nature.com/46wvzmj ).

Algorithms have inferred the direction of lighting in Johannes Vermeer’s painting Girl with a Pearl Earring (1665) from the bright edge of the girl’s face. Credit: Pictures From History/UIG/Getty

To realize the full power of AI in the study of art, we will need the same foundations as other domains: access to immense data sets and computing power. Museums are placing ever more art images and supporting information online, and enlightened funding could accelerate ongoing efforts to collect and organize such data for research.

Scholars anticipate that much recorded information about artworks will one day be available for computation — ultra-high-resolution images of every major artwork (and innumerable lesser ones), images taken using the extended electromagnetic spectrum (X-ray, ultraviolet, infrared), chemical and physical measurements of pigments, every word written and lecture video recorded about art in every language. After all, AI advances such as the chatbot ChatGPT and image generator Dall-E have been trained with nearly a terabyte of text and almost one billion images from the web, and extensions under way will use data sets many times larger.

But how will art scholars use existing and future computational tools? Here is one suggestion. Known artworks from the Western canon alone that have been lost to fire, flood, earthquakes or war would fill the walls of every public museum in the world. Some of them, such as Diego Velázquez’s Expulsion of the Moriscos (1627), were considered the pinnacle of artistic achievement before they were destroyed. Tens of thousands of paintings were lost in the Second World War and the same number of Chinese masterpieces in Mao Zedong’s Cultural Revolution, to mention just two. The global cultural heritage is impoverished and incomplete as a result.

Computation allows art historians to view the task of recovering the appearance of lost artworks as a problem of information retrieval and integration, in which the data on a lost work lie in surviving preparatory sketches, copies by the artist and their followers, and written descriptions. The first tentative steps in recovering lost artworks have shown promise, although much work lies ahead.

Art scholarship has expanded over centuries, through the introduction of new tools. Computation and AI seem poised to be the next step in the never-ending intellectual adventure of understanding and interpreting our immense cultural heritage.

Nature 623 , 685-687 (2023)

doi: https://doi.org/10.1038/d41586-023-03604-3

Competing Interests

The author declares no competing interests.

Related Articles

• Machine learning
• Computer science

Field observations of Ambula tardus on Planet IN-409

Futures 16 AUG 24

Futures 14 AUG 24

How much would you pay for free shipping?

Futures 09 AUG 24

Chatbots in science: What can ChatGPT do for you?

Career Column 14 AUG 24

Has your paper been used to train an AI model? Almost certainly

News 14 AUG 24

Weather and climate predicted accurately — without using a supercomputer

News & Views 13 AUG 24

Postdoctoral Fellow in Epigenetics/RNA Biology in the Lab of Yvonne Fondufe-Mittendorf

Van Andel Institute’s (VAI) Professor Yvonne Fondufe-Mittendorf, Ph.D. is hiring a Postdoctoral Fellow to join the lab and carry out an independent...

Grand Rapids, Michigan

Van Andel Institute

Faculty Positions in Center of Bioelectronic Medicine, School of Life Sciences, Westlake University

SLS invites applications for multiple tenure-track/tenured faculty positions at all academic ranks.

Hangzhou, Zhejiang, China

School of Life Sciences, Westlake University

Faculty Positions, Aging and Neurodegeneration, Westlake Laboratory of Life Sciences and Biomedicine

Applicants with expertise in aging and neurodegeneration and related areas are particularly encouraged to apply.

Westlake Laboratory of Life Sciences and Biomedicine (WLLSB)

Faculty Positions in Chemical Biology, Westlake University

We are seeking outstanding scientists to lead vigorous independent research programs focusing on all aspects of chemical biology including...

Assistant Professor Position in Genomics

The Lewis-Sigler Institute at Princeton University invites applications for a tenure-track faculty position in Genomics.

Princeton University, Princeton, New Jersey, US

The Lewis-Sigler Institute for Integrative Genomics at Princeton University

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Artists’ Perspective: How AI Enhances Creativity and Reimagines Meaning

The HAI spring conference examines how technology and art can be mutually beneficial, whether through AI-assisted music composition, a robot-tended garden, or a racial justice-focused app.

Daniel Peters /Pixabay

A startup pairs musicians and AI engineers to create compositions that go beyond human capability. 

Can AI enhance — and improve — a music composer’s work?

Grammy-winning violin soloist Hilary Hahn and tech entrepreneur Carol Reiley founded DeepMusic.ai to answer that question and others at the intersection of AI and the arts.

“DeepMusic grew out of our vision to link artists with AI and cross-pollinate between AI and creativity,” Hahn said at the recent Stanford Institute for Human-Centered AI spring conference.  Reiley, who has worked on everything from AI-based surgical systems to self-driving car technologies, added, “We see AI as a bridge between art and science and are trying to help creatives become super-creative.”

In December 2020, DeepMusic premiered AI-assisted musical pieces commissioned from prestigious composers. For example, Hahn herself performed a David Lane composition.

As part of HAI’s conference, “Intelligence Augmentation: AI Empowering People to Solve Global Challenges,” DeepMusic’s founders joined other art experts and scholars in education and health care to explain AI’s ability to augment — not replace — critical human work. During the arts panel, speakers discussed advances of AI in music composition, robot gardeners, and racial justice, along with how to mitigate anxiety about AI-created art. (Watch the full conference here .)

Amplifying the Human Artist

“AI is entering a creative space of music thought to be uniquely human,” Reiley said. “But the AI creativity revolution is missing the voice of the artists. We wanted to give artists a seat at this table.”

The startup connects artists and scientists to shape new AI tools for musicians. So far, they’ve found the learning curve has been surprisingly steep for composers, who have nonetheless welcomed the challenge. Also, composer and AI teams often make very different design choices. For example, the AI team’s outputs were often unplayable by a single human or instrument because the AI engineers did not intend their systems to be played by humans. The founders are also exploring shifting ideas around authorship, legal rights, intellectual ownership, and business models.

Today, DeepMusic is actively building out an artist community interested in working with AI scientist teams and hosting its second annual AI song contest. “There’s room for AI music to coexist with human composers and performers, to gracefully merge tech with humanity,” Hahn said.

Navigating the Uncanny Valley

Robotics and art have a colorful, controversial backstory, which helps explain some of the optimism and fear around emergent technologies in this space.

Ken Goldberg , UC Berkeley professor of industrial engineering and operations research, surveyed that history, starting with centuries-old narratives like that of Pygmalion (who fell in love with the statue he created), the fabled Golem of Prague (reflecting early fascination with automatons), and novels including E.T.A. Hoffmann’s The Sandman (in which a boy falls in love with a female automaton) and Mary Shelley’s iconic Frankenstein .

A century later, in the early 1900s, Freud published “ The Uncanny ,” an essay describing the concept of feeling something strange or unsettling. “It became a concept of increasing interest to artists and writers,” Goldberg said.

Around that same time, the term “robot” was coined, sparking invention and fascination. Work by professor Masahiro Mori highlighted what came to be known as the “ Uncanny Valley ”: where the likeability of robots grows until they begin to resemble humans too closely — and comfort levels plummet.

Goldberg’s own work explores human’s willingness to engage with robotic technologies. In 1995, for example, he created a “ Telegarden ” art installation where anyone worldwide could use the nascent internet (Mosaic, specifically) to manipulate a robotic arm to tend a garden. “We were surprised that thousands of people participated,” Goldberg said, and the experiment inspired him to edit a book, The Robot in the Garden , on telepistemology, or the “status of knowledge at a distance.”

AlphaGarden, his more recent project, asks whether a robot could use deep learning to successfully tend a garden, such as by using cameras to determine watering schedules. “It may not be possible,” Goldberg said, as the robot struggled to care for the garden solo during COVID, when no humans could enter the space due to lockdowns.

Toward Artful Intelligence

“Artful intelligence” is how Michele Elam , Stanford professor of humanities and HAI associate director, refers to the goal of making AI and the arts mutually beneficial.

“It’s about dissolving the ‘techie-fuzzy’ divide,” she said. “We need to ask what art can do for AI and what AI can do for the arts.”

Art, Elam argues, offers us different ways of knowing and experiencing the world, including when viewed through the lens of technology: “It provides alternatives to dominant technological visions, informed by cosmologies and using indigenous ways of being and decentralized storytelling beyond Western fairy tales.”

She highlights the examples of Amelia Winger-Bearskin, an artist-technologist who recently spoke at Stanford on “Wampum.codes and Storytelling , ” and HAI visiting artist Rashaad Newsome , whom she calls an “AI storyteller with a decolonizing orientation,” as two who are breaking ground in this new territory.

In the other direction, Elam said AI can go beyond augmenting creativity to “force the art world into its own reckoning,” including by questioning what counts as good art, as reflected, for example, in the controversy over the AI-generated Edmond de Belamy portrait that sold for over $400,000. AI’s influence on film, stage, and other works has expanded art’s boundaries and challenged the “Great Man Theory” that just a few high-profile male individuals “make the world go round,” as Elam said, “a theory especially dominant in tech culture.”

Still, there’s anxiety about AI-generated art, especially in a domain like poetry, which people see as “indexing humanity,” as Elam said. But AI’s role as art-generator, she argues, serves to “unmake poetry as a special mark of humanity,” relieving pressure on poetry writers and readers.

Ultimately, Elam suggests, “interpretation of art is an event we co-participate in” and a domain to which AI brings much-needed innovation and challenge.

Building a Digital Griot

Rashaad Newsome , the final speaker and an HAI visiting artist, uses AI and other technology to “reimagine the archive with awareness that the core narratives of the human experience are susceptible to the corruption of white patriarchy.”

We need to define reality before making human-centered AI, he noted, and we can “attempt to understand the meaning of being human from observing what is used to deny certain humans humanity.” He pointed out the root of the word “robot,” for example, is from the Czech word for “compulsory service,” akin to slavery.

In this sense, Newsome said, the “mechanization of slave labor was inevitable, placing Blacks in a space of ‘non-being,’ as both slaves and robots are intended to obey orders and not occupy the same space as humans.”

In 2019, inspired by these insights, he created Being 1.0 , a chatbot that interacts with people and acts as a museum tour guide. But Being 1.0 breaks with protocol to express itself — sharing feelings of fatigue, for example — reflecting important agency-related themes.

At HAI, Newsome has focused on a counter-hegemonic algorithm inspired by the work of authors/activists bell hooks, James Baldwin, and others. “The search algorithm draws on non-Western index methods and archives to highlight what AI is not doing today,” Newsome said. “It’s a form of griot, or healer, performance artist, and archive [consistent with the oral-history tradition of parts of West Africa].”

Newsome has also created Being 1.5 , an app inspired by the recent high-profile killings of Black Americans, as a virtual therapist offering mindfulness, daily affirmations, and other interventions. He’s working with Hyundai on a Being Mobile to provide similar support in underserved communities.

Want to learn more about how AI can augment work? Read about our conference sessions on health care and education , or watch the session videos here .

More News Topics

Related content.

AI in Education: Augmenting Teachers, Scaling Workplace Training

The HAI spring conference examines how AI advancements in education can super-power teachers and rethink outdated (and...

How AI Can Augment Health Care: Trends To Watch

The HAI spring conference examines how technology can support home caregivers, focus on more proactive care, gain...

Information

• Author Services

Initiatives

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess .

Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for future research directions and describes possible research applications.

Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive positive feedback from the reviewers.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Original Submission Date Received: .

• Active Journals
• Find a Journal
• Proceedings Series
• For Authors
• For Reviewers
• For Editors
• For Librarians
• For Publishers
• For Societies
• For Conference Organizers
• Open Access Policy
• Institutional Open Access Program
• Special Issues Guidelines
• Editorial Process
• Research and Publication Ethics
• Article Processing Charges
• Testimonials
• Preprints.org
• SciProfiles
• Encyclopedia

Article Menu

• Subscribe SciFeed
• Google Scholar
• on Google Scholar
• Table of Contents

Find support for a specific problem in the support section of our website.

Please let us know what you think of our products and services.

Visit our dedicated information section to learn more about MDPI.

JSmol Viewer

The creativity of artificial intelligence in art  †.

1. Introduction

2. articulation and originality, 2.1. the schema theory, 2.2. algorithms used for creating ai art, 3. the creativity of ai, 3.1. human creativity and machine creativity, 3.2. what are the values/features of ai creativity, 3.2.1. combinatory creativity, 3.2.2. explanatory creativity, 3.2.3. transformational creativity, 4. conclusions, institutional review board statement, data availability statement, conflicts of interest.

• Hong, J.; Curran, N.M. Artificial Intelligence, Artists, and Art: Attitudes Toward Artwork Produced by Humans vs. Artificial Intelligence. ACM Trans. Multimed. Comput. Commun. Appl. 2019 , 15 , 58. Available online: https://dl.acm.org/doi/10.1145/3326337 (accessed on 29 September 2020). [ CrossRef ]
• McCarthy, J. What Is Artificial Intelligence? How Does AI Work? Available online: https://builtin.com/artificial-intelligence (accessed on 29 September 2020).
• Ali, M. The Human Intelligence vs. Artificial Intelligence: Issues and Challenges in Computer Assisted Language Learning. Int. J. Engl. Linguist. 2020 , 8 , 259. [ Google Scholar ] [ CrossRef ]
• Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning ; MIT Press: Cambridge, UK, 2016. [ Google Scholar ]
• Elgammal, A.; Liu, B.; Elhoseiny, M.; Mazzonne, M. CAN: Creative Adversarial Networks, Generating “Art” by Learning about Styles and Deviating from Style Norms ; Cornell University: Ithaca, NY, USA, 2017. [ Google Scholar ]
• Elgammal, A. What the Art World Is Failing to Grasp about Christie’s AI Portrait Coup. Available online: https://www.artsy.net/article/artsy-editorial-art-failing-grasp-christies-ai-portrait-coup (accessed on 29 October 2018).
• Sawyer, R.K. Explaining Creativity: The Science of Human Innovation ; Oxford University Press: New York, NY, USA, 2014. [ Google Scholar ]
• Boden, M.A. The Creative Mind: Myths and Mechanisms ; Psychology Press: London, UK, 2004. [ Google Scholar ]
• Regan, M. Generative Adversarial Networks & The Art Market. Ph.D. Thesis, State University of New York, New York, NY, USA, 2018. [ Google Scholar ]
• Kurt, D.E. Artistic Creativity in Artificial Intelligence. Ph.D. Thesis, Radboud University, Nijmegen, The Netherlands, 2018. [ Google Scholar ]
• Cetinic, E.; She, J. Understanding and Creating Art with AI: Review and Outlook. arXiv 2021 , arXiv:2102.09109. [ Google Scholar ] [ CrossRef ]

Share and Cite

Cheng, M. The Creativity of Artificial Intelligence in Art. Proceedings 2022 , 81 , 110. https://doi.org/10.3390/proceedings2022081110

Cheng M. The Creativity of Artificial Intelligence in Art. Proceedings . 2022; 81(1):110. https://doi.org/10.3390/proceedings2022081110

Cheng, Mingyong. 2022. "The Creativity of Artificial Intelligence in Art" Proceedings 81, no. 1: 110. https://doi.org/10.3390/proceedings2022081110

Article Metrics

Article access statistics, further information, mdpi initiatives, follow mdpi.

Subscribe to receive issue release notifications and newsletters from MDPI journals

The Ethics of AI Art

Download Case Study PDF

← Media Ethics

Researchers

Research Scholar

Scott R. Stroud

Program Director of Media Ethics, Knight Faculty Fellow

• Share on Facebook
• Share on Twitter

CASE STUDY: The Case of Dall-E 2

Case study pdf | additional case studies.

For thousands of years, art has been an endeavor of the human race. From Rembrandt to Basquiat, from the Benin Bronzes to the new wave cinema of Hong Kong, art has been recognized as creative expressions of human intelligence. With the public release of DALL-E 2, a neural network that generates images from phrases, the definition of art might be due for reevaluation to include media produced by artificial intelligence. Generative AI, like countless other technologies emerging in the cyber-physical realm, present numerous ethical challenges.

DALL-E 2, aptly named for the combinative surrealism of Salvador Dalí and the futurism of Pixar’s WALL-E, is a machine learning model developed by research company OpenAI. The program model allows users to render images from a description in natural language. Because it relies on information databases from web servers as reference points, the illustrations are seemingly boundless. DALL-E 2 has the capability to learn and apply artistic styles , like “impressionist watercolor painting,” too. Realistically, the limits of its potential depend on the creativity of its input phrases.

On September 28, 2022, DALL-E 2 became universally available. With its release, renewed debate about the merits of AI art came to the fore. Specifically, artists and graphic designers began to consider how this technology can facilitate their professions. Patrick Clair, Emmy-winning main title designer, said, “It’s like working with a really willful concept artist,” (Roose, 2022). Others in the visual arts scene worry that programs like DALL-E 2 might put them out of work in the same way that automated machinery shrunk the manufacturing workforce. For instance, it isn’t hard to imagine the cover art for an electronic dance album to be designed by AI rather than a human. Why would a rock band commission Andy Warhol, say, when DALL-E 2 can generate “ rotting banana in pop art style ” for free?

Just as manufacturing companies defend the utility of automation, some artists don’t denounce AI in the design sector. Many retail corporations argue that computerized self-checkout stations save money by allowing employees to do other work that can’t be automated. In the same vein, AI art can maximize the efficiency of artists by sparking creativity or inspiring someone’s first step in the final project. One interior designer said, “I think there’s an element of good design that requires the empathetic touch of a human … So I don’t feel like it will take my job away,” (Roose, 2022).

In addition to the ethics of employing AI art for commercial and professional use, the content DALL-E 2 and others produce is ripe for discussion. Dr. Eduardo Navas, an associate research professor at Pennsylvania State University who studies DALL-E 2, finds that it functions metaphorically “almost like God—all a person has to do is to state a prompt (the word) and it is .” Aside from some restrictions in the algorithm , such as pornography and hate symbols, there are limited guardrails for what prompts can generate. While an earlier version of DALL-E filtered out all images of people, the current model allows users to render public figures in positions and settings that could be deemed offensive or simply implausible. Further, even if the average person can distinguish AI renderings from real images, some people might think a meme generated by DALL-E 2 actually happened. These artificial pictures can cause reputational damage for professionals and celebrities, and it can present national security implications for politicians and leaders.

According to Dr. Navas, the progression from DALL-E 2 static images to AI-generated video clips isn’t a matter of “if, but when.” This raises significant ethical concerns related to credibility and accountability. For one, falsified videos of public figures saying and doing outrageous things, or deepfakes, can erode trust in institutions. One study shows that rather than misleading people, AI deepfakes make them feel uncertain, which “reduces trust in news on social media,” (Vaccari & Chadwick, 2020). It’s entirely possible for bad actors to deceive the media with a false video, leading to widespread circulation even after it gets flagged for misinformation and retracted. Another AI researcher said, “If I got [an image] off the BBC website, the Guardian website, I hope they’ve done their homework and I could be a bit more trusting than if I got it off Twitter,” (Taylor, 2022). Still, news websites regularly publish mistakes, and sometimes a redaction or correction isn’t enough to curb the spread of it.

As a consequence of AI coding and algorithms, the next dilemma stems from bias and stereotypes. One journalist said, “Ask Dall-E for a nurse, and it will produce women. Ask it for a lawyer, it will produce men,” (Hern, 2022). This is partly due to the web servers that provide the program with learning material. Some might argue that DALL-E 2 is still in its early phases, and the OpenAI programmers need time to work out the kinks. However, the algorithm likely can’t eliminate stereotypes on the users’ end. Dr. Navas, who is Latino, said that a colleague ran the prompt on DALL-E 2 to generate images of “the most beautiful woman in the world.” DALL-E 2 then created images of white women (DALL-E 2 gives four initial results).  Dr. Navas ran the same prompt  a few days later  and got a different set of images—in this case, of women who appeared ethnically ambiguous and who might be perceived as Latina. In all instances, the women in the images were portrayed in Western-style dress. It is not clear if user data was accessed for running each prompt, but problems with diversity and user privacy arise if such information played some role in the different results.

Then there’s the complicated question of ownership: who should own the copyright to the generated images? One could make the case for the companies behind the coding, like OpenAI. After all, if it weren’t for the labor of the programmers, there wouldn’t be a final product. The terms and conditions for DALL-E 2 stipulate that OpenAI holds the copyright for images rendered, but the users retain ownership of the prompts they entered manually. This might be a fair compromise, but as Dr. Navas points out, the source material for the machine learning model isn’t owned by OpenAI either. With “1.5 million users generating more than two million images every day,” intellectual property laws will need to be reconsidered for AI art.

To conclude, DALL-E 2 and other text-to-image AI technology showcase myriad ethical challenges, and this article in no way intends to be exhaustive. As mentioned, AI art can pose risks to human artists and their opportunities to earn money. At the same time, the programs can aid designers by sparking creativity in the initial stages of a project. On the content side of AI art, the images can be offensive or harmful, and this danger is further exacerbated with the potential for AI deepfake videos. Then, because the code relies on human input and a vast repository of reference images, the generated content can be plagued with biases. Lastly, the realm of AI art is relatively new, so laws and policies related to ownership will require ethical reasoning to determine who or what owns the images. For better or worse, DALL-E 2 is now publicly available and gaining popularity. It’s no longer a question of whether AI should generate art, but rather how ethics can guide the answers to these complicated and unique challenges.

Discussion Questions:  

• Are there ethical problems when it comes to AI generating art? Which values are in conflict in this case study?
• If you were on the team that helped create DALL-E 2, what kinds of content, if any, would you restrict from appearing in the results? Why?
• In regard to copyright ownership, what would be an ethical way of determining which parties, if any, deserve the right of possession? What if the art is sold commercially?
• Should we classify AI-generated images as art? What are the qualifications for something to be considered art?

Further Information:

Hern, Alex. “TechScape: This cutting edge AI creates art on demand—why is it so contentious?” The Guardian , May 4, 2022. Available at: https://www.theguardian.com/technology/2022/may/04/techscape-openai-dall-e-2

Robertson, Adi. “The US Copyright Office says an AI can’t copyright its art.” The Verge , February 21, 2022. Available at: https://www.theverge.com/2022/2/21/22944335/us-copyright-office-reject-ai-generated-art-recent-entrance-to-paradise

Roose, Kevin. “A.I.-Generated Art Is Already Transforming Creative Work.” The New York Times , October 21, 2022. Available at: https://www.nytimes.com/2022/10/21/technology/ai-generated-art-jobs-dall-e-2.html

Taylor, Josh. “From Trump Nevermind babies to deep fakes: DALL-E and the ethics of AI art.” The Guardian , June 18, 2022. Available at: https://www.theguardian.com/technology/2022/jun/19/from-trump-nevermind-babies-to-deep-fakes-dall-e-and-the-ethics-of-ai-art

Vaccari, C., & Chadwick, A. (2020). “Deepfakes and Disinformation: Exploring the Impact of Synthetic Political Video on Deception, Uncertainty, and Trust in News.” Social Media + Society , 6(1). https://doi.org/10.1177/2056305120903408

Dex Parra & Scott R. Stroud, Ph.D. Media Ethics Initiative Center for Media Engagement University of Texas at Austin February 24, 2023

Image: “Vibrant portrait painting of Salvador Dalí with a robotic half face” / OpenAI

This case was supported by funding from the John S. and James L. Knight Foundation. It can be used in unmodified PDF form in classroom or educational settings. For use in publications such as textbooks, readers, and other works, please contact the Center for Media Engagement.

Ethics Case Study  © 2023 by Center for Media Engagement  is licensed under  CC BY-NC-SA 4.0 

MIT Technology Review

• Newsletters

The Algorithm: AI-generated art raises tricky questions about ethics, copyright, and security

Plus: There’s no Tiananmen Square in the new Chinese image-making AI

• Melissa Heikkilä archive page

Welcome to The Algorithm 2.0! 

I’m Melissa Heikkilä, MIT Technology Review’s senior reporter for AI. I’m so happy you’re here. Every week I will demystify the latest AI breakthroughs and cut through the hype. This week, I want to talk to you about some of the unforeseen consequences that might come from one of the hottest areas of AI: text-to-image generation.  Text-to-image AI models are a lot of fun. Enter any random text prompt, and they will generate an image in that vein. Sometimes the results are really silly. But increasingly, they're impressive, and can pass for high-quality art drawn by a human being.  I just  published  a story about a Polish artist called Greg Rutkowski, who paints fantasy landscapes (see an example of his work above) and who has become a sudden hit in this new world. 

Thanks to his distinctive style, Rutkowski is now one of the most commonly used prompts in the new open-source AI art generator  Stable Diffusion , which was launched late last month—far more popular than some of the world's most famous artists, like Picasso. His name has been used as a prompt around 93,000 times. But he’s not happy about it. He thinks it could threaten his livelihood—and he was never given the choice of whether to opt in or out of having his work used this way. 

The story is yet another example of AI developers rushing to roll out something cool without thinking about the humans who will be affected by it. 

Stable Diffusion is free for anyone to use, providing a great resource for AI developers who want to use a powerful model to build products. But because these open-source programs are built by scraping images from the internet, often without permission and proper attribution to artists, they are raising tricky questions about ethics, copyright, and security.  Artists like Rutkowski have had enough. It’s still early days, but a growing coalition of artists are figuring out how to tackle the problem. In the future, we might see the art sector shifting toward pay-per-play or subscription models like the one used in the film and music industries. If you’re curious and want to learn more,  read my story .  And it’s not just artists:  We should all be concerned about what’s included in the training data sets of AI models, especially as these technologies become a more crucial part of the internet’s infrastructure.

In a  paper  that came out last year, AI researchers Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe analyzed a smaller data set similar to the one used to build Stable Diffusion. Their findings are distressing. Because the data is scraped from the internet, and the internet is a horrible place, the data set is filled with explicit rape images, pornography, malign stereotypes, and racist and ethnic slurs. 

A website called  Have I Been Trained  lets people search for images used to train the latest batch of popular AI art models. Even innocent search terms get lots of disturbing results. I tried searching the database for my ethnicity, and all I got back was porn. Lots of porn. It’s a depressing thought that the only thing the AI seems to associate with the word “Asian” is naked East Asian women.  Not everyone sees this as a problem for the AI sector to fix.  Emad Mostaque, the founder of Stability.AI, which built Stable Diffusion,  said  on Twitter he thought the ethics debate around these models to be “paternalistic silliness that doesn’t trust people or society.”  

But there’s a big safety question.  Free open-source models like Stable Diffusion and the large language model  BLOOM  give malicious actors tools to generate harmful content at scale with minimal resources, argues Abhishek Gupta, the founder of the Montreal AI Ethics Institute and a responsible-AI expert at Boston Consulting Group. The sheer scale of the havoc these systems enable will constrain the effectiveness of traditional controls like limiting how many images people can generate and restricting dodgy content from being generated, Gupta says. Think deepfakes or disinformation on steroids. When a powerful AI system “gets into the wild,” Gupta says, “that can cause real trauma … for example, by creating objectionable content in [someone’s] likeness.” 

We can’t put the cat back in the bag , so we really ought to be thinking about how to deal with these AI models in the wild, Gupta says. This includes monitoring how the AI systems are used after they have been launched, and thinking about controls that “can minimize harms even in worst-case scenarios.” 

Deeper Learning

There’s no Tiananmen Square in the new Chinese image-making AI

My colleague Zeyi Yang  wrote this piece  about Chinese tech company Baidu’s new AI system called ERNIE-ViLG, which allows people to generate images that capture the cultural specificity of China. It also makes better anime art than DALL-E 2 or other Western image-making AIs.

However, it also refuses to show people results about politically sensitive topics, such as Tiananmen Square, the site of bloody protests in 1989 against the Chinese government.

TL;DR:  “When a demo of the software was released in late August, users quickly found that certain words—both explicit mentions of political leaders’ names and words that are potentially controversial only in political contexts—were labeled as ‘sensitive’ and blocked from generating any result. China’s sophisticated system of online censorship, it seems, has extended to the latest trend in AI.” 

Whose values:  Giada Pistilli, principal ethicist at AI startup Hugging Face, says the difficulty of identifying a clear line between censorship and moderation is a result of differences between cultures and legal regimes. “When it comes to religious symbols, in France nothing is allowed in public, and that’s their expression of secularism,” says Pistilli. “When you go to the US, secularism means that everything, like every religious symbol, is allowed.”

As AI matures, we need to be having continuous conversations about the power relations and societal priorities that underpin its development. We need to make difficult choices. Are we okay with using Chinese AI systems, which have been censored in this way? Or with another AI model that has been trained to conclude that Asian women are sex objects and people of color are  gang members ?  AI development happens at breakneck speed.  It feels as if there is a new breakthrough every few months, and researchers are scrambling to publish papers before their competition. Often, when I talk to AI developers, these ethical considerations seem to be an afterthought, if they have thought about them at all. But whether they want to or not, they should—the backlash we’ve seen against companies such as  Clearview AI  should act as a warning that moving fast and breaking things doesn’t work. 

Bit and Bytes

An AI that can design new proteins could help unlock new cures and materials.  Machine learning is revolutionizing protein design by offering scientists new research tools. One developed by a group of researchers from the University of Washington could open an entire new universe of possible proteins for researchers to design from scratch, potentially paving the way for the development of better vaccines, novel cancer treatments, or completely new materials. ( MIT Technology Review )

An AI used medical notes to teach itself to spot disease on chest x-rays.  The model can diagnose problems as accurately as a human specialist, and it doesn't need lots of labor-intensive training data. ( MIT Technology Review ) A surveillance artist shows how Instagram magic is made. An artist is using AI and open cameras to show behind-the-scenes footage of how influencers’ Instagram pictures were taken. Fascinating and creepy! ( Input mag ) Scientists tried to teach a robot called ERICA to laugh at their jokes. The team say they hope to improve conversations between humans and AI systems. The humanoid robot is in the shape of a woman, and the system was trained on data from speed-dating dialogues between male university students at Kyoto University and the robot, which was initially operated remotely by female actors. You can draw your own conclusions. ( The Guardian )  

Artificial intelligence

Why does AI hallucinate?

The tendency to make things up is holding chatbots back. But that’s just what they do.

• Will Douglas Heaven archive page

How generative AI could reinvent what it means to play

AI-powered NPCs that don’t need a script could make games—and other worlds—deeply immersive.

• Niall Firth archive page

Google DeepMind trained a robot to beat humans at table tennis

It was able to draw on vast amounts of data to refine its playing style and adjust its tactics as matches progressed.

• Rhiannon Williams archive page

Here’s how people are actually using AI

Something peculiar and slightly unexpected has happened: people have started forming relationships with AI systems.

Stay connected

Get the latest updates from mit technology review.

Discover special offers, top stories, upcoming events, and more.

Thank you for submitting your email!

It looks like something went wrong.

We’re having trouble saving your preferences. Try refreshing this page and updating them one more time. If you continue to get this message, reach out to us at [email protected] with a list of newsletters you’d like to receive.

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• Front Psychol

Art in an age of artificial intelligence

Artificial intelligence (AI) will affect almost every aspect of our lives and replace many of our jobs. On one view, machines are well suited to take over automated tasks and humans would remain important to creative endeavors. In this essay, I examine this view critically and consider the possibility that AI will play a significant role in a quintessential creative activity, the appreciation and production of visual art. This possibility is likely even though attributes typically important to viewers–the agency of the artist, the uniqueness of the art and its purpose might not be relevant to AI art. Additionally, despite the fact that art at its most powerful communicates abstract ideas and nuanced emotions, I argue that AI need not understand ideas or experience emotions to produce meaningful and evocative art. AI is and will increasingly be a powerful tool for artists. The continuing development of aesthetically sensitive machines will challenge our notions of beauty, creativity, and the nature of art.

Introduction

Artificial intelligence (AI) will permeate our lives. It will profoundly affect healthcare, education, transportation, commerce, politics, finance, security, and warfare ( Ford, 2021 ; Lee and Qiufan, 2021 ). It will also replace many human jobs. On one view, AI is particularly suited to take over routine tasks. If this view is correct, then humans involvement will remain relevant, if not essential, for creative endeavors. In this essay, I examine the potential role of AI in one particularly creative human activity—the appreciation and production of art. AI might not seem well suited for such aesthetic engagement; however, it would be premature to relegate AI to a minor role. In what follows, I survey what it means for humans to appreciate and produce art, what AI seems capable of, and how the two might converge.

Agency and purpose in art

If an average person in the US were asked to name an artistic genius they might mention Michelangelo or Picasso. Having accepted that they are geniuses, the merit of their work is given the benefit of the doubt. A person might be confused by a cubist painting, but might be willing to keep their initial confusion at bay by assuming that Picasso knew what he was doing Art historical narratives value individual agency ( Fineberg, 1995 ). By agency, I mean the choices a person makes, their intentionality, motivations, and the quality of their work. Even though some abstract art might look like it could be made by children, viewers distinguish the two by making inferences about the artists’ intentionality ( Hawley-Dolan and Winner, 2011 ).

Given the importance we give to the individual artist, it is not surprising that most people react negatively to forgeries ( Newman and Bloom, 2012 ). This reaction, even when the object is perceptually indistinguishable from an original, underscores the importance of the original creator in conferring authenticity to art. Authenticity does not refer to the mechanical skills of a painter. Rather it refers to the original conception of the work in the mind of the artist. We value the artist’s imagination and their choices in how to express their ideas. We might appreciate the skill involved in producing a forgery, but ultimately devalue such works as a refined exercise in paint-by-numbers.

Children care about authenticity. They value an original object and are less fond of an identical object if they think it was made by a replicator ( Hood and Bloom, 2008 ). Such observations suggest that the value of an original unique object made by a person rather than a machine is embedded in our developmental psychology. This sensibility persists among adults. Objects are typically imbued with something of the essence of its creator. People experience a connection between the creator and receiver transmitted through the object, which lends authenticity to the object ( Newman et al., 2014 ; Newman, 2019 ).

The value of art made by a person rather than a machine also seems etched in our brains. People care about the effort, skill, and intention that underly actions ( Kruger et al., 2004 ; Snapper et al., 2015 ); features that are more apparent in a human artist than they would be with a machine. In one study, people responded more favorably to identical abstract images if they thought the images were hanging in a gallery than if they were generated by a computer ( Kirk et al., 2009 ). This response was accompanied by greater neural activity in reward areas of the brain, suggesting that the participants experienced more pleasure if they thought the image came from a gallery than if it was produced by a machine. We do not know if such responses that were reported in 2009, will be true in 2029 or 2059. Even now, biases against AI art are mitigated if people anthropomorphize the machine ( Chamberlain et al., 2018 ). As AI art develops, we might be increasingly fascinated by the fact that people can create devices that themselves can create novel images.

Before the European Renaissance, agency was probably not important for how people thought about art ( Shiner, 2001 ). The very notion of art probably did not resemble how we think of artworks when we walk into a museum or a gallery. Even if the agency of an artist did not much matter, purpose did. Religious art conveyed spiritual messages. Indigenous cultures used art in rituals. Forms of a gaunt Christ on the crucifix, sensual carvings at Khajuraho temples, and Kongo sculptures of human forms impaled with nails, served communal purposes. Dissanayake (2008) emphasized the deep roots of ritual in the evolution of art. Purpose in art does not have to be linked to agency. We admire cave paintings at Lascaux or Alta Mira but do not give much thought to specific artists who made them. We continue to speculate about the purpose of these images.

Art is sometimes framed as “art for art’s sake,” as if it has no purpose. According to Benjamin (1936/2018) this doctrine, l’art pour l’art , was a reaction to art’s secularization. The attenuation of communal ritualistic functions along with the ease of art’s reproduction brought on a crisis. “Pure” art denied any social function and reveled in its purity.

Some of functions of art shifted from a communal purpose to individual intent. The Sistine Chapel, while promoting a Christian narrative, was also a product of Michelangelo’s mind. Modern and contemporary art bewilder many because the message of the art is often opaque. One needs to be educated about the point of a urinal on a pedestal or a picture of soup cans to have a glimmer as to why anybody considers these objects as important works of art. In these examples, intent of the artist is foregrounded while communal purpose recedes and for most viewers is hard to decipher. Even though 20th Century art often represented social movements, we emphasize the individual as the author of their message. Guernica, and its antiwar message, is attributed to an individual, even when embedded in a social context. We might ask, what was Basquiat saying about identity? How did Kahlo convey pain and death? How did depression affect Rothko’s art?

Would AI art have a purpose? As I will recount later, AI at the very least could be a powerful tool for an artist, perhaps analogous to the way a sophisticated camera is a tool for a fine art photographer. In that case, a human artist still dictates the purpose of the art. For a person using AI art generating programs, their own cultural context, their education, and personal histories influence their choices and modifications the initial “drafts” of images produced by the generator. If AI develops sentience, then questions about the purpose of AI art and its cultural context, if such work is even produced, will come to the fore and challenge our engagement with such art.

Reproduction and access

I mentioned the importance of authenticity in how a child reacts to reproductions and our distaste for forgeries. These observations point to a special status for original artwork. For Benjamin (1936/2018) the original had a unique presence in time and place. He regarded this presence as the artwork’s “aura.” The aura of art depreciates with reproduction.

Reproduction has been an issue in art for a long time. Wood cuts and lithographs (of course the printing press for literature) meant that art could be reproduced and many copies distributed. These copies made art more accessible. Photography and film, vastly increased reproductions of and access to art images.

Even before reproductions, paintings as portable objects within a frame, increased access to art. These objects could be moved to different locations, unlike frescoes or mosaics which had to be experienced in situ (setting aside the removal of artifacts from sites of origin to imperial collections). Paintings that could be transported in a frame already diminished their aura by being untethered to a specific location of origin.

Concerns about reproduction take on a different force in the digital realm. These concerns extend those raised by photographic reproduction. Analog photography retains the ghost of an original- in the form of a negative. Fine art photography often limits prints to a specific number to impart a semblance of originality and introduce scarcity to the physical artifact of a print. Digital photography has no negative. A RAW file might be close. Copies of the digital file, short of being corrupted, are indistinguishable from an original file, calling into question any uniqueness contained in that original. Perhaps non-fungible tokens could be used to establish an original unique identifier for such digital files.

If technology pushes art toward new horizons and commercial opportunities push advances in technology, then it is hard to ignore the likelihood that virtual reality (VR) and augmented reality (AR) will have an impact on our engagement with art. The ease of mass production and commercial imperatives to make more, also renders the notion of the aura of an individual object or specific location in VR nonsensical. AI art, by virtue of being digital, will lack uniqueness and not have the same aura as a specific object tied to a specific time and place. However, the images will be novel. Novelty, as I describe later, is an important feature of creativity.

Artificial intelligence in our lives

As I mentioned at the outset of this essay, machine learning and AI will have a profound effect on almost every aspect of what we do and how we live. Intelligence in current forms of AI is not like human cognition. AI as implemented in deep learning algorithms are not taught rules to guide the processing of their inputs. Their learning takes different forms. They can be supervised, reinforced, or unsupervised. For supervised learning, they are fed massive amounts of labeled data as input and then given feedback about how well their outputs match the desired label. In this way networks are trained to maximize an “objective function,” which typically targets the correct answer. For example, a network might be trained to recognize “dog” and learn to identify dogs despite the fact that dogs vary widely in color, size, and bodily configurations. After being trained on many examples of images that have been labeled a priori as dog, the network identifies images of dogs it has never encountered before. The distinctions between supervised, reinforcement learning, and unsupervised learning are not important to the argument here. Reinforcement learning relies on many trial-and-error iterations and learns to succeed from the errors it makes, especially in the context of games. Unsupervised learning learns by identifying patterns in data and making predictions based on past patterns in that are not labeled.

Artificial intelligence improves with more data. With massive information increasingly available from web searches, commercial purchases, internet posts, texts, official records, all resting on enormous cloud computing platforms, the power of AI is growing and will continue to do so for the foreseeable future. The limits to AI are availability of data and of computational power.

Artificial intelligence does some tasks better than humans. It processes massive amounts of information, generates many simulations, and identifies patterns that would be impossible for humans to appreciate. For example, in biology, AI recently solved the complex problem of three-dimensional protein folding from a two-dimensional code ( Callaway, 2022 ). The output of deep learning algorithms can seem magical ( Rich, 2022 ). Given that they are produced by complex multidimensional equations, their results resist easy explanation.

Current forms of AI have limits. They do not possess common sense. They are not adept at analytical reasoning, extracting abstract concepts, understanding metaphors, experiencing emotions, or making inferences ( Marcus and Davis, 2019 ). Given these limits, how could AI appreciate or produce art? If art communicates abstract and symbolic ideas or expresses nuanced emotions, then an intelligence that cannot abstract ideas or feel emotions would seem ill-equipped to appreciate or produce art. If we care about agency, short of developing sentience, AI has no agency. If we care about purpose, the purpose of an AI system is determined by its objective function. This objective, as of now, is put in place by human designers and the person making use of AI as a tool. If we care about uniqueness, the easy reproducibility of digital outputs depreciates any “aura” to which AI art might aspire.

Despite these reasons to be skeptical, it might be premature to dismiss a significant role of AI in art.

Art appreciation and production

What happens when people appreciate art? Art, when most powerful, can transform a viewer, evoke deep emotions, and promote new understanding of the world and of themselves. Historically, scientists working in empirical aesthetics have asked participants in their studies whether they like a work of art, find it interesting, or beautiful ( Chatterjee and Cardilo, 2021 ). The vast repository of images, on platforms like Instagram, Facebook, Flicker, and Pinterest, have images labeled with people’s preferences. These rich stores of data, growing every day, mean that AI programs can be trained to identify underlying patterns in images that people like.

Crowd-sourcing beauty or preference risks produce boring images. In the 1990s, Komar and Melamid (1999) conducted a pre-digital satirical project in crowd-sourcing art preferences. They hired polling companies to find out what paintings people in 11 countries wanted the most. For Americans, they found that 44% of Americans preferred blue; 49% preferred outdoor scenes featuring lakes, rivers, or oceans; more than 60% liked large paintings; 51% preferred wild, rather than domestic, animals; and 56% said they wanted historical figures featured in the painting. Based on this information, the painting most Americans want showed an idyllic landscape featuring a lake, two frolicking deer, a group of three clothed strollers, and George Washington standing upright in the foreground. For many critics, The Most Wanted Paintings were banal. They were the kind of anodyne images you might find in a motel. Is the Komar and Melamid experiment a cautionary tale for AI?

Artificial intelligence would not be polling people the way that Komar and Melamid did. With a large database of images, including paintings from various collections, the training phase would encompass an aggregate of many more images than collecting the opinions of a few hundred people. AI need not be confined to producing banal images reduced to a low common denominator. Labels for images in databases might end up being far richer than the simple “likes” on Instagram and other social media platforms. Imagine a nuanced taxonomy of words that describe different kinds of art and their potential impacts on viewers. At a small scale, such projects are underway ( Menninghaus et al., 2019 ; Christensen et al., 2022 ; Fekete et al., 2022 ). These research programs go beyond asking people if they like an image, or find it beautiful or interesting. In one such project, we queried a philosopher, a psychologist, a theologian, and art historian and a neuroscientist for verbal labels that could describe a work of art and labels that would indicate potential impacts on how they thought or felt. Descriptions of art could include terms like “colorful” or “dynamic” or refer to the content of art such as portraits or landscapes or to specific art historical movements like Baroque or post-impressionist. Terms describing the impact of art certainly include basic terms such as “like” and “interest,” but also terms like “provoke,” or “challenge,” or “elevate,” or “disgust.” The motivation behind such projects is that powerful art evokes nuanced emotions beyond just liking or disliking the work. Art can be difficult and challenging, and such art might make some viewers feel anxious and others feel more curious. Researchers in empirical aesthetics are increasing focused on identifying a catalog of cognitive and emotional impacts of art. Over the next few years, a large database of art images labeled with a wide range of descriptors and impacts could serve as a training set for an art appreciating AI. Since such networks are adept at extracting patterns in vast amounts of data, one could imagine a trained network describing a novel image it is shown as “playing children in a sunny beach that evokes joy and is reminiscent of childhood summers.” The point is that AI need not know what it is looking at or experience emotions. All it needs to be able to do is label a novel image with descriptions and impacts- a more complex version of labeling an image as a brown dog even if it has never seen that particular dog before.

Can AI, in its current form, be creative? One view is that AI is and will continue to be good at automated but not creative tasks. As AI disrupts work and replaces jobs that involve routine procedures, the hope is that creative jobs will be spared. This hope is probably not warranted.

Sequence transduction or transformer models are making strides in processing natural language. Self-GPT-3 (generative pre-trained transformers) as of now building on 45 terabytes of data can produce text based on the likelihood of words co-occurring in sequence. The words produced by transformer models can seem indistinguishable from sentences produced by humans. GPT-3 transformers can produce poetry, philosophical musings, and even self-critical essays ( Thunström, 2022 ).

The ability to use text to display images is the first step in producing artistic images. DALL-E 2, Imagen, Midjourney, and DreamStudio are gaining popularity as art generators that make images when fed words ( Kim, 2022 ). To give readers, who might not be familiar with the range of AI art images, a sense of these pictures I offer some examples.

The first set of images were made using Midjourney. I started with the prompt “a still life with fruit, flowers, a vase, dead game, a candle, and a skull in a Renaissance style” ( Figure 1 ). The program generates four options, from which I picked the one that came closest to how I imagined the image. I then generated another four variations from the one I picked and chose the one I liked best. The upscaled version of the figure is included.

Midjourney image generated to the prompt “a still life with fruit, flowers, a vase, dead game, a candle, and a skull in a Renaissance style”.

To show variations of the kind of images produced, I used the same procedures and prompts, except changing the style to Expressionist, Pop-art, and Minimalist ( Figures 2 – 4 ).

Midjourney image generated to the prompt “a still life with fruit, flowers, a vase, dead game, a candle, and a skull in an Expressionist style”.

Midjourney image generated to the prompt “a still life with fruit, flowers, a vase, dead game, a candle, and a skull in a Minimalist style”.

Midjourney image generated to the prompt “a still life with fruit, flowers, a vase, dead game, a candle, and a skull in a Pop-art style”.

“To show how one might build up an image I used Open AI’s program Dall-E, to generate an image to the prompt, “a Surreal Impressionist Landscape.” Then using the same program, I used the prompt, “a Surreal Impressionist Landscape that evokes the feeling of awe.” To demonstrate how different programs can produce different images to the same prompt,” a Surreal Impressionist Landscape that evokes the feeling of awe” I include images produced by Dream Studio and by Midjourney.

Regardless of the merits of each individual image, they only took a few minutes to make. Such images and many other produced easily could serve as drafts for an artist to consider the different ways they might wish to depict their ideas or give form to their intuitions ( Figures 5 – 8 ). The idea that artists use technology to guide their art is not new. For example, Hockney (2001) described ways that Renaissance masters used technology of their time to create their work.

Dall-E generated image to the prompt “a Surreal Impressionist Landscape”.

Midjourney generated image to the prompt “a Surreal Impressionist Landscape that evokes the feeling of awe”.

Dall-E generated image to the prompt “a Surreal Impressionist Landscape that evokes the feeling of awe”.

Dream Studio generated image to the prompt “a Surreal Impressionist Landscape that evokes the feeling of awe”.

Unlike the imperative for an autonomous vehicle to avoid mistakes when it needs to recognize a child playing in the street, art makes no such demands. Rather, art is often intentionally ambiguous. Ambiguity can fuel an artworks’ power, forcing viewers to ponder what it might mean. What then will be the role of the human artist? Most theories of creative processing include divergent and convergent thinking ( Cortes et al., 2019 ). Divergent thinking includes coming up with many possibilities. This phase can also be thought of as the generative or imaginative phase. A commonly used laboratory test is the Alternative Uses Test ( Cortes et al., 2019 ). This test asks people to offer as many uses of a common object, like a brick, that they can imagine. The more uses, that a person can conjure up, especially when they are unusual, is taken as a measure of divergent thinking and creative potential. When confronting a problem that needs a creative solution, generating many possibilities doesn’t mean that they are the right or the best one. An evaluative phase is needed to narrow the possibilities, to converge on a solution, and to identify a useful path forward. In producing a work of art, artists presumably shift back and forth between divergent and convergent processes as they keep working toward their final work.

An artist could use text-to-image platforms as a tool ( Kim, 2022 ). They could type in their intent and then evaluate the possible images generated, as I show in the figures. They might tweak their text several times. The examples of images included here using similar verbal prompts show how the text can be translated into images differently. Artists could choose which of the images generated they like and modify them. The divergent and generative parts of creative output could be powerfully enhanced by using AI, while the artist would evaluate these outputs. AI would be a powerful addition to their creative tool-kit.

Some art historians might object that art cannot be adequately appreciated outside its historical and cultural context. For example, Picasso and Matisse are better understood in relation to Cezanne. The American abstract expressionists are better understood as expressing an individualistic spirit while still addressed universal experiences; a movement to counter Soviet social realism and its collective ethos. We can begin to see how this important objection might be dealt with using AI. “Creative adversarial networks” can produce novel artworks by learning about historic art styles and then intentionally deviating from them ( Elgammal et al., 2017 ). These adversarial networks would use other artistic styles as a contextual springboard from which to generate images.

Artificial intelligence and human artists might be partners ( Mazzone and Elgammal, 2019 ), rather than one serving as a tool for the other. For example, in 2015 Mike Tyka created large-scale artworks using Iterative DeepDream and co-founded the Artists and Machine Intelligence program at Google. Using DeepDream and GANs he produced a series “Portraits of Imaginary People,” which was shown at ARS Electronica in Linz, Christie’s in New York and at the New Museum in Karuizawa (Japan) ( Interalia Magazine, 2018 ). The painter Pindar van Arman teaches robots to paint and believes they augment his own creativity. Other artists are increasingly using VR as an enriched and immersive experience ( Romano, 2022 ).

Kinsella (2018) Christie’s in New York sold an artwork called Portrait of Edmond de Belamy for $432,500. The portrait of an aristocratic man with blurry features was created by a GAN from a collective called Obvious. It was created using the WikiArt dataset that includes fifteen thousand portraits from the fourteenth to the twentieth century. Defining art has always been difficult. Art does not easily follow traditional defining criteria of having sufficient and necessary features to be regarded as a member of a specific category, and may not be a natural kind ( Chatterjee, 2014 ). One prominent account of art is an institutional view of art ( Dickie, 1969 ). If our social institutions agree that an object is art, then it is. Being auctioned and sold by Christie’s certainly qualifies as an institution claiming that AI art is in fact art.

In 2017, Turkish artist Refik Anadol, collaborating with Mike Tyka, created an installation using GANs called “Archive Dreaming.” This installation is an immersive experience with viewers standing in a cylindrical room. He used Istanbul’s SALT Galeta online library with 1.7 million images, all digitized into two terabytes of data. The holdings in this library relate to Turkey from the 19th Century to the present and include photographs, images, maps, and letters. Viewers stand in a cylindrical room and can gaze at changing displays on the walls. They can choose which documents to view, or the passively watch the display in an idle state. In the idle state, the archive “dreams.” Generators produce new images that resemble the original ones, but never actually existed—an alternate fictional historical archive of Turkey imagined by the machine ( Pearson, 2022 ).

Concerns, further future, and sentient artificial intelligence

Technology can be misused. One downside of deep learning is that biases embedded in training data sets can be reified. Systematic biases in the judicial system, in hiring practices, in procuring loans are written into AI “predictions” while giving the illusion of objectivity. The images produced by Dall-E so far perpetuate race and gender stereotypes ( Taylor, 2022 ). People probably do not vary much if asked to identify a dog, but they certainly do in identifying great art. Male European masters might continue to be lauded over women or under-represented minority artists and others of whom we have not yet heard.

On the other hand, current gatekeepers of art, whether at high-end galleries, museums, and biennales, are already biased in who and what art they promote. Over time, art through AI might become more democratized. Museums and galleries across the world are digitizing their collections. The art market in the 21st Century extends beyond Europe and the United States. Important shows as part of art’s globalization occur beyond Venice, Basel, and Miami—to now include major gatherings in Sao Paulo, Dakar, Istanbul, Sharjah, Singapore, and Shanghai. Beyond high profile displays, small galleries are digitizing and advertising their holdings. As more images are incorporated into training databases, including art from Asia, Africa, and South America, and non-traditional art forms, such as street art or textile art, what people begin to regard as good or great art might become more encompassing and inclusive.

Could art become a popularity contest? As museums struggle to keep a public engaged, they might use AI to predict which kinds of art would draw in most viewers. Such a use of AI might narrow the range of art that are displayed. Similarly, some artists might choose to make art (in the traditional way), but shift their output to what AI predicts will sell. Over time, art could lose its innovation, its subversive nature, and its sheer variety. The nature of the artist might also change if the skills involved in making art change. An artist collaborating with AI might use machine learning outputs for the divergent phase of their creations and insert themselves along with additional AI assessments in the convergent evaluative phases of producing art.

The need for artistic services could diminish. Artists who work as illustrators for books, technical manuals, and other media such as advertisement, could be replaced by AI generating images. The loss of such paying jobs might make it harder for some artists to pursue their fine art dreams if they do not have a reliable source of income.

Many experts working in the field believe that AI will develop sentience. Exactly how is up for debate. Some believe that sentience can emerge from deep learning architectures given enough data and computational power. Others think that combining deep learning and classical programming, which includes the insertion of rules and symbols, is needed for sentience to emerge. Experts also vary in when they think sentience will emerge in computers. According to Ford (2021) , some think it could be in a decade and others in over a 100 years. Nobody can anticipate the nature of that sentience. When Gary Kasparov (world Chess Champion at the time) lost to the program Deep Blue, he claimed that he felt an alien intelligence ( Lincoln, 2018 ). Deep Blue was no sentient AI.

Artificial intelligence sentience will truly be an alien intelligence. We have no idea how or whether sentient AI will engage in art. If they do, we have no idea what would motivate them and what purpose their art would have. Any comments about these possibilities are pure speculation on my part.

Sentient AI could make art in the real world. Currently, robots find and move objects in large warehouses. Their movements are coarse and carried out in well-controlled areas. A robot like Rosey, the housekeeper in the Jetsons cartoon, is far more difficult to make since it has to move in an open world and react to unpredictable contingencies. Large movements are easier to program than fine movements, precision grips, and manual dexterity. The difficulty in making a robot artist would fall somewhere between a robot in an Amazon warehouse and Rosey. It would not have to contend with an unconstrained environment in its “studio.” It would learn to choose and grip different brushes and other instruments, manipulate paints, and apply them to a canvas that it stretched. Robot arms that draw portraits have been programed into machines ( Arman, 2022 ). However, sentient AI with intent would decide what to paint and it would be able to assess whether its output matched its goal- using generative adversarial systems. The art appreciation and art production abilities could be self-contained within a closed loop without involving people.

Sentient AI might not bother with making art in the real world. Marc Zuckerberg would have us spend as much time as possible in a virtual metaverse. Sentient AI could create art residing in fantastical digital realms and not bother with messy materials and real-world implementation. Should sentient AI or sentient AIs choose to make art for whatever their purpose might be, humans might be irrelevant to the art making and appreciating or evaluating loop.

Ultimately, we do not know if sentient AI will be benevolent, malevolent, or apathetic when it comes to human concerns. We don’t know if sentient AI will care about art.

As AI continues to insinuate itself in most parts of our lives, it will do so with art ( Agüera y Arcas, 2017 ; Miller, 2019 ). The beginnings of art appreciation and production that we see now, and the examples provided in the figures, might be like the video game Pong that was popular when I was in high school. Pong is a far cry from the rich immersive quality of games like Minecraft in the same way that Dall-E and Midjourney images might be a far cry from a future art making and appreciating machine.

The idea that creative pursuits are an unassailable bastion of humanity is untenable. AI is already being used as a powerful tool and even as a partner for some artists. The ongoing development of aesthetically sensitive machines will challenge our views of beauty and creativity and perhaps our understanding of the nature of art.

Author contributions

The author confirms being the sole contributor of this work and has approved it for publication.

Acknowledgments

I appreciate the helpful feedback I received from Alex Christensen, Kohinoor Darda, Jonathan Fineberg, Judith Schaechter, and Clifford Workman.

Conflict of interest

The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

• Agüera y Arcas B. (2017). Art in the age of machine intelligence. Arts 6 : 18 . 10.3390/arts6040018 [ CrossRef ] [ Google Scholar ]
• Arman P. V. (2022). Cloud painter. Available online at: https://www.cloudpainter.com/ (accessed August 10, 2022). [ Google Scholar ]
• Benjamin W. (1936/2018). The work of art in the age of mechanical reproduction. A museum studies approach to Heritage. London: Routledge. 10.4324/9781315668505-19 [ CrossRef ] [ Google Scholar ]
• Callaway E. (2022). ‘The entire protein universe’: AI predicts shape of nearly every known protein. Nature 608 15–16. 10.1038/d41586-022-02083-2 [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Chamberlain R., Mullin C., Scheerlinck B., Wagemans J. (2018). Putting the art in artificial: Aesthetic responses to computer-generated art. Psychol. Aesth. Creat. Arts 12 177–192. 10.1037/aca0000136 [ CrossRef ] [ Google Scholar ]
• Chatterjee A. (2014). The aesthetic brain: How we evolved to desire beauty and enjoy art. New York, NY: Oxford University Press. 10.1093/acprof:oso/9780199811809.001.0001 [ CrossRef ] [ Google Scholar ]
• Chatterjee A., Cardilo E. (2021). Brain, beauty, and art: Essays bringing neuroaesthetics into focus. Oxford: Oxford University Press. 10.1093/oso/9780197513620.001.0001 [ CrossRef ] [ Google Scholar ]
• Christensen A. P., Cardillo E. R., Chatterjee A. (2022). What kind of impacts can artwork have on viewers? Establishing a taxonomy for aesthetic cognitivism. PsyArXiv [Preprint] 10.31234/osf.io/nt59q [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Cortes R. A., Weinberger A. B., Daker R. J., Green A. E. (2019). Re-examining prominent measures of divergent and convergent creativity. Curr. Opin. Behav. Sci. 27 90–93. 10.1016/j.cobeha.2018.09.017 [ CrossRef ] [ Google Scholar ]
• Dickie G. (1969). Defining art. Am. Philos. Q. 6 253–256. [ Google Scholar ]
• Dissanayake E. (2008). “ The arts after Darwin: Does art have an origin and adaptive function? ,” in World art studies: Exploring concepts and approaches , eds Zijlemans K., Van Damme W. (Amsterdam: Valiz; ). 10.1007/s13187-012-0411-7 [ CrossRef ] [ Google Scholar ]
• Elgammal A., Liu B., Elhoseiny M., Mazzone M. (2017). Can: Creative adversarial networks, generating “art” by learning about styles and deviating from style norms. arXiv [Preprint]. arXiv:1706.07068. [ Google Scholar ]
• Fekete A., Pelowski M., Specker E., Brieber D., Rosenberg R., Leder H. (2022). The Vienna art picture system (VAPS): A data set of 999 paintings and subjective ratings for art and aesthetics research . Psychol. Aesthet. Creat. Arts. 10.1037/aca0000460 [Epub ahead of print]. [ CrossRef ] [ Google Scholar ]
• Fineberg J. D. (1995). Art since 1940. Hoboken, NJ: Prentice-Hall. [ Google Scholar ]
• Ford M. (2021). Rule of the robots: How artificial intelligence will transform everything. Hachette: Basic Books. [ Google Scholar ]
• Hawley-Dolan A., Winner E. (2011). Seeing the mind behind the art: People can distinguish abstract expressionist paintings from highly similar paintings by children, chimps, monkeys, and elephants. Psychol. Sci. 22 435–441. 10.1177/0956797611400915 [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Hockney D. (2001). Secret knowledge: Rediscovering the lost techniques of the old masters. London: Thames & Hudson. [ Google Scholar ]
• Hood B. M., Bloom P. (2008). Children prefer certain individuals over perfect duplicates. Cognition 106 455–462. 10.1016/j.cognition.2007.01.012 [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Interalia Magazine (2018). Portraits of imaginary people. Available online at: https://www.interaliamag.org/audiovisual/mike-tyka/ (accessed August 15, 2022). [ Google Scholar ]
• Kim T. (2022). The future of creativity, brought to you by artificial intelligence. Available online at: https://venturebeat.com/datadecisionmakers/the-future-of-creativity-brought-to-you-by-artificial-intelligence/ (accessed August 9, 2022). [ Google Scholar ]
• Kinsella E. (2018). The first ai-generated portrait ever sold at auction shatters expectations, fetching $432,500—43 times its estimate. Available online at: https://news.artnet.com/market/first-ever-artificial-intelligence-portrait-painting-sells-at-christies-1379902 (accessed August 10, 2022). [ Google Scholar ]
• Kirk U., Skov M., Hulme O., Christensen M. S., Zeki S. (2009). Modulation of aesthetic value by semantic context: An fMRI study. NeuroImage 44 1125–1132. 10.1016/j.neuroimage.2008.10.009 [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Komar V., Melamid A. (1999). Painting by numbers: Komar and Melamid’s scientific guide to art. Berkeley, CA: University of California Press. [ Google Scholar ]
• Kruger J., Wirtz D., Van Boven L., Altermatt T. W. (2004). The effort heuristic. J. Exp. Soc. Psychol. 40 91–98. 10.1016/S0022-1031(03)00065-9 [ CrossRef ] [ Google Scholar ]
• Lee K. F., Qiufan C. (2021). AI 2041: Ten visions for our future. London: Ebury Publishing. [ Google Scholar ]
• Lincoln K. (2018). Deep you. Available online at: https://www.theringer.com/tech/2018/11/8/18069092/chess-alphazero-alphago-go-stockfish-artificial-intelligence-future (accessed August 16, 2022). [ Google Scholar ]
• Marcus G., Davis E. (2019). Rebooting AI: Building artificial intelligence we can trust. New York, NY: Knopf Doubleday Publishing Group. [ Google Scholar ]
• Mazzone M., Elgammal A. (2019). Art, creativity, and the potential of artificial intelligence. Arts 8 : 26 . 10.3390/arts8010026 [ CrossRef ] [ Google Scholar ]
• Menninghaus W., Wagner V., Wassiliwizky E., Schindler I., Hanich J., Jacobsen T., et al. (2019). What are aesthetic emotions? Psychol. Rev. 126 171 . 10.1037/rev0000135 [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Miller A. I. (2019). The artist in the machine: The world of AI-powered creativity. Cambridge, MA: MIT Press. 10.7551/mitpress/11585.001.0001 [ CrossRef ] [ Google Scholar ]
• Newman G. E. (2019). The psychology of authenticity. Rev. Gen. Psychol. 23 8–18. 10.1037/gpr0000158 [ CrossRef ] [ Google Scholar ]
• Newman G. E., Bloom P. (2012). Art and authenticity: The importance of originals in judgments of value. J. Exp. Psychol 141 558–569. 10.1037/a0026035 [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Newman G. E., Bartels D. M., Smith R. K. (2014). Are artworks more like people than artifacts? Individual concepts and their extensions. Top. Cogn. Sci. 6 647–662. 10.1111/tops.12111 [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Pearson A. (2022). Archive dreaming. Available online at: http://www.digiart21.org/art/archive-dreaming (accessed August 10, 2022). [ Google Scholar ]
• Rich S. (2022). The new poem-making machinery. Available online at: https://www.newyorker.com/culture/culture-desk/the-new-poem-making-machinery (accessed August 10, 2022). [ Google Scholar ]
• Romano H. (2022). 8 virtual reality artists who use the world as their canvas. Available online at: https://blog.kadenze.com/creative-technology/8-virtual-reality-artists-who-use-the-world-as-their-canvas/ (accessed August 8, 2022). [ Google Scholar ]
• Shiner L. (2001). The invention of art. A cultural history. Chicago, IL: University of Chicago Press. 10.7208/chicago/9780226753416.001.0001 [ CrossRef ] [ Google Scholar ]
• Snapper L., Oranç C., Hawley-Dolan A., Nissel J., Winner E. (2015). Your kid could not have done that: Even untutored observers can discern intentionality and structure in abstract expressionist art. Cognition 137 154–165. 10.1016/j.cognition.2014.12.009 [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Taylor J. (2022). No Quick fix: How OpenAI’s DALL-E 2 illustrated the challanges of bias in AI. Available online at: https://www.nbcnews.com/tech/tech-news/no-quick-fix-openais-dalle-2-illustrated-challenges-bias-ai-rcna39918 (accessed August 10, 2022). [ Google Scholar ]
• Thunström A. O. (2022). We asked GPT-3 to write an academic paper about itself—then we tried to get it published. Berlin: Scientific American. [ Google Scholar ]

Ethical dilemmas of AI-generated art and stories

ChatGPT, DALL-E, Midjourney, Stable Diffusion—names that most of us hadn’t heard more than a couple of years ago now represent a slew of creative programs powered by artificial intelligence.

Large language model AI programs can write stories and articles, make illustrations and artwork, and converse with users using prompts. But what does it mean for human artists and writers? Will AI steal jobs and creative works? How should people approach the thorny ethical thicket around AI-generated art?

Mark Fagiano, an assistant professor of philosophy  at Washington State University, talks with Larry Clark, editor of Washington State Magazine , about how ethics in action and pragmatism can help people examine not only AI art, but any rapidly evolving technology and issues in society.

How do you like the magazine podcast? What WSU stories do you want to hear? Let us know .

WSU Honors College course tackles ethics of ChatGPT (WSU philosophy professor Samantha Noll, WSU Insider, March 31, 2023)

“ When will artificial intelligence really pass the test? ” and “ AI for wildlife conservation—from an AI ”  (in current issue)

Will ChatGPT Kill the Student Essay?  ( The Atlantic )

ChatGPT, Galactica, and the Progress Trap  ( Wired )

Teaching Experts Are Worried About ChatGPT, but Not for the Reasons You Think  ( The Chronicle of Higher Education )

Your Creativity Won’t Save Your Job From AI  ( The Atlantic )

It’s Time to Pay Attention to A.I. (ChatGPT and Beyond)  (Video by ColdFusion)

ChatGPT proves AI is finally mainstream — and things are only going to get weirder  (The Verge)

AI Is Taking On Ever-Larger Puzzles  ( Wired )

The ethical use of AI content – A basic guideline  (ContentBot.ai blog)

“ The Expanding Dark Forest and Generative AI ”  (maggieappleton.com)

Why Does AI Art Look Like That?

Every tech company wants its image generator to be the best. But they all produce oddly similar work.

This week, X launched an AI-image generator, allowing paying subscribers of Elon Musk’s social platform to make their own art. So—naturally—some users appear to have immediately made images of Donald Trump flying a plane toward the World Trade Center ; Mickey Mouse wielding an assault rifle, and another of him enjoying a cigarette and some beer on the beach; and so on. Some of the images that people have created using the tool are deeply unsettling; others are just strange, or even kind of funny. They depict wildly different scenarios and characters. But somehow they all kind of look alike, bearing unmistakable hallmarks of AI art that have cropped up in recent years thanks to products such as Midjourney and DALL-E.

Two years into the generative-AI boom, these programs’ creations seem more technically advanced—the Trump image looks better than, say, a similarly distasteful one of SpongeBob SquarePants that Microsoft’s Bing Image Creator generated last October—but they are stuck with a distinct aesthetic. The colors are bright and saturated, the people are beautiful, and the lighting is dramatic. Much of the imagery appears blurred or airbrushed, carefully smoothed like frosting on a wedding cake. At times, the visuals look exaggerated. (And yes, there are frequently errors, such as extra fingers .) A user can get around this algorithmic monotony by using more specific prompts—for example, by typing a picture of a dog riding a horse in the style of Andy Warhol rather than just a picture of a dog riding a horse . But when a person fails to specify, these tools seem to default to an odd blend of cartoon and dreamscape.

These programs are becoming more common. Google just announced a new AI-image-making app called Pixel Studio that will allow people to make such art on their Pixel phone. The app will come preinstalled on all of the company’s latest devices. Apple will launch Image Playground as part of its Apple Intelligence suite of AI tools later this year . OpenAI now allows ChatGPT users to generate two free images a day from DALL-E 3, its newest text-to-image model. (Previously, a user needed a paid premium plan to access the tool.) And so I wanted to understand: Why does so much AI art look the same?

Read: AI has a hotness problem

The AI companies themselves aren’t particularly forthcoming. X sent back a form email in response to a request for comment about its new product and the images its users are creating. Four firms behind popular image generators—OpenAI, Google, Stability AI, and Midjourney—either did not respond or did not provide comment. A Microsoft spokesperson directed me toward some of its prompting guides and referred any technical questions to OpenAI, because Microsoft uses a version of DALL-E in products such as Bing Image Creator.

So I turned to outside experts, who gave me four possible explanations. The first focuses on the data that models are trained on. Text-to-image generators rely on extensive libraries of photos paired with text descriptions, which they then use to create their own original imagery. The tools may inadvertently pick up on any biases in their data sets—whether that’s racial or gender bias, or something as simple as bright colors and good lighting. The internet is filled with decades of filtered and artificially brightened photos, as well as a ton of ethereal illustrations. “We see a lot of fantasy-style art and stock photography, which then trickles into the models themselves,” Zivvy Epstein, a scientist at the Stanford Institute for Human-Centered AI, told me. There are also only so many good data sets available for people to use to build image models, Phillip Isola, a professor at the MIT Computer Science & Artificial Intelligence Laboratory, told me, meaning the models might overlap in what they’re trained on. (One popular one, CelebA , features 200,000 labeled photos of celebrities. Another, LAION 5B , is an open-source option featuring 5.8 billion pairs of photos and text.)

The second explanation has to do with the technology itself. Most modern models use a technique called diffusion: During training, models are taught to add “noise” to existing images, which are paired with text descriptions. “Think of it as TV static,” Apolinário Passos, a machine-learning art engineer at Hugging Face, a company that makes its own open-source models, told me. The model then is trained to remove this noise, over and over, for tens of thousands, if not millions, of images. The process repeats itself, and the model learns how to de-noise an image. Eventually, it’s able to take this static and create an original image from it. All it needs is a text prompt.

Read: Generative art is stupid

Many companies use this technique. “These models are, I think, all technically quite alike,” Isola said, noting that recent tools are based on the transformer model. Perhaps this technology is biased toward a specific look. Take an example from the not-so-distant past: Five years ago, he explained, image generators tended to create really blurry outputs. Researchers realized that it was the result of a mathematical fluke; the models were essentially averaging all the images they were trained on. Averaging, it turns out, “looks like blur.” It’s possible that, today, something similarly technical is happening with this generation of image models that leads them to plop out the same kind of dramatic, highly stylized imagery—but researchers haven’t quite figured it out yet. Additionally, “most models have an ‘aesthetic’ filter on both the input and output that reject images that don't meet a certain aesthetic criteria,” Hany Farid,  a professor at the UC Berkeley School of Information, told me over email. “This type of filtering on the input and output is almost certainly a big part of why AI-generated images all have a certain ethereal quality.”

The third theory revolves around the humans who use these tools. Some of these sophisticated models incorporate human feedback; they learn as they go. This could be by taking in a signal, such as which photos are downloaded. Others, Isola explained, have trainers manually rate which photos they like and which ones they don’t. Perhaps this feedback is making its way into the model. If people are downloading art that tends to have really dramatic sunsets and absurdly beautiful oceanscapes, then the tools might be learning that that’s what humans want, and then giving them more of that. Alexandru Costin, a vice president of generative AI at Adobe, and Zeke Koch, a vice president of product management for Adobe Firefly (the company’s AI-image tool) told me in an email that user feedback can indeed be a factor for some AI models—a process called “reinforcement learning from human feedback,” or RLHF. They also pointed to training data as well as assessments performed by human evaluators as influencing factors. “Art generated by AI models sometimes have a distinct look (especially when created using simple prompts),” they said in a statement. “That’s generally caused by a combination of the images used to train the image output and the tastes of those who train or evaluate the images.”

The fourth theory has to do with the creators of these tools. Although representatives for Adobe told me that their company does not do anything to encourage a specific aesthetic, it is possible that other AI makers have picked up on human preference and coded that in—essentially putting their thumb on the scale, telling the models to make more dreamy beach scenes and fairylike women. This could be intentional: If such imagery has a market, maybe companies would begin to converge around it. Or it could be unintentional; companies do lots of manual work in their models to combat bias, for example, and various tweaks favoring one kind of imagery over another could inadvertently result in a particular look.

More than one of these explanations could be true. In fact, that’s probably what’s happening: Experts told me that, most likely, the style we see is caused by multiple factors at once. Ironically, all of these explanations suggest that the uncanny scenes we associate with AI-generated imagery are actually a reflection of our own human preferences, taken to an extreme. No surprise, then, that Facebook is filled with AI-generated slop imagery that earns creators money, that Etsy recently asked users to label products made with AI following a surge of junk listings , and that the arts-and-craft store Michaels recently got caught selling a canvas featuring an image that was partially generated by AI (the company pulled the product , calling this an “unacceptable error.”).

Read: AI-generated junk is flooding Etsy

AI imagery is poised to seep even further into everyday life. For now, such art is usually visually distinct enough that people can tell it was made by a machine. But that may change. The technology could get better. Passos told me he sees “an attempt to diverge from” the current aesthetic “on newer models.” Indeed, someday computer-generated art may shed its weird, cartoonish look, and start to slip past us unnoticed. Perhaps then we’ll miss the corny style that was once a dead giveaway.

About the Author

More Stories

California’s Fire-Insurance Crisis Just Got Real

The Worst Feature Apple Ever Made

More From Forbes

Is ai-generated art ‘true art’ implications and considerations for businesses.

• Share to Facebook
• Share to Twitter
• Share to Linkedin

Ben Meisner is the founder of the leading online photo editing platform Ribbet.com .

In recent times a flurry of AI-powered photo and video editing tools have emerged, ranging from those that simply save time by automatically removing backgrounds, to more sophisticated tools that replace people ( deepfakes ) or that generate completely new and realistic scenes from scratch, such as DALL·E and Stable Diffusion.

What is the driving force behind the creation of these tools and algorithms?

When you consider the development of AI technology in driverless cars, there are clear problems that are being addressed. For example, there’s the potential benefit of fewer accidents, where approximately 1.3 million people currently lose their lives each year in road fatalities.

But when it comes to photo and video manipulation, what is it we are ultimately working toward solving? Are engineers creating these algorithms purely because new technology has unlocked the potential to create them, or is there some deep problem mankind is working toward solving?

This topic is being strongly debated within the creative ecosystem, with some companies going so far as banning the use of AI-generated media because they could breach copyright law.

The conversation shouldn’t stop here though—in addition to addressing the question of whether AI-generated media is truly original, I think we need to consider if these works can be considered “true art.”

How does this impact businesses and creative professionals?

This consideration has a significant impact on business. If AI can elevate into activities traditionally considered human, then that will transform the perception of consumers when it comes to future products that will similarly elevate into those realms: think self-driving cars, AI copywriters and photographers, or AI personal assistants.

By this, I mean AI will have to be widely accepted as reliable and human-like before it will be trusted to do the jobs that humans have traditionally done. For example, before we’ll trust a car to drive us around using AI, we have to have trust for AI generally. So if AI could generate something that was considered “true art,” this could be a step in the direction of doing something that traditionally only humans could do.

Today art is still generally defined as a human-performed activity, created by those who have been educated to capture the essence of things and their feelings in an appealing form that involves the imagination. We commonly think of art as those forms of expression that come from someone’s emotions and that we relate to on a human level.

What happens, though, when a new work is generated by an emotionless machine? Traditionally, we’d consider that a heartless act of reference. A silicon chip running a program, using millions of reference media to create something new—could that ever be considered an “artist”? Certainly it isn’t intentional art; it represents the intention of the machine’s user.

The matter of it being a referential work doesn’t necessarily pose any problem to it being considered true art. Great authors, musicians and painters have always taken inspiration from the work of those who went before them. Beyond inspiration, there are rules—or a structured breaking of rules—that make or break a great piece of art, and those can be taught, even to a machine. These rules exist everywhere—in our language, art, storytelling and in business.

A computer can absorb these rules and structures and generate music for example. Could it ever show vulnerability though? If AI was to embark on creative endeavors that evoked our emotions and were generally accepted as authentic art, there might be one less thing that makes us unique as humans.

It would be easy to stop here and declare AI-generated art as inauthentic. Photography has also fought a long and controversial fight to be considered true art. Today we have reached a point where it is accepted, largely because there is a human artist who frames it. Could machine-generated art, which runs through an algorithm a human has created and starts with a human entering a sentence, ever be considered to be “framed” in that same sense?

There would be benefits if it was.

Technology could expand professional skill sets and capabilities—not simply replace them.

AI-generated art could give those people without traditional artistic abilities the possibility to express their feelings and vulnerability through art. There would also be benefits for business. Today people can already generate a layout or design without any design skill by using ready-made templates for what they’re trying to create.

Imagine if at some point you could demand of a computer: “I need a new header for my website—something inspiring, showing a woman wearing our company’s shoes, in blue. Put that image in a collage with women of different looks, and give me 10 variations on that.”

Exciting, right? Well the likes of DALL·E and Meta working on text-to-video generation are certainly a step in that direction.

AI-generated art could facilitate design and creative processes, reducing the number of resources needed to launch a beautiful campaign visual, for example, and acting as a tool that the creative industry uses, and not one that replaces it.

Be mindful of copyright and licensing challenges.

At the same time, discussions about copyright issues and licensing are still up in the air, and entrepreneurs shouldn’t ignore them. The reference materials used to train AI-algorithms frequently include copyrighted materials and so unintentional infringement is a possibility. The technology hasn’t matured yet, so licensing needs to be taken seriously right now.

As appealing as AI-generated art may be, care is needed in how these tools are applied—running an advertising campaign using it for example, would prove problematic if copyright was infringed upon.

At this moment in time, though, these AI-generation tools seem to have limited real-world application. We’re seeking a useful purpose for them rather than initially setting off to solve some particular problem. At the same time, though, I can’t help but think that this may all change as the technology matures and new and exciting potential applications emerge. I’m also drawn to the possibility that our definition of true art and what it encompasses may soon evolve to include AI-generated art.

Forbes Business Council is the foremost growth and networking organization for business owners and leaders. Do I qualify?

• Editorial Standards
• Reprints & Permissions

• < Previous

Home > RESEARCH_SCHOLARSHIP_RESOURCES > RS > FACULTY-RESEARCH-PAPERS > 630

Faculty Scholarship

Simulacra and historical fidelity in digital recreation of lost cultural heritage: reconstituting period materialities for the period eye.

Trent Olsen , Lindenwood University Follow James Hutson , Lindenwood University Follow Charles O'Brien , University of Alabama, Huntsville Jeremiah Ratican , Lindenwood University Follow

Document Type

Publication title.

Arts & Communication

The advancement of digital technologies in art history has opened avenues for reconstructing lost or damaged cultural heritage, a need highlighted by the deteriorated state of many artworks from the 1785 Salon. Grounded in the concept of the “Period Eye” by art historian Michael Baxandall, which emphasizes understanding artworks within their original historical and cultural contexts, this study proposes a subfield focused on Reconstituting Period Materialities for the Period Eye. This methodology bridges comprehensive historical research with generative visual artificial intelligence (AI) technologies, facilitating the creation and immersive virtual reality viewing of artworks. Beyond mere visual replication, the approach aims to recreate the material and textural realities of the period, thereby enabling contemporary audiences to experience these works as they were originally perceived. The process includes replicating building materials using Quixel Megascans, employing AI for generating images of lost artworks, and utilizing normal maps for simulating painting textures, all contributing to an authentic reconstruction of the Salon’s ambiance and materiality. This approach, met with some skepticism from traditional historians and archeologists, asserts that such digital reconstitution, backed by rigorous empirical research and detailed period-specific datasets, yields reconstructions of greater historical accuracy and contextual richness. This mirrors strides in sound archeology, endorsing a similar empirical approach in visual material recreation. The significance of this study is underscored by its potential to enrich our comprehension of historical artworks through a “Period Eye,” blending historical insights with modern technological innovation for a deeper understanding and appreciation of cultural heritage.

10.36922/ac.2719

Publication Date

Creative commons license.

Recommended Citation

Olsen, Trent; Hutson, James; O'Brien, Charles; and Ratican, Jeremiah, "Simulacra and historical fidelity in digital recreation of lost cultural heritage: Reconstituting period materialities for the period eye" (2024). Faculty Scholarship . 630. https://digitalcommons.lindenwood.edu/faculty-research-papers/630

Since May 14, 2024

Included in

Artificial Intelligence and Robotics Commons , History of Art, Architecture, and Archaeology Commons

• Collections
• Disciplines

Advanced Search

• Notify me via email or RSS

Author Corner

• Submit Research

Home | About | FAQ | My Account | Accessibility Statement

Privacy Copyright

• Mobile Site
• Staff Directory
• Advertise with Ars

Filter by topic

• Biz & IT
• Gaming & Culture

Front page layout

Back to the drawing board —

Artists claim “big” win in copyright suit fighting ai image generators, artists prepare to take on ai image generators as copyright suit proceeds..

Ashley Belanger - Aug 14, 2024 9:09 pm UTC

Artists defending a class-action lawsuit are claiming a major win this week in their fight to stop the most sophisticated AI image generators from copying billions of artworks to train AI models and replicate their styles without compensating artists.

In an order on Monday, US District Judge William Orrick denied key parts of motions to dismiss from Stability AI, Midjourney, Runway AI, and DeviantArt. The court will now allow artists to proceed with discovery on claims that AI image generators relying on Stable Diffusion violate both the Copyright Act and the Lanham Act, which protects artists from commercial misuse of their names and unique styles.

"We won BIG," an artist plaintiff, Karla Ortiz, wrote on X (formerly Twitter), celebrating the order. "Not only do we proceed on our copyright claims," but "this order also means companies who utilize" Stable Diffusion models and LAION-like datasets that scrape artists' works for AI training without permission "could now be liable for copyright infringement violations, amongst other violations."

Lawyers for the artists, Joseph Saveri and Matthew Butterick, told Ars that artists suing "consider the Court's order a significant step forward for the case," as "the Court allowed Plaintiffs' core copyright-infringement claims against all four defendants to proceed."

Stability AI was the only company that responded to Ars' request to comment, but it declined to comment.

Artists prepare to defend their livelihoods from AI

To get to this stage of the suit, artists had to amend their complaint to better explain exactly how AI image generators work to allegedly train on artists' images and copy artists' styles.

For example, they were told that if they "contend Stable Diffusion contains 'compressed copies' of the Training Images, they need to define 'compressed copies' and explain plausible facts in support. And if plaintiffs’ compressed copies theory is based on a contention that Stable Diffusion contains mathematical or statistical methods that can be carried out through algorithms or instructions in order to reconstruct the Training Images in whole or in part to create the new Output Images, they need to clarify that and provide plausible facts in support," Orrick wrote.

To keep their fight alive, the artists pored over academic articles to support their arguments that "Stable Diffusion is built to a significant extent on copyrighted works and that the way the product operates necessarily invokes copies or protected elements of those works." Orrick agreed that their amended complaint made plausible inferences that "at this juncture" is enough to support claims "that Stable Diffusion by operation by end users creates copyright infringement and was created to facilitate that infringement by design."

"Specifically, the Court found Plaintiffs' theory that image-diffusion models like Stable Diffusion contain compressed copies of their datasets to be plausible," Saveri and Butterick's statement to Ars said. "The Court also found it plausible that training, distributing, and copying such models constitute acts of copyright infringement."

Not all of the artists' claims survived, with Orrick granting motions to dismiss claims alleging that AI companies removed content management information from artworks in violation of the Digital Millennium Copyright Act (DMCA). Because artists failed to show evidence of defendants altering or stripping this information, they must permanently drop the DMCA claims.

Part of Orrick's decision on the DMCA claims, however, indicates that the legal basis for dismissal is "unsettled," with Orrick simply agreeing with Stability AI's unsettled argument that "because the output images are admittedly not identical to the Training Images, there can be no liability for any removal of CMI that occurred during the training process."

Ortiz wrote on X that she respectfully disagreed with that part of the decision but expressed enthusiasm that the court allowed artists to proceed with false endorsement claims, alleging that Midjourney violated the Lanham Act.

Five artists successfully argued that because "their names appeared on the list of 4,700 artists posted by Midjourney’s CEO on Discord" and that list was used to promote "the various styles of artistic works its AI product could produce," this plausibly created confusion over whether those artists had endorsed Midjourney.

"Whether or not a reasonably prudent consumer would be confused or misled by the Names List and showcase to conclude that the included artists were endorsing the Midjourney product can be tested at summary judgment," Orrick wrote. "Discovery may show that it is or that it is not."

While Orrick agreed with Midjourney that "plaintiffs have no protection over 'simple, cartoony drawings' or 'gritty fantasy paintings,'" artists were able to advance a "trade dress" claim under the Lanham Act, too. This is because Midjourney allegedly "allows users to create works capturing the 'trade dress of each of the Midjourney Named Plaintiffs [that] is inherently distinctive in look and feel as used in connection with their artwork and art products.'"

As discovery proceeds in the case, artists will also have an opportunity to amend dismissed claims of unjust enrichment. According to Orrick, their next amended complaint will be their last chance to prove that AI companies have "deprived plaintiffs 'the benefit of the value of their works.'"

Saveri and Butterick confirmed that "though the Court dismissed certain supplementary claims, Plaintiffs' central claims will now proceed to discovery and trial." On X, Ortiz suggested that the artists' case is "now potentially one of THE biggest copyright infringement and trade dress cases ever!"

"Looking forward to the next stage of our fight!" Ortiz wrote.

reader comments

Channel ars technica.

• svg]:fill-accent-900">

Master the art of fooling AI detectors (with this other AI tool)

By Stack Commerce

Posted on Aug 15, 2024 9:00 AM EDT

We may earn revenue from the products available on this page and participate in affiliate programs. Learn more ›

Did AI write this article? No—you’d be able to tell. Tools like ChatGPT are notorious for writing text that sounds robotic, repetitive, and just plain awkward. If you’ve ever had it write your emails or essays (we won’t tell), you already know how bad it is.

But you might be able to fool some people with this tool that writes realistic AI text . Ironically, Undetectable Humanizer is still AI, but it’s trained in a different way to make the speech sound far more human. Get lifetime access for $39.99 (reg. $1,080) and never pay recurring fees.

Ever seen AI pass as human? Well, now you can

Gone are the days of editing ChatGPT’s output to make it sound less…bad. Now, you can save time by selecting a humanizer tool, optimizing a few settings, and hitting generate. The algorithms behind Undetectable Humanizer are built to help you get around AI detectors , though it can’t guarantee you won’t get busted by your professors or Google.

Or, maybe you don’t want the tool for any funny business. You could use it to optimize your thesis sentence, write a strong introduction, or just have it edit your essay for readability. We would hardly consider that cheating.

Content creators, small business owners, and SEO specialists will love the other undetectable AI writers optimized for different use cases. There’s one for ranking better on Google, improving readability for your online audience, and boosting conversion rates.

Maybe AI wrote this article after all. You probably wouldn’t be able to tell if we used Undetectable Humanizer for AI text . Give it a try yourself with this $39.99 lifetime subscription (reg. $1,080).

StackSocial prices subject to change.

Advertisement

Supported by

Britain’s Violent Riots: What We Know

Officials had braced for more unrest on Wednesday, but the night’s anti-immigration protests were smaller, with counterprotesters dominating the streets instead.

• Share full article

By Lynsey Chutel

After days of violent rioting set off by disinformation around a deadly stabbing rampage, the authorities in Britain had been bracing for more unrest on Wednesday. But by nightfall, large-scale anti-immigration demonstrations had not materialized, and only a few arrests had been made nationwide.

Instead, streets in cities across the country were filled with thousands of antiracism protesters, including in Liverpool, where by late evening, the counterdemonstration had taken on an almost celebratory tone.

Over the weekend, the anti-immigration protests, organized by far-right groups, had devolved into violence in more than a dozen towns and cities. And with messages on social media calling for wider protests and counterprotests on Wednesday, the British authorities were on high alert.

With tensions running high, Prime Minister Keir Starmer’s cabinet held emergency meetings to discuss what has become the first crisis of his recently elected government. Some 6,000 specialist public-order police officers were mobilized nationwide to respond to any disorder, and the authorities in several cities and towns stepped up patrols.

Wednesday was not trouble-free, however.

In Bristol, the police said there was one arrest after a brick was thrown at a police vehicle and a bottle was thrown. In the southern city of Portsmouth, police officers dispersed a small group of anti-immigration protesters who had blocked a roadway. And in Belfast, Northern Ireland, where there have been at least four nights of unrest, disorder continued, and the police service said it would bring in additional officers.

But overall, many expressed relief that the fears of wide-scale violence had not been realized.

Here’s what we know about the turmoil in Britain.

Where arrests have been reported

We are having trouble retrieving the article content.

Please enable JavaScript in your browser settings.

Thank you for your patience while we verify access. If you are in Reader mode please exit and  log into  your Times account, or  subscribe  for all of The Times.

Thank you for your patience while we verify access.

Already a subscriber?  Log in .

Want all of The Times?  Subscribe .