"Science" Magazine: Change is not "the end of art", generative AI will reshape contemporary media aesthetics

By: Ziv Epstein (MIT), Aaron Hertzmann (Adobe Research), The Investigators Of Human Creativity (Adobe)

Source: Science

Image source: Generated by Unbounded AI tool

Understanding the transformation of creative work will help guide the impact of AI on the media ecosystem.

Generative artificial intelligence (AI) is a hotly debated topic. A prominent application to date is the production of high-quality artistic media for visual art, concept art, music and literature, as well as video and animation. For example, diffusion models can synthesize high-quality images (1), and large language models (LLMs) can produce plausible-sounding, impressive prose and poetry in a wide range of contexts (2). The generative capabilities of these tools could fundamentally change the creative process by which creators form ideas and bring them into production. As creativity is reimagined, many areas of society may also be reimagined. Understanding the impact of generative AI, and making policy decisions around it, will require new interdisciplinary scientific investigations of culture, economics, law, algorithms, and the interaction of technology and creativity.

Moments of change did not signal the 'end of art' but had more complex effects, reshaping the roles and practices of creators and changing the aesthetics of contemporary media (3). For example, some 19th-century artists saw the advent of photography as a threat to painting. Photography did not replace painting, however, but eventually liberated it from realism, giving rise to the Impressionism and modern art movements. Portrait photography, by contrast, did largely replace portraiture. Likewise, the digitization of music production (for example, digital sampling and sound synthesis) has been condemned as "the end of music". But in reality, it changed the way people make and listen to music, and helped spawn new genres, including hip-hop and bass drum. Like these historical parallels, generative AI is not a harbinger of the death of art, but a new medium with its own unique capabilities. As a set of tools used by human creators, generative AI is positioned to disrupt many areas of the creative industry and threaten existing models of work and labor in the short term, while ultimately enabling new models of creative labor and reconfiguring the media ecosystem system.

Unlike past disruptions, however, generative AI relies on training data that people do. These models "learn" generative art by extracting statistical patterns from existing art media. And this reliance raises new questions -- such as where the data comes from, how it affects the output, and how authorship is determined. By leveraging existing work to automate the creative process, generative AI challenges traditional definitions of authorship, ownership, creative inspiration, sampling, and remixing, thereby complicating existing notions of media production. It is therefore important to consider the aesthetic and cultural impact of generative AI, legal issues of ownership and credit, the future of creative work, and implications for contemporary media ecosystems. Among these topics, there are some key research questions that could inform policy and the beneficial use of this technology (4).

About "artificial intelligence"

In order to properly examine these topics, it is first necessary to understand how the language used to describe AI affects perceptions of the technology. The term "artificial intelligence" can be misleading, suggesting that these systems exhibit human-like intentions, agency, and even self-awareness. Natural language-based interfaces for generative AI models, including chat interfaces that use "me," may give users a human-like feel for interacting with them. These perceptions can undermine the credibility of creators whose labor underpins the output of systems (5), and shift responsibility from developers and policymakers when these systems cause harm (6). Future work is needed to understand how perceptions of generative processes influence attitudes toward output and authors. This will help in the design of systems that disclose the generative process and avoid misleading interpretations.

Generative AI and Aesthetics

The special capabilities of generative AI, in turn, generate new aesthetics that could have long-term effects on art and culture. As these tools proliferate, and their use becomes ubiquitous (as photography did a century ago), it remains an open question how the aesthetics they produce will affect artistic output. A low barrier to entry for generative AI could increase the overall diversity of artistic output by expanding the pool of creators involved in artistic practice. At the same time, aesthetic and cultural norms and biases embedded in training data may be captured, reflected, and even amplified, thereby reducing diversity (7). AI-generated content may also provide fodder for future models, creating a self-referential aesthetic flywheel that perpetuates AI-driven cultural norms. Future research should explore ways to quantify and increase output diversity, and examine how generative AI tools affect aesthetics and aesthetic diversity.

The opaque, engagement-maximizing recommendation algorithms of social media platforms can further enforce aesthetic norms through feedback loops (8), producing sensational, shareable content. This could further homogenize content as algorithms and content creators try to maximize engagement. However, some preliminary experiments (9) suggest that incorporating engagement metrics when curating AI-generated content can, in some cases, diversify the content. It remains an open question as to which styles are amplified by recommendation algorithms, and how this prioritization affects the types of content creators produce and share. Future work must explore the complex, dynamic systems formed by the interactions between generative models, recommendation algorithms, and social media platforms, and their impact on aesthetic and conceptual diversity.

Generative AI and Copyright

Generative AI's reliance on training data to automate creation also presents legal and ethical challenges that prompt technical research into the nature of these systems. Copyright law must balance the interests of creators, users of generative AI tools, and society at large. The law can treat the use of training data as non-infringing if the protected work has not been directly copied; as fair use if the training involves a substantial transformation of the underlying data; and only if the creator gives explicit permission Allow use; or, where the creator is compensated, a statutory mandatory license that allows the data to be used for training. Much of copyright law relies on judicial interpretation, so it's unclear whether collecting third-party data for training or imitating an artist's style would violate copyright. Legal and technical issues are entangled: does the model directly replicate elements in the training data, or produce something entirely new? Even if the model doesn't directly reproduce existing work, it's unclear whether and how an artist's personal style should be protected. What mechanisms would protect and compensate artists whose work was used for training, or even allow them to opt out, while still allowing new cultural contributions to be made with generative AI models? Answering these questions and determining how copyright law should treat training data will require substantial technical research to develop and understand AI systems, social science research to understand perceptions of similarity, and legal research to apply existing precedents to new ones. technology. Of course, these views represent only the legal views of the United States.

An obvious legal question is who can claim ownership of the output of the model. Answering this question requires understanding the creative contributions of users of the system and other stakeholders, such as the developers of the system and the creators of the training data. AI developers can claim ownership of output through terms of use. In contrast, users of the system may be considered default copyright holders if they participate in meaningful creative ways (for example, the process is not fully automated, or a particular work is not parodied). But to what extent does a user's creative influence warrant claiming ownership? These questions involve studying the creative process using AI-based tools, which may become more complex if users are given more direct control.

Generative AI and Creative Careers

Regardless of the legal outcome, generative AI tools have the potential to transform creative work and employment. Popular economic theory [i.e. skills-based technological change (SBTC)] posits that cognitive and creative workers face less labor disruption from automation because creativity is not easily codified into specific rules (i.e. Polish Ni's paradox) (10). However, the new tools have raised employment concerns for creative occupations such as composers, graphic designers and writers. This conflict arises because SBTC fails to differentiate cognitive activities such as analytical work from creative ideation. We need a new framework to describe the specific steps of the creative process, which of these steps may be influenced by generative AI tools, and the workplace requirements and activities of different cognitive occupations (11).

While these tools may threaten some professions, they can increase the productivity of others and perhaps create new ones. For example, music automation technology has historically enabled more musicians to create, even with income skewed (12). Generative AI systems can create hundreds of outputs per minute, potentially accelerating the creative process through rapid ideation. However, this acceleration can also disrupt aspects of creativity, as it eliminates the design period of shaping an initial prototype from scratch. In either case, production time and costs are likely to drop. The production of creative products may become more efficient, achieving the same output with less labor. In turn, the demand for creative work may increase. In addition, many employment occupations that use traditional tools, such as illustration or stock photography, may be displaced. Some historical examples bear this out. Most notably, the Industrial Revolution enabled traditional crafts such as ceramics, textiles, and steelmaking to be mass-produced by non-artisan labor; handcrafted goods became exceptional items. Likewise, photography replaced portraiture. The digitization of music removes the constraints of learning to physically operate an instrument, enabling more contributors to more complex arrangements. These tools could change who can be an artist, in which case employment of artists could rise even as average wages fall.

Generative AI and Media Ecology

As these tools impact creative labor, they also pose potential downstream harms to the wider media ecosystem. As the cost and time of producing media at scale decreases, the media ecosystem may become vulnerable to AI-generated misinformation through the creation of synthetic media, especially media that provide evidence for claims (13). These new possibilities for generating realistic synthetic media may undermine trust in truth-captured media through the so-called “liar’s dividend” (false content benefits liars by undermining trust in truth) (14) and increase fraud and Threats of non-consensual sexual images. This raises important research questions: what is the role of platform interventions, such as tracking provenance and detecting downstream synthetic media, in terms of governance and building trust (15)? And how does the proliferation of synthetic media, such as unedited news photos, affect trust in real media? As content production increases, collective attention may decrease (16). The explosion of AI-generated content, in turn, could hamper society’s ability to collectively discuss and act on important areas like climate and democracy.

Every art medium reflects and comments on the issues of its time, and the debate around contemporary AI-generated art mirrors current issues around automation, corporate control, and the attention economy. Ultimately, we express our humanity through art, so understanding and shaping AI's impact on creative expression is central to broader questions about its impact on society. New research into generative AI should inform policy and the beneficial uses of the technology, while engaging key stakeholders, especially artists and creative workers themselves, many of whom are at the vanguard of actively engaging in solving hard problems for social change .

Translator's Note: There are 16 annotations in the text, for related reading, please refer to the original text