Stanford University


The Investigators of Human Creativity

A team of researchers set out to examine how AI is disrupting the creative industry

Tim Hatton
Contributing Writer

September 25, 2023
7-minute read

Show buttons

Generative artificial intelligence tools like ChatGPT and Stable Diffusion have soared in popularity over the past several months—and so have the challenges surrounding them,  including issues of authorship and credit, concerns about misinformation and the role AI could have in the media ecosystem, and ultimately the impact the technology will have on skills and jobs. 

Because no matter how magical, mysterious, or automatic generative AI may seem, the content didn’t come from nowhere, and it didn’t come from the AI itself. Text-based tools are trained on the writing of human authors while image-generation software sources photos, paintings, and other illustrations made by people. At the same time, humans are ultimately responsible for how these AI tools are used and to what ends they’re applied.

In other words, the science of generative AI is similar to the science of human creativity. Now there’s a team of interdisciplinary researchers working to study how the two sciences interact. The collaboration recently spawned a new paper published in Science titled “Art and the Science of Generative AI.” The researchers, who dub themselves “The Investigators of Human Creativity,” have also produced an extended whitepaper and website (that shares the same title as the Science paper) devoted to explorations of the relationship between art and AI.

“With such a big and complex topic with so many promises and perils, like AI has, everyone has their own perspective that’s founded in their own discipline—but how can we build bridges and try to understand this together?” said Ziv Epstein, the lead author of the papers and a postdoctoral scholar at the Stanford Digital Economy Lab. “In presenting a snapshot of where we are right now, we want to toe the line between optimism and pessimism, and we want to build coalitions of people who are thinking about this so that we can pool our resources and create the space for really valuable conversations.”

“If you can get labor economists, AI experts, and artists to all look at a complex issue like AI in the same way, you’re going to have a pretty good picture of what is possible to know at this point.”

Sandy Pentland

Sandy Pentland

Stanford Visiting Scholar
Director, MIT Human Dynamics Laboratory, MIT Media Lab Entrepreneurship Program

The project spans several disciplines and features fourteen experts, including legal scholars, computer scientists, designers, and other leaders studying how AI is impacting their various fields. Epstein is one of several Stanford Digital Economy Lab researchers involved in the project. 

Morgan Frank, a Stanford Digital Fellow, led the section on the labor economics of the creative industry and the impact new AI tools will have on the field. “Our main goal was to touch on these different economic, cultural, and legal aspects of generative AI, and highlight where we think we’re going,” Frank said. “We were thinking about policymakers, the media, and the public, and we were trying to highlight what is known, and what is still yet to be known.”

The idea for the initiative arose last year at the International Conference on Computational Creativity in Bolzano, Italy. Aaron Hertzmann, who ultimately became one of the project’s contributors, spoke about the challenge of anthropomorphizing AI, making the case that the AI tools themselves produce no art—they can only put out reinterpretations of the training data that were inputted by humans. In conversations after that presentation, Epstein and others discussed how Hertzmann’s concern about misinterpretation and agency spoke to a larger issue: since the field of generative AI was developing so fast, how could experts and the public alike wrap their heads around what’s happening? These tools have enormous potential, but what are the immediate questions about how they are being created and deployed?

In the months that followed, those conversations coalesced along a few main themes: aesthetics and culture, legality and copyright, labor economics, and the media ecosystem and misinformation. The Art and the Science of Generative AI project presents a survey of those key themes, allowing for an introduction to these complex issues in hopes of creating a shared understanding going forward. Since the AI landscape touches so many industries and areas of expertise, it requires a common language to enable computer scientists to talk to artists, and for labor economists to talk to copyright lawyers. That’s what these papers seek to provide.

“This whole project was so satisfying to me because there was a consensus and a synergy to the way all of these experts were approaching the subject,” said Sandy Pentland, a visiting scholar at Stanford and one of the project’s contributors. “If you can get labor economists, AI experts, and artists to all look at a complex issue like AI in the same way, you’re going to have a pretty good picture of what is possible to know at this point.”

This collaborative approach provides more than a shared understanding of where the world of generative AI sits currently—it also creates clarity in thinking about what to study in the future. For each of the four key themes—culture, misinformation, law, and the job market—the Art and the Science of Generative AI project offers a snapshot of current concerns and trends, as well as a list of potential questions for researchers to pursue.

Theme #1: Aesthetics and Culture

Thinking about the visual quality of AI art, one possibility is that AI outputs would be trained on other AI outputs, which could create a cycle where the same style of content is produced again and again. 

Recommendation algorithms, such as those that select what you see on social media feeds, add more nuance to that concern. If online creators are generating the same type of information or imagery to please an internet algorithm, it’s also likely to give us the same output repeatedly.

More worryingly, if an increasing amount of content is drawn from biased AI training data, then the technology may capture, reflect, and even amplify cultural norms and biases, potentially reducing diversity in art and design. On the other hand, using an AI tool presents a lower barrier to entry than creating similar images manually, and this may increase diversity by allowing more people to participate in creative processes. A diverse set of outputs from AI tools could also keep the art from becoming boring and potentially do a better job of reflecting human diversity. 

These and other concerns suggest that future research is needed to measure diversity in AI output in order to curb homogeneity.  

Theme #2: Media and Misinformation

The issue of recommendation also impacts the relationship between generative AI and the broader media landscape, including misinformation. Generative tools use far less resources than conventional media and lack many of the guardrails that consumers are used to. And production time and costs decrease, concerns about AI-generated misinformation, synthetic media, fraud, and nonconsensual imagery have risen. Deepfakes—videos edited using falsified images—are a prime example of this phenomenon. 

Many relevant research questions naturally follow: How should platforms intervene to stop the spread of misinformation? How can the provenance of a piece of content be traced to ensure its reliability? And as the quality of AI-generated imagery improves, what can be done to detect that content in the first place so that audiences can understand the kind of media they’re consuming?

Theme #3: Legality and Copyright

The use of training data poses perhaps the most prominent legal issue of generative AI and art since existing copyright laws are insufficient when dealing with AI-generated content. For example, there is no legislation currently on the books that outlines compensation for the training data that AI tools consume. More research and legislation will need to establish whether AI copies what it’s taught or creates something new. These guidelines would also determine how to protect an artist’s style, keep their work protected, and compensate them appropriately. 

The researchers also present a possible alternative: artists could opt-out of having their work used as training data for AI.

“The streaming era has also changed the entire music industry, and that’s what we’re worried about with AI,” Pentland said. “You get a few megastars and then lots of people making pennies. So that’s a good thing for some people, especially the streaming companies, but it’s unequal in a way that is not very positive. That’s what this article was about—there’s a lot that we’re excited about, but there are also threats. We want to be clear about what’s good and what’s bad.”

Theme #4: The Future of Work

Finally, and perhaps most practically, generative AI is set to transform work and employment, especially in creative roles. For a while now, economists have assumed that automation would not affect creative work the same way it has disrupted fields like manufacturing. 

But generative AI has upended this assumption. When it comes to creative production, AI is remarkably efficient, which may lead to the same output with fewer workers. This could threaten certain creative occupations, such as writers, composers, and graphic designers—or at the very least, restructure how that work is done.

“We know that technology rarely automates whole occupations, and it’s much more common that technology automates specific tasks within a job,” Frank said. “People might be tempted to look at tools like MidJourney and assume that automation will make graphic designers a thing of the past, and I don’t think that’s likely true. Workers usually have to perform many tasks within their job, and it’s very rare to see technology replace all of them at once.”

‘The real work begins’

The four themes addressed by the project are astoundingly complex, especially given how fast generative AI tools are being developed and released—and the authors know their work can’t answer all the questions that AI raises. However, by providing a common language and a shared understanding of the current moment, Art and the Science of Generative AI can lead the way for those questions to be answered.

“This is very much the first step,” Epstein said. “There’s nothing empiric here, there’s no engineering, we’re just trying to build a common foundation for future science and research to sit on top of. And with this paper out there, now the real work begins.”

Stanford University