Overview
The rise of artificial intelligence (AI) has extended its reach into creative domains, such as fiction writing and art (Du Sautory, 2019). However, the subjective perception of AI-created artworks remains a fascinating and complex topic. While AI algorithms can produce impressive artistic works, the public's perception of these creations as true art might differ from that of human-made art.
Creativity is in the eye of the beholder. But what we consider to be genuinely “art” may be socially limited to human creative endeavors. We're seeing this with ChatGPT and people's reactions to the creative works it produces. While these works may objectively pass as art, the public's subjective perceptions of them seems to differ.
This begets an important question: Do we consider works created by AI as truly art?
The Experiment
To answer this question, we conducted an experiment with 400 participants on Amazon Mechanical Turk. Each participant was shown a painting attributed to either a human artist or an AI algorithm (randomly assigned) and asked to evaluate whether they considered it "art" and how much they liked it.
Participants were told, “The painting below was created by [Alfred Iodice / an AI algorithm].” This painting, illustrated below, was actually created by an AI algorithm and made available courtesy of Wikimedia Commons.
After viewing the painting, participants were asked two questions, “To what extent do you think this painting is art? (1 = Not at all, 7 = Definitely art)” and “How much do you like this painting? (1 = Not at all, 7 = Very much),” both measured via a 1-7 survey scale.
After viewing the painting, participants were asked two questions, “To what extent do you think this painting is art? (1 = Not at all, 7 = Definitely art)” and “How much do you like this painting? (1 = Not at all, 7 = Very much),” both measured via a 1-7 survey scale.
Results
Our analysis revealed that the painting was significantly less likely to be considered art when attributed to an AI algorithm (avg. = 5.26) rather than a human (avg. = 5.91), (p < 0.001). Additionally, participants liked the art less when it was attributed to AI (avg. = 4.76) than a human (avg. = 5.10), (p = 0.023).
Importantly, the results differed based on age. Younger participants were more accepting of the AI-attributed painting as art compared to their older counterparts. For every additional year of age, participants’ "is art" ratings were 0.03 points lower (1-7 scale) for the AI-attributed painting than the human-attributed one (p = 0.027). For example, we’d expect a 20-year-old’s “is art” rating to only differ by about 0.12 points, whereas a 60-year-old’s rating would differ by about 1.33 points.
The age difference was only marginally significant for how much participants like the painting (p = 0.088), but there was still an additional 0.024-point difference for each additional year of age. Whereas younger individuals seemed to like the painting about the same regardless of its authorship attribution, older participants seemed to like the painting less when it was attributed to an AI algorithm. As an example, we’d expect 20-year-old’s and 60-year-old’s painting likability difference to be +0.09 and -0.89, respectively, using our 1-7 scale.
Conclusion
In the realm of art perceptions, there seems to be a bias against AI-generated art, with people viewing it as inferior to human-made art. However, such perceptions are wont to change over time, especially in the tech space. Younger generations appear more accepting of AI art, hinting at a potential shift in attitudes towards creative works authored by artificial intelligence.
References
Du Sautoy, M. (2019). The Creativity Code: Art and Innovation in the Age of AI. Harvard University Press.
Methods Note
We used an independent samples t-test to test for significant differences in perceptions between the AI-and Human-attributed art conditions. For significant differences, the difference between the two groups' averages would be large and the corresponding “p-value” would be small. A p-value below 0.05 indicates statistical significance. We used OLS regression analyses with interaction terms to test whether the results differed by age.
The data and survey materials used for this study are available upon request.