Artificial Intelligence (AI) has become ubiquitous in our everyday lives. With each iteration in technological capabilities, the gap between AI and human ability seems to narrow. One such advancement has been the recent wave of image-synthesis models; AI image-generation tools that have evolved to a level of sophistication such that it is nearly impossible to distinguish between photographs of real human faces and those generated by a computer (Moshel et al., 2022; Nightingale and Farid, 2022). The fact that AI is able to fool the human visual system’s perception of faces—one of our most deeply-rooted and evolutionarily relevant brain functions—is certainly cause for concern (Westerlund, 2019), but how does it fare against what has arguably been the cultural benchmark of human creativity throughout history: Art?

Addressing this question has increasing importance for understanding the changing landscape of the art world and the role of technology in shaping artistic production and consumption. Separately from concerns that centre on questions of intellectual property and privacy violations (Smits and Borghuis, 2022; Zhou and Nabus, 2023), AI-generated art also raises fundamental questions about how we might (re)define creativity (Epstein et al., 2023; Oksanen et al., 2023), an ability considered until now to be essentially human-specific. This issue has captured interdisciplinary interest, with researchers across computer science, philosophy, and psychological and social sciences joining forces to investigate how we process and interact with the AI-generated works that we are increasingly exposed to in advertising, social media, and scams (Baidoo-Anu and Ansah, 2023; Epstein et al., 2023; Millet et al., 2023; Oksanen et al., 2023; Ooi et al., 2023).

Previous research has shown a consistent trend of negative bias towards AI-generated art (Agudo et al., 2022; Chamberlain et al., 2018; Chiarella et al., 2022; Gangadharbatla, 2022; Horton et al., 2023; Millet et al., 2023; Ragot et al., 2020), however, this research has primarily examined the impact of knowing whether the art was created by AI or humans (rather than assessing artworks’ intrinsic aesthetic qualities). Whether or not humans exhibit a reliable preference for the intrinsic aesthetic qualities of (unlabelled) human- or AI-generated art has received comparatively less attention. Therefore, there is lack of understanding in current literature of how people respond to AI art separately from biases related to authorship attribution, which will be increasingly important as AI-generated art continues to proliferate, and contextual information such as authorship attribution becomes more scarce.

Here, we aimed to address this critical gap in the literature by examining observer preferences for human-made vs. AI-generated art in the absence of any knowledge of the artworks’ origins. The novelty of this study also resides in providing an objective and quantitative assessment of the human perception of artificial art made using OpenAI’s DALL·E 2, one of the most advanced AI tools for art generation.¹ With high-level performance for representational image generation, DALL·E 2 represents a step change in the field, as technology and synthetic representational artworks publicly accessible before its release were often much more rudimentary in complexity of composition and general verisimilitude (Pan et al., 2019). Here we focus on how observers perceive and appreciate the aesthetic qualities of DALL·E 2 artworks by comparing human observers’ appreciation of unlabelled AI-generated and human-made art and testing observers’ ability to distinguish between the two.

To assess human observers’ appreciation and discrimination of non-abstract artworks generated by both humans and AI, we paired 50 lesser-known real artworks by famous representational artists (absent in the WikiArt’s “famous work” section) with 50 artificial artworks generated in a similar style using OpenAI’s DALL·E 2 with comparable visual features (e.g., colour, style, composition), as depicted in Figure 1. Online observers viewed these image pairs in either a preference judgement task (Experiment 1, “Which artwork do you like the most?”, 127 participants) or a real-artificial discrimination task (Experiment 2, “Which artwork was generated by a computer?”, 137 participants) (see Figure 2A). To minimise bias and conceal the true purpose of each experiment (i.e., comparing human- vs. AI-generated), in both cases the 50 matched image pairs appeared randomly intermingled with random pairs (e.g., Human-made vs. Human-made) drawn from the full image set. We also assessed online observers’ art experience to test its potential influence on the appreciation and discrimination of these images.

Experiment 1 revealed a significant preference for AI-generated artworks. Without being provided with any information about the origin/authorship of the artworks, participants in the preference task indicated they preferred the AI-generated artworks significantly more often than the human-made artworks (AI-preference scores significantly above 50% chance-level) [t(126) = 5.39, p < 0.001, d = 0.48; Figure 2B].

Interestingly, when a separate group of participants in Experiment 2 were asked to detect which one of the two artworks was made by a computer, they could do so significantly better than chance. A one-sample t-test indicated AI detection accuracy was significantly above 50% [t(136) = 3.62, p < 0.001, d = 0.31] although a smaller effect size was observed compared to Experiment 1 (Figure 2B).

Image-wise correlational analysis revealed a positive relationship between AI preference and AI detection accuracy scores associated with each pair [t(48) = 3.23, p = 0.002, r = 0.42]. As shown in Figure 3, the AI-generated artworks that Experiment 1 participants tended to prefer were also those that Experiment 2 participants were better able to detect, suggesting there may be features in the artworks driving both preference for and detection of AI-generated art. Importantly, this correlation remains almost identical when controlling for potential differences in mean luminance, root mean square contrast and entropy between human-made and AI-generated stimuli [t(48) = 3.15, p = 0.003, r = 0.42] suggesting that there are other more complex visual features influencing participants’ preference and discrimination.

Regarding the influence of expertise, our results indicated no significant correlations between participants’ experience in art (i.e., interest and knowledge in art, see methods for further details on how this was assessed) and AI preference and detection accuracy in either experiment [r(125) = 0.02, p = 0.79, and r(135) = 0.16, p = 0.06 for Experiments 1 and 2, respectively].

There is a burgeoning sentiment that AI-image generation technology has reached a point of refinement that challenges our traditional understanding of the human perception and appreciation of art (Epstein et al., 2023; Oksanen et al., 2023). Our results evidence this claim, revealing that in the absence of attribution labels, human observers systematically prefer AI-generated artworks over stylistically similar artworks painted by real people. Should this effect prove universal, it could constitute a paradigm shift in art appreciation, favouring synthetic works over those created by human artists, which has the potential to transform the art world, while also raising new questions about authorship, authenticity, and the role of human creativity in the age of generative AI.

Our findings stand in contrast to prior research on subjective evaluations of computer-generated artwork, which have largely reported a negative bias towards AI art (Agudo et al., 2022; Caporusso et al., 2020; Chiarella et al., 2022; Gangadharbatla, 2022; Kirk et al., 2009). This work has primarily examined the role of authorship attribution in AI art perception, rather than the aesthetic value of the artworks themselves. Thus, the observed negative bias in these studies appears to relate to our explicit prejudice against artificially-generated content (i.e., if an artwork is labelled as computer-generated, we tend not to like it). In contrast, here we obtained observer preference decisions in the absence of any authorship label—a neutral presentation format that encouraged observers to judge the inherent aesthetic qualities of the artworks—and assessed authorship discrimination in a separate experiment. This approach allowed us to obtain a quantitative assessment of the degree to which observers prefer the inherent artistic qualities of real artworks made by humans vs. those created by new AI image-generation models, free from external biases.

Although observers in the first experiment consistently preferred artworks generated by DALL·E 2 over those made by human artists, it was not the case that these AI artworks were indistinguishable from human creations. In Experiment 2, a separate group of observers were asked to explicitly judge which of the two artworks in each pair was generated by a computer. We found they could reliably do so above chance-level. Moreover, there was a positive correlation between the image-pairs’ AI-preference and AI-detection scores, suggesting that the same visual features that made the AI-generated artworks more detectable to participants in Experiment 2 also made those artworks more appealing to participants in Experiment 1 (Van Geert and Wagemans, 2020). This intriguing pattern underscores the role that explicit bias against artificial creations has likely played in prior investigations (Agudo et al., 2022; Caporusso et al., 2020; Chiarella et al., 2022; Gangadharbatla, 2022; Kirk et al., 2009) of the aesthetic appeal of AI-generated artworks: When participants do not know the artworks are computer-generated, they freely prefer them. Interestingly, we found no evidence that these effects were moderated by observers’ art expertise, suggesting that the features in question are broadly accessible. However, our results suggest that they are not related to luminance, contrast and entropy, and that other visual features might have a more important role; a possibility which future research will no doubt explore in detail. This is an important research avenue together with examining the influence of contextual and authorship information on these effects to better understand the human perception of AI-generated artworks.

More generally, these results suggest that the technology behind DALL·E 2, in striving for stronger verisimilitude in computer-generated art, has evolved to do so by extrapolating (or exploiting) existing known biases in human cognition. On this thinking, DALL·E’s capacity to produce works that observers tend to prefer over human artworks could possibly be explained by the fact that its training dataset comprises images of artworks that are broadly popular and likely considered to be aesthetically pleasing. This is in line with recent research on “deepfakes,” wherein AI-generated faces not only fool observers with their hyper-realistic nature, but are also associated with enhanced perceptions of trustworthiness (Boháček and Farid, 2022; Groh et al., 2022; Nightingale and Farid, 2022). These findings raise critical concerns about the exact nature of the cognitive processes that could be targeted and manipulated using generative AI, and therefore, about its large-scale deployment without detailed investigation.

In a world increasingly shaped by the algorithms around us, the current findings suggest that as AI continues to evolve towards human capabilities, it may be poised to redefine our understanding of creative expression altogether. If AI-generated content can reach or even surpass aesthetic equivalence with human creation, the question of whether something can truly be considered “art” if it has no human architect becomes more complicated. Although AI image diffusion models currently still depend on some level of human intervention to produce artwork (from the initial production of training data to programming to user input), it seems increasingly likely that this necessity will decrease in future as the technologies progress. Our results are an initial step towards untangling the complex interaction between generative AI and human aesthetic preference; clearly, systematic examinations of AI-generated artworks’ features are needed to fully understand the mechanisms and implications of AI preferences. A clear future avenue for this field will be to combine parametric feature variations in AI-artwork with systematic manipulations of artwork-attribution (AI or human), to determine how low and mid-level visual featural differences interact with the documented negative bias associated with AI attribution (Chamberlain et al., 2018; Chiarella et al., 2022; Horton et al., 2023; Millet et al., 2023; Ragot et al., 2020). Indeed, this bias may well be shifting with the rapid adoption of AI technology in the wider population (Epstein et al., 2023; Oksanen et al., 2023; Prasad Agrawal, 2023).

In addition, it will be important to consider additional presentation parameters that may contribute to or modulate observer preferences: For example, here we presented artworks at a relatively small size, cropped from their original dimensions to conform to a 1:1 ratio. Since these aspects necessarily change the composition (and therefore appreciation) of the artwork, it will be necessary to explore whether our results hold under different presentation formats, as well as across real-world and digital presentation formats. Digital presentation might have favoured AI-made works. Some critical features in human-made works such as texture might have been lost with their digital representation. Furthermore, while specific efforts have been made to account for selection bias using image statistics, our design required some degree of curation to omit clear outliers produced by DALL-E 2 that did not match the criteria laid out in advance. Critically, this means that our results do not automatically apply to the state of AI-image generation as a whole, but rather illustrate the height of its capabilities. In juxtaposing samples of the acme of traditional art with that which AI is capable of producing, we get a better view of the perception of AI as it continues to progress. Future studies should expand on this further by including a broader selection of images and removing the potential for selection bias through human intervention entirely. Importantly, excluding famous human-made artworks might have favoured AI artworks. Including famous human-made artworks assuming they could have been tested on completely naïve participants with no previous exposure could have increased preference for human-made works. There might be features in these artworks resulting in higher preference independently of how famous they are.

Moreover, future research could investigate the effects of expertise further as this was only succinctly assessed in this online experiment across three generic questions. Variations in the type of expertise and the number of years could be more deeply examined. Demographics such as age and cultural background could also be explored. They might influence how participants perceive and appreciate these artworks and the features they might focus on. It can also be noted that the AI-preference we report here is, by design, derived from a constrained set of artistic styles. As such, future work expanding this line of investigation could look to systematic comparisons of different styles. These are necessary steps for future research to reach a broader understanding of the perception of digital artworks in the general population. This has implications for the art market, museums, curators, and future artists, with new insights into the type of artworks, in the digital space in particular, people might prefer. It might also be important to extend this research in future studies to other creative domains. The production of movies and music, for instance, are all changing quickly with generative AI. Understanding similarities and differences across domains, could help identify universal mechanisms in this new generation of AI models biassing human perception and preference.

To conclude, our findings suggest that we might be observing with recent advances in AI technology a shift in art preference to favour synthetic creations, raising critical questions about the way we think about art and its value to our society. As the field of generative-AI continues to accelerate—spurring equal parts concern and excitement—there can be no doubt as to the urgency in this challenge. DALL·E 2 will soon be superseded by the next generation of algorithms with as-yet unknown capabilities. Understanding how the human experience intersects with this technology will be critical to ensuring its positive impact on our society.