A study in Japan found that individuals tend to look longer at paintings when they believe that they were made by humans compared to paintings they believe were AI-generated. There were, however, no differences in subjective evaluations of AI-generated and human-made paintings on average. The paper was published in Perception.
Recent years have witnessed a surge in the utilization of artificial intelligence (AI) tools across various domains previously thought to be exclusive to human expertise. The AI revolution, as some refer to it, is led by generative AI models. Generative AI models are a class of artificial intelligence tools designed to create new content, whether that be text, images, music, or other forms of media, based on the patterns and information they have learned from their training data. The most popular generative AI tools currently include OpenAI’s ChatGPT and DALL-E, Google’s BERT, Bard, and LaMDA, NVIDIA’s StyleGAN, Facebook’s BlenderBot, and others.
One type of generative AI models that are gaining particular popularity are those that generate pictures from textual prompts. AI models that create pictures like OpenAI’s DALL-E, Google’s Imagen, Midjourney, or Stable Diffusion are used by more and more individuals for generating pictures of all kinds. This rapidly increasing popularity of AI art has also fostered an interest in studying people’s attitudes towards it. In general, previous studies indicate that people often have difficulty recognizing AI art, but tend to perceive the AI-generated artwork as worse than human-made art.
Study authors Yizhen Zhou and Hideaki Kawabata wanted to further explore the negative bias toward AI art. They were particularly interested in finding out whether there is an implicit bias towards it. These authors conducted a study in which they tracked how much time people spend looking at AI- and human-made art, but also how they subjectively evaluate it i.e., how they see its beauty, emotional valence, emotional arousal, familiarity, concreteness, and how much they like it.
The study involved 34 undergraduate students from universities in the greater Tokyo area, all of whom lacked experience in art criticism. The group had an average age of 21 years, and 22 were women.
Utilizing 20 landscape paintings from the Vienna Art Picture System dataset and 20 AI-generated paintings created with Disco Diffusion, the research involved three tasks. Initially, participants viewed a series of paintings (both human-made and AI-generated) displayed on a screen for 20 seconds each, followed by a 1-second blank screen, while their eye movements were tracked. Subsequently, they rated each painting on various scales, such as beauty, and attempted to identify whether the artworks were human or AI-created.
Results showed that there was no difference in the average time participants spent looking at AI-generated pictures and at human-created ones — total fixation times, as detected by the eye tracker, were the same for the two types of pictures in the free-viewing tasks. The same was the case for the number of fixations (the number of times eyes looked at a specific place in the picture) and the average duration of a fixation.
In a similar fashion, there were no differences in any of the subjective evaluations between human-made and AI-generated pictures. However, when participants were asked to classify the paintings into AI-made and human-made, they spent more time looking at pictures that they considered to be human-made.
Participants classified 68% of human-made paintings correctly (i.e., classified them as human-made). However, they correctly classified only 43% of AI-generated images.
“Our results indicate an implicit bias toward AI art. Although participants were unable to identify whether the paintings were made by AI and evaluated human- and AI-made paintings equivalently in terms of perceived aesthetic values, they spent more time viewing the paintings they categorized as human-made than AI-generated. This finding suggests that a negative bias toward AI art can be reflected at an implicit level. Although AI is now capable of performing creative tasks typically undertaken by humans, artistic creativity is still considered a human-exclusive ability,” the study authors concluded.
The study sheds light on the way people perceive AI-made artwork. However, the study used a limited set of pictures, all representing landscapes. Additionally, study participants were a small group of Japanese students. Studies using different types of pictures and larger and more diverse groups of participants might not yield equal results.
The paper, “Eyes can tell: Assessment of implicit attitudes toward AI art”, was authored by Yizhen Zhou, Hideaki Kawabata.