Creativity has long been regarded as one of the advanced cognitive abilities unique to humans. Within the realm of creativity research, the perception of creativity in GAI-generated works and individuals’ attitudes to GAI creativity, particularly when these creations are judged to have creative characteristics, are key issues worthy of in-depth exploration. This study focuses on the artistic products generated by GAI, specifically three-line poems and oil paintings, and uses three progressive experiments to investigate individual perceptions and attitudes toward GAI creativity among college students. Additionally, the study explores key factors moderating human evaluations of GAI creations, aiming to provide empirical evidence to support the widespread adoption, evaluation, and societal utility of GAI.

In Experiment 1, a 2 (products’ actual author: human or GAI) × 2 (product quality: high or low) within-subjects design was employed. Participants’ likability of the products was considered as a covariate, and the perceived creativity of the products served as the dependent variable. During the experiment, participants viewed four three-line poems and four oil paintings, all created by humans or GAI. For each type of author, there was 1 high-creativity piece and 1 low-creativity piece. These works were presented in a random order within each type. The actual authorship of the works was hidden, and participants were asked to rate the works on a 7-point Likert scale for the creativity (obtained by calculating the average value of originality and appropriateness scores) and likability based on their subjective impressions. The results showed that the creative score of the three-line poems created by GAI was higher than that of those created by humans, but the main effect of the actual author of oil paintings was not significant. Overall, college students rated GAI’s works as equally creative as human-created works, though this rating was affected by the type and quality of the works.

In Experiment 2, a 2 (speculated author label: human or GAI) × 2 (product quality: high or low) within-subjects design was utilized. The dependent variables consisted of participants’ inferred authorship of the works and their perceptual evaluations (creativity and likability). During the experiment, participants were presented with four three-line poems and eight oil paintings. Among human-created works, there was 1 high-creativity and 1 low-creativity three-line poem, as well as 2 high-creativity and 2 low-creativity oil paintings; the categories and quantities of works created by GAI were consistent with those by humans. All works were displayed in a random order within each category (three-line poems and oil paintings). Without revealing the actual authorship, participants were asked to speculate on the authorship of each piece (Human or GAI) after viewing the works. Subsequently, participants rated the works on a 7-point Likert scale for creativity and likability. The results indicated that participants had moderate accuracy in distinguishing the creators of the three-line poems (54.64% accuracy), whereas their ability to identify the creators of the oil paintings (40.62% accuracy) was below chance level, suggesting an inability to accurately identify authorship. Whether they are three-line poems or oil paintings, the perceived score of the works created by the speculated author for “GAI” are always lower than those of the works created by the speculated author for “Human”. The above results show that for the three-line poems and oil paintings, whether the actual creator is GAI or humans, as long as the participants think the creator is GAI, their creativity ratings are lower. It indicates that people have speculative bias and evaluative bias on the creativity of GAI works.

In Experiment 3, a 2 (products’ actual author: human or GAI) × 2 (exoteric author label: human or GAI) within-subjects design was employed. Participants’ likability of the products was again considered as a covariate, and the perceived creativity of the products was the dependent variable. In this experiment, participants sequentially viewed eight three-line poems and eight oil paintings. The three-line poems and oil paintings were presented randomly within their respective categories. The exoteric author labels were displayed either at the top (for three-line poems) or at the bottom (for oil paintings) of the works. These labels were pseudo-randomized: within each pair of works from the same condition (the types and quality of the works are the same), one was randomly labeled as “Creator: Human/GAI”, while the other received the opposite label (Creator: GAI/Human). The remaining procedure adhered to the steps of Experiment 1. The results show that for both three-line poems and oil paintings, regardless of whether their actual author are GAI or humans, if the exoteric author label is “GAI”, their creativity ratings are lower. This suggests that evaluative bias persists even when authorship is explicitly disclosed. Furthermore, the evaluative bias was moderated by participants’ exposure to AI and their professional background. Specifically, evaluative bias was reduced in participants with higher levels of exposure to AI and those with a background in STEM fields.

In summary, this study highlights the perceptual characteristics of GAI creativity, the biases in college students’ perceptions of GAI-generated works, and the potential factors moderating these biases. The findings offer valuable insights into human-AI interactions and contribute to the advancement of human-computer collaboration and innovative behavior in the AI era.