Comparative Designs Reveal Preferences for Human-Generated Rather Than AI-Generated art

Area of Science:

• Artificial Intelligence
• Art Evaluation
• Human-Computer Interaction

Background:

• Widespread access to AI art generators (e.g., DALL-E, Stable Diffusion) has increased interest in evaluating AI-generated art.
• Previous research on preferences for human versus AI art yields contradictory findings, potentially due to differing experimental designs.

Purpose of the Study:

• To investigate how experimental design (comparative vs. non-comparative) influences the evaluation and preference of AI-generated art.
• To determine if traditional art appraisal methods are sensitive enough to detect subtle preferences between human and AI art.

Main Methods:

• Two experiments were conducted: one utilizing a Likert scale (N=250) for non-comparative appraisal.
• A second experiment employed a 2-alternative forced-choice design (N=102) for comparative appraisal.

Main Results:

• Conflicting results emerged between the two experimental designs.
• The forced-choice (comparative) design revealed preferences for human art that the Likert scale (non-comparative) design did not reliably detect.
• Participants showed a preference for human art in liking and valuation appraisals when assessed via comparative methods.

Conclusions:

• Non-comparative appraisal methods, like Likert scales, may lack the sensitivity to reliably detect art preferences.
• Comparative designs better approximate real-world art interaction and are more effective in revealing preferences for human art over AI-generated art.
• As AI art becomes more prevalent, understanding evaluation methodologies is crucial for accurate preference assessment.