AI design tools suffer from 'central tendency bias' in user selection

A new study from Huiyang Chen and Keqing Jiao, accepted at the 2026 Human-AI Interaction and Experience Design (HAXD) conference, reveals a systematic bias in how humans select from AI-generated design variations. Drawing on ensemble perception theory, the researchers hypothesized that presenting multiple options simultaneously might induce a 'central tendency bias' — a preference for designs closest to the average of the set. To test this, they ran a controlled experiment with image-generation AI outputs, manipulating the variance of the design set (high vs. low) and measuring selections in both aesthetic preference and representativeness tasks.

The results confirmed their hypothesis: when the set variance was high, participants were significantly more likely to select designs near the center of the distribution. This effect held across both tasks, indicating that the bias is not task-specific. The study highlights a critical tension in human-AI co-creation: while generating diverse outputs is intended to spark creativity, the multi-option interface may inadvertently push users toward safe, average choices. This 'central tendency bias' could lead to homogenized results in tools like DALL-E, Midjourney, or Stable Diffusion, potentially undermining the exploratory benefits of offering many variations. For AI design tools, the finding suggests that interface design — not just model capability — plays a key role in shaping creative outcomes.