Meta-analysis finds viewers rate art lower once they know AI made it

The pooled data covered 6 studies (16 effect sizes) for sensory-motor, 27 studies (94 effect sizes) for emotion-valuation and 26 studies (49 effect sizes) for knowledge-meaning. The analysis tested whether bias is built into aesthetic experience or varies with circumstances.

The size of the bias varied with age, art style, image source and situatedness. The authors read the age pattern as suggesting a generational shift in how audiences think about AI in art. Large leftover variation led them to conclude the bias is unlikely to be structural.

The meta-analysis grouped findings by the three systems involved in an aesthetic experience: sensory-motor, emotion-valuation and knowledge-meaning. The bias was weakest where responses are more perceptual and strongest where viewers interpret and assign meaning to the work.

Moderator analyses found that the bias was not uniform. Age, art style, image source and situatedness each changed its magnitude, and the age effect is read by the authors as pointing to a generational shift in audience attitudes.

Because substantial heterogeneity remained after accounting for these moderators, the authors conclude the bias is unlikely to be a fixed, structural feature of how people experience art. They recommend further work on invisible human and machine authorship.