The Algorithmic Gaze of Image Quality Assessment: An Audit and Trace Ethnography of the LAION-Aesthetics Predictor

Researchers audited LAION-Aesthetics Predictor (LAP), an algorithmic model widely used to filter training datasets for visual generative AI systems like Stable Diffusion. The audit reveals LAP systematically biases toward images of women while filtering out men and LGBTQ+ individuals, and reinforces Western artistic preferences, raising critical questions about whose aesthetic values shape AI-generated imagery.

The LAION-Aesthetics Predictor represents a fundamental infrastructure layer in modern generative AI, yet operates as a black box with undisclosed aesthetic criteria. By scoring 1.2 billion images during dataset curation, LAP has become the de facto arbiter of visual quality for models affecting billions of users globally. The audit's findings expose how algorithmic systems can mathematize subjective cultural preferences, embedding specific aesthetic hierarchies into AI systems at scale.

This research extends longstanding critiques of AI bias beyond demographics into aesthetic representation. The model's preference for Western realistic art and disproportionate filtering based on gender reveals how technical choices made during development propagate through downstream applications. The researchers traced these biases to LAP's training data sourced primarily from English-speaking photographers and Western enthusiasts, demonstrating how developer populations directly influence algorithmic outputs.

For the generative AI industry, this audit signals potential reputational and functional risks. Companies relying on LAP for dataset curation may inadvertently create systems that produce biased outputs reflecting embedded Western male perspectives. This threatens inclusivity claims and exposes developers to criticism from artists and communities underrepresented in training data. The findings also complicate efforts to develop "objective" quality metrics in AI, suggesting aesthetic evaluation fundamentally requires pluralistic approaches rather than singular algorithmic measures.

The research emphasizes that infrastructure decisions in AI development, often invisible to end users, carry downstream consequences. Future generative AI systems may require transparent aesthetic criteria and diverse input for dataset curation to avoid perpetuating representational harms at scale.

• →LAION-Aesthetics Predictor filters datasets based on hidden aesthetic biases, disproportionately removing images of men and LGBTQ+ individuals while over-representing women
• →The model systematically rates Western and Japanese realistic art highest, reinforcing imperial and male-centered art historical perspectives in AI training
• →LAP's biases originate from its training data sourced primarily from Western English-speaking photographers, demonstrating how developer demographics shape algorithmic outputs
• →Aesthetic evaluation in AI should shift toward pluralistic approaches rather than one-size-fits-all measures to prevent representational harms
• →This infrastructure-level bias affects billions of users of generative AI systems like Stable Diffusion trained on LAP-filtered datasets