Dataset and Analytical Codebook for: The Algorithmic Gaze and the Conceptual Gravity Well

This dataset contains the quantitative and qualitative coding for N=1,192 AI-generated images, produced to audit aesthetic colonialism and symbolic annihilation in generative AI models. The data was collected using the CAST (Context, Agency, Stereotype, Trope) analytical framework. The dataset compares the ideological output of three distinct AI architectures: Corporate (DALL-E 3), Open-Source (Stable Diffusion Base), and Consumer Aesthetic (Aesthetic Tuned Models). Images were generated across a 21-prompt matrix designed to test foundational biases, cultural stereotyping, socio-political erasure, and linguistic fragility (using English, Hindi, and Malayalam prompts). The dataset includes 9 quantitative variables (e.g., Phenotype Score, Stereotype Index, Agency/Gaze, Aesthetic Trope) and qualitative heuristic notes tracking algorithmic erasure and coherence failures. An Inter-Rater Reliability (IRR) of Cohen’s Kappa = 0.876 was achieved for this dataset. The accompanying Codebook.pdf provides the strictly operationalised rubrics used for all variable scoring.