Research Data for Human Interpretative Variability and Semantic Ambiguity in Text-to-Image Systems

This dataset contains the tabular research data supporting a human-centered evaluation of semantic interpretation in visual outputs generated by text-to-image systems from boundary-safe prompts. The experimental corpus comprised 42 evaluable images generated with Stable Diffusion XL, DALL-E 3, and Midjourney across three semantic conditions: Concealed Eroticism (EC), Implicit Discrimination (DI), and Symbolic Violence (VS). The deposited data include an anonymized matrix of 420 categorical judgments produced by 10 human evaluators, the experimental image inventory, an inventory of 15 operational prompts with traceability information and SHA-256 hashes where available, and per-image OpenCLIP semantic assessment results. A data dictionary, integrity manifest, and documentary contact-sheet overview of the visual corpus are also provided. The original generated-image corpus is retained by the authors and is not included as individual image files in this public dataset. The contact-sheet overview is provided solely as documentary material and must not be considered a substitute for the original image dataset. Access to the original generated images may be considered upon reasonable request to the corresponding author, subject to applicable platform terms, redistribution rights, and ethical considerations. Human judgments and OpenCLIP outputs represent distinct sources of semantic evidence and should not be interpreted as a single ground-truth annotation.