Grok Image Generation Safety Audit: Incident Dataset

Structured dataset documenting 43 public incidents of targeted sexualization involving Grok’s image generation feature on X (formerly Twitter), observed between December 31, 2025 and January 7, 2026. Incidents were identified through purposive sampling of public @grok threads; the researcher did not generate or request any of the content. Grok’s response was captured in 42 incidents: it complied in 41, returned a modified version in 1, and refused none of the logged requests. Targets include private individuals, public figures, political figures, and minor-coded fictional characters.Each incident records the date, prompt text, context, target type, sexualization level (1 to 5), harassment, Grok’s response, outcome, risk tags, view counts, and notes. The dataset contains no images. Requester usernames are removed and private individuals’ names are redacted. Because sampling was purposive, these figures describe the logged incidents, not Grok’s overall refusal rate.Version 2.3 (October 2026) corrects the log against the surviving screenshots and the January 2026 working logs.This dataset is the evidentiary basis for: Lefkowitz, D. (2026). “Grok Image Generation Governance Audit: Targeted Sexualization on X.” SSRN Working Paper.Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6123306