Who Is the Safety Model Validated For? A Management-Governance Audit of Demographic Transparency in AI-Based Construction Safety Prediction

The dataset enables a dual-coded systematic audit of 45 peer-reviewed studies published between 2015 and 2026 on AI-based construction safety prediction, monitoring, and risk assessment. This audit examines if published work offers sufficient demographic transparency to evaluate the extent of AI safety model validation by gender. The dataset includes bibliographic details, study stream, AI method, source-data characteristics, mention of gender or sex, reporting of sample composition by gender, use of gender as a model feature, reporting of gender-disaggregated model performance, acknowledgement of gender as a validity limitation, and determination of the presence of a sex or gender field in the source data. It also contains failure-mode classifications, coder decisions, agreement indicators, and consensus notes. The dataset supports the paper’s finding that the field lacks demographic transparency to assess subgroup validity: no study reported gender-disaggregated model performance, no study acknowledged gender as a validity limitation, only one study used gender as a model feature, and in 28 of the 45 studies it could not be determined whether the source data had a sex or gender field. This dataset is not comprised of individual worker-level data, nor is it any newly collected personal data. It is a literature-audit dataset from published studies. The dataset is intended to support reproducibility, peer review, and future research on demographic transparency, validation scope, and governance of AI-based construction-safety tools.