A Systematic Review of Open Datasets Used in Text-to-Image (T2I) Gen AI Model Safety

Dataset composition directly impacts safety research validity: The paper finds that the quality and composition of publicly available datasets are "crucial" for T2I safety research, meaning flawed datasets can lead to incomplete or misleading conclusions about whether these models are actually safe.

Significant gaps exist in harm coverage: The systematic review identifies gaps in how different types of harms are represented across datasets, suggesting that current safety evaluations may miss important risks or fail to test models comprehensively across all potential problem areas.

Prompt diversity matters for real-world safety: The paper examines "semantic and syntactic diversity of prompts," indicating that datasets with limited variety in how requests are phrased may not adequately prepare models for the full range of ways users will actually interact with T2I systems.

Researchers need transparency about dataset limitations: By documenting collection methods and distributions, the paper enables researchers to "critically assess the downstream impacts of their work," meaning safety claims should be evaluated in light of which specific datasets were used and what they may or may not represent.

Future dataset development is essential: The review identifies specific gaps in "dataset coverage and quality that future research may address," suggesting that responsible AI development requires ongoing investment in better, more comprehensive safety evaluation datasets rather than relying on existing imperfect tools.