Data-Centric Safety and Ethical Measures for Data and AI Governance

Datasets are foundational to AI safety, yet often overlooked: The paper identifies that while much attention goes to monitoring AI models themselves, the quality and ethics of the data used to train these models has received significantly less focus—despite datasets being crucial to determining what capabilities (both beneficial and harmful) AI systems develop.

A framework for responsible dataset design across the full lifecycle is needed: The authors propose a domain-agnostic framework that addresses dataset safety and ethics at multiple stages, from creation through use and sharing, rather than treating data governance as a single point in time.

Low-quality and unethical data directly enables AI misuse: The paper emphasizes that datasets containing unsafe or unethical content can introduce harmful capabilities into AI foundation models, meaning responsible dataset design is a practical tool for reducing risks of AI misuse before models are even deployed.

Data-centric approaches can support red teaming and build trust: By implementing safety and ethical measures at the dataset level, organizations can better prepare for adversarial testing (red teaming) and increase stakeholder confidence in AI systems—addressing both technical safety and public trust simultaneously.

One framework can work across different domains and applications: Rather than requiring separate governance approaches for each industry, the proposed framework is designed to be adaptable across various sectors, making responsible data practices more scalable and standardized.