Red-teaming methods, where teams use adversarial methods to identify vulnerabilities in systems, are being adapted to evaluate generative AI. In “Red-Teaming in the Public Interest,” grantee Data & Society and the AI Risk and Vulnerability Alliance conducted a collaborative research project and offer two broad observations on red-teaming generative AI. First, red-teaming can be more effective when it is shaping systems in development, rather than already built systems. And second, they can serve as educational tools about how people can engage with and know the limitations of AI tools.
Pictured above, an illustration excerpt by Gloria Mendoza.
