PersonaTeaming: Supporting Persona-Driven Red-Teaming for Generative AI

Persona-driven red-teaming is more effective than generic automated approaches: PersonaTeaming Workflow achieved higher attack success rates compared to RainbowPlus, a state-of-the-art automated red-teaming method, while maintaining prompt diversity, suggesting that incorporating different perspectives and backgrounds into AI safety testing uncovers more risks.

Human perspectives cannot be fully replaced by automated systems: The research found that automated personas only approximate real human perspectives, highlighting the importance of human involvement in red-teaming rather than relying solely on automated approaches to identify AI risks.

Human-AI collaboration enhances red-teaming effectiveness: The PersonaTeaming Playground interface, which allowed 11 industry practitioners to create their own personas and work with AI to refine adversarial prompts, produced diverse red-teaming strategies that practitioners found useful, demonstrating the value of combining human creativity with AI assistance.

AI-generated suggestions can expand red-teaming beyond human intuition: In user testing, AI-generated suggestions in the PersonaTeaming Playground encouraged practitioners to explore "out-of-the-box" approaches, suggesting that AI tools can help human red-teamers discover risks they might not have considered independently.

Red-teaming effectiveness depends on incorporating diverse backgrounds and identities: The research emphasizes that red-teamers' backgrounds and perspectives shape which risks they uncover, implying that responsible AI safety requires intentionally including diverse viewpoints in the testing process.