Enabling External Scrutiny of AI Systems with Privacy-Enhancing Technologies

Privacy-enhancing technologies can enable independent audits of AI systems without exposing proprietary information or sensitive data, addressing a major barrier to external scrutiny that has previously prevented researchers from accessing AI systems for evaluation.

Real-world deployments have already demonstrated feasibility, including audits of social media recommendation algorithms through the Christchurch Call initiative and evaluations of frontier AI models by the UK AI Safety Institute, showing these approaches work beyond theoretical frameworks.

External scrutiny is essential for responsible AI governance but has been hindered by legitimate corporate concerns about security, intellectual property, and privacy—technical infrastructure from OpenMined offers a practical solution to this longstanding tension.

Policymakers can accelerate progress by focusing on legal and regulatory frameworks that support this approach, rather than requiring companies to choose between transparency and protecting sensitive information, making external audits a more feasible governance tool.

This infrastructure represents a maturation of privacy-enhancing technologies that can be configured in multiple ways to enable different types of AI system scrutiny, suggesting broader potential applications beyond the current case studies.