Private, Verifiable, and Auditable AI Systems

Privacy and accountability don't have to conflict: The research demonstrates that AI systems can simultaneously protect user data privacy and allow for verification that the system is working correctly—two goals that are often seen as competing but can be achieved together through cryptographic techniques.

Zero-knowledge proofs enable trustworthy AI claims: The work introduces methods using zero-knowledge cryptography that allow AI developers to prove their systems meet safety and performance standards without revealing proprietary details, addressing a key tension between transparency and competitive interests.

Secure computation protects sensitive data during AI deployment: By using secure multi-party computation and trusted execution environments, the research shows how organizations can audit large language models and information retrieval systems while keeping confidential data protected throughout the process.

Access controls matter for autonomous AI systems: As AI systems become more autonomous and work together in multi-agent environments, the research emphasizes that robust credentialing systems and access controls are essential safeguards to prevent misuse and ensure accountability.

Technical solutions must align with policy needs: The thesis bridges the gap between technical innovation and real-world governance by identifying risks in the AI pipeline and designing solutions that can inform both system design and policy discussions on AI safety.