What Cryptography can do for AI

BAIOSPHERE's Munich AI Signature Lecture brought one of the most influential figures in computer science to Munich: Prof. Shafi Goldwasser, co-founder of modern cryptography and winner of the 2012 ACM Turing Award. Her lecture, "A Cryptographic Perspective on Machine Learning Trustworthiness Challenges," demonstrated how cryptographic methods can enhance machine learning trustworthiness with mathematical guarantees.

The lecture addressed one of the central questions of the current state of AI: can we trust systems that even their own developers do not fully understand, and that are already being used in health, finance, law, and science? Goldwasser approaches this question from the perspective of cryptography, which has kept digital platforms reliable for decades. Rather than testing models and patching problems post facto, she argued for systems that are trustworthy by design: their guarantees are mathematically proven against clearly defined adversaries, and where no such guarantee is possible, that limit is proven, pointing to where regulation has to step in.

Drawing on her recent research, Prof. Goldwasser showed how tools such as interactive proofs, digital signatures and time-lock puzzles bear on concrete challenges along the machine learning pipeline: verifying individual model outputs, detecting and mitigating hidden backdoors, and understanding the fundamental limits of safety filters. Concluding in her own words: “Ask not what AI can do for you, but what cryptography can do for AI.”

The lecture took place on 20 October 2026 in the Plenary Hall of the Bavarian Academy of Sciences and Humanities in Munich. The full recording is available on YouTube, and more information is on the BAIOSPHERE event page.