Neurocryptography

How can we make it harder to misuse powerful AI systems, such as GPT-4, for propaganda, fraud, academic cheating, copyright violation, and so on? In this talk, I’ll survey one class of approaches, which I’ve called “neurocryptography”: namely, putting cryptographic functionalities inside or on top of AI models. I’ll detail some of the progress we’ve made on making AI outputs detectable as such, with the help of mathematical tools. I’ll also explain the interrelated technical, conceptual, and social challenges in actually deploying these sorts of solutions.

Speaker: Scott Aaronson, University of Texas at Austin