Formal Confinement

Quinn Dougherty

Keywords: AI Boxing, AI Containment, Formal Verification

TL;DR: Policing the interface between AI and the world with formal verification

Abstract: We would like to put the AI in a box. We show how to create an _interface_ between the box and the world out of specifications in Lean. It is the AI's responsibility to provide a proof that its (restricted) output abides by the spec. The runnable prototype is at https://github.com/quinn-dougherty/formal-confinement.

Confirmation: I confirm that I and my co-authors have read the policies are releasing our work under a CC-BY 4.0 license.

Submission Number: 6