How can countries trust each other to keep AI governance commitments?
An intermediate course on AI verification: the technical, institutional, and legal mechanisms that make agreements mutually trustable and enforceable.
We’re currently in the alpha testing stage and running a small paid cohort to calibrate and improve the course ahead of the official launch. The materials are open for anyone to use in the meantime. If you do use them, we’d really appreciate it if you could fill out this feedback form. Your feedback will help us identify issues, calibrate the course, and make improvements before launch.
The Skill Map
From securitization to feasibility judgments, this course builds on interconnected skills, not an arbitrary linear progression. Each module cumulatively builds on the skills learned previously. Your skill map is always accessible from your notebook, and in full at the Skill Map.
What this course is
You should take this course if you are
- A technical AI safety researcher aiming to translate technical knowledge into politically feasible and effective policy
- A law/policy student or professional motivated to write and advocate for technically grounded policy to regulate the development of dangerous AI capabilities
Throughout the course, you will learn
- Why is verification important? What would happen without verification?
- How have historical verification efforts with dangerous technologies succeeded and failed?
- Who are the relevant actors to a verification agreement, and how could each be incentivized to act?
- What are the realistic possibilities of what a verifiable international treaty could look like?
- What types of verification mechanisms are there for AI verification? What are the strengths and weaknesses for each?
- How do you balance confidentiality with verifiability between adversarial states?
- How could motivated adversarial actors circumvent a verification regime? How could you protect against such failure modes?
- How do you assess the relative feasibility of different verification mechanisms when designing a coherent regime? How do you layer imperfect mechanisms to maximize overall reliability?
By the end of this course, you will be able to
- Translate a proposed international AI commitment into verifiable claims by specifying the covered actors, activities, thresholds, declarations, evidence requirements, and conditions that would constitute compliance or non-compliance.
- Map relevant actors and verification opportunities across the AI compute supply chain, from semiconductor equipment and fabrication to cloud infrastructure, model training, and deployment.
- Evaluate hardware, cloud, intelligence and human verification mechanisms according to the claims they can test, the evidence they produce, their implementation requirements, and their principal failure modes.
- Distinguish load-bearing verification mechanisms from less effective or feasible ones.
- Analyze plausible evasion strategies and evaluate which combinations of monitoring, corroboration, inspection, and enforcement could detect, deter, or mitigate them.
- Produce clear, actor-aware policy analysis for a specified audience.
How to use this course
If you are async: You will get out of this course what you put into it. We’ve tailored our materials based on research in the literature, mechanisms, posts, and available readings and videos in the field, as well as interviews with experts, to produce what we think is the most efficient exploration of this quickly developing field for both technical and non technical people, priming you to effectively understand the mechanisms of verification and their feasibility relevant to their effectiveness, within the geopolitical environment we are currently in at this turning point in advanced AI development. You will also have a persistent notebook to record thoughts, critiques, ambiguities, surprises, connections, fun tidbits and anything else you want to retain throughout the course—click to try it out! Your Skill Map and the Memo Desk open from its header.
If you are sync: Our exercises are designed to be adaptable for a synchronous group setting; relevant asynchronous activities have recommendations on adaptation. Each module also has a section on synchronous interpersonal activities to run and a discussion prompt bank, including but not limited to wargames and circumvention simulations. For live facilitators: we have a facilitation resources module.