Now that we've established the foundational motivations for why verification is important, we'll dive into building key intuitions around the timelines, scope, and urgency of what a verification regime might actually look like. We'll do this by analyzing AI 2040: Plan A, the most detailed and well-known public attempt to forecast a successful AI slowdown and complement verification regime.
AI 2040: Plan A?
What is the ideal end state? What agreement reaches it? What would verification have to cover for the agreement to hold? The most detailed public attempt to answer all three is AI 2040: Plan A, published by the AI Futures Project, the team behind the earlier AI 2027 scenario. While AI 2027 dramatized how a race ends badly, Plan A tells a dated, concrete story in which a US–China deal, layered verification, and a managed slowdown deliver a good outcome by 2040. Its verification supplement specifies the machinery: mutual compute declarations checked by inspections, datacenters retrofitted so that large-scale training is detectable, optical network taps feeding trusted recomputation servers, secure R&D facilities, and production caps on unverified hardware.
AI 2040: Plan A — verification supplement
The machinery behind Plan A, specified in full.
AI Futures Project
Below, you will read and gauge the feasibility and robustness of Plan A yourself. Read with a critical but thoughtful eye. For every date, percentage, and recommendation, ask yourself: what assumptions have to be true in order for this to hold? What existing or historical precedents have pointed to this mechanism succeeding or failing? What alternative solutions or estimates could there be, and how does this hold up against them? What is left uncertain? Remember: there is no correct answer, but you want to back your opinions with logic and evidence.
Your task is to write one of two open-ended essay prompts, consisting of shorter, prompted questions linking into a more cohesive essay at the end:
A: stress-test the Plan A verification supplement, or
B: compare Plan A (verified slowdown) and Plan S (complete shutdown).
Clarify the source of each claim you make, and explicitly identify gaps, uncertainties, or ambiguities as they appear. You won't know everything, and you aren't expected to.
Option A: Stress-Test Plan A's Verification Regime
Read the AI 2040 Verification Plan, paying attention to:
- Summary of the Plan
- Concrete inference-only retrofitting proposal
- 2029–2030: Deal Implementation
- Third-party countries begin joining the deal
Plan A's verification problem has two broad parts. First, the regime needs confidence that declared compute is being used in permitted ways. Second, it needs to keep undeclared or covert compute small enough that it cannot overturn the agreement. Your job, as a critical reader, is to interrogate whether Plan A's assumptions are reasonable; outlooks are optimistic, realistic, or pessimistic; the key load-bearing mechanisms of their proposal; any unrealistic aspects of the proposal that you should give less weight, if at all; and any important considerations you do not see.
A1. Identify the Regime's Strongest Mechanism or Recommendation
A2. Identify the Regime's Weakest Link(s)
A3. Stress-Test the Timeline
A4. Assess the Covert-Compute Margin
A5. Final Essay
You have now approached the AI 2040 Verification Supplement from four angles: its most load-bearing strength, its weakest link, its biggest timeline bottleneck, and the amount of failure the regime can tolerate.
Now imagine Plan A is moving from scenario to serious policy proposal. Decision-makers are asking whether its verification regime is strong enough to rely on as written, or whether it needs major changes before anyone should build an agreement around it. They have asked you for an assessment.
At the end, select one of the three recommendations, and briefly justify why:
Option B: Compare Plan A and Plan S as Verification Problems
Plan A and Plan S aim at different kinds of AI agreements. Plan A permits substantial AI activity under a layered verification regime; Plan S calls for a much stronger halt on frontier AI development. Those choices shape the verification problem: what counts as compliance, what evidence inspectors can collect, how much of the AI ecosystem must remain visible, and which actors must cooperate.
Skim the AI 2040 Verification Supplement, the discussion of Plan S, and the relevant Plan S section of the AI 2040 FAQ. Plan S has no equivalent verification supplement, so you will need to infer what a credible verification regime for it would require using mechanisms from this course.
B1. Which Plan Gives Verification the More Tractable Target?
B2. Which Plan Could Provide Stronger Evidence of Compliance?
B3. Which Plan Creates the Harder Monitoring Problem?
B4. Which Regime Could States Actually Cooperate On?
B5. Final Essay
Policymakers are deciding whether a serious international AI agreement should more closely resemble Plan A or Plan S. They have asked which approach provides a verification regime strong enough to rely on.
Which creates the more robust verification regime: Plan A or Plan S? Make a recommendation:
Optional: Additional Exercises and Materials
The earlier scenario from the same team — how a race ends badly, dramatized. Read both of its endings.
AI Futures Project (2025)
Curated Readings
You have now taken a position on a concrete verification regime. Use these readings to push on whichever part of your argument still feels least settled. Skim broadly; deep-read the one or two closest to the question you found hardest.
Nuclear Arms Control Verification and Lessons for AI Treaties
If you want the historical reality check. Read for how adversarial states built confidence without eliminating uncertainty, and where the nuclear analogy becomes strained when applied to AI.
Baker (2023)
If your Plan A assessment hinged on whether layered verification can actually produce enough assurance. Read for redundancy, correlated failure, privacy, and what still has to be built before a multilayer system deserves high confidence.
Baker et al. (2025)
Verifying Restrictions on Frontier AI Research
If you chose Plan S — or remain unsure whether a stronger halt is actually easier to verify. Read for the uncomfortable parts of a halt that compute monitoring alone does not solve: algorithmic research, experiments, personnel, code, and other less physically legible activity.
Scher (2026)
Verification for International AI Governance
If your argument turned on political feasibility or inspection access. Read for how the feasible verification architecture changes with political access, privacy constraints, and the type of agreement states are trying to verify.
Oxford Martin AI Governance Initiative (2025)
An International Agreement to Prevent the Premature Creation of Artificial Superintelligence
If you want to compare your own architecture against a concrete proposal for an international halt — especially its verification sections and Appendix D. Read for one attempt to translate a pause posture into specific restrictions on compute and dangerous AI research.
Scher et al. (2025)
Mechanisms to Verify International Agreements About AI Development
If you want the mechanism catalog. Read selectively: choose the policy goal or mechanism closest to your essay and ask whether its evidence, access requirements, and remaining R&D gaps change your conclusion.
Scher and Thiergart, MIRI Technical Governance Team (2025)
Drill Bench
The primer bench: inspection games, credible commitment, and two-level games. One step at a time: commit, read why, then Continue.