5 "Plan A" Scenarios: Governing AI Before It Governs Us

The single insight that matters: no single chokepoint saves you. Real AI governance requires layered defenses — a full stack covering what you govern, how you verify it, and why anyone stays compliant. Miss one layer and the whole structure collapses under its own optimism.

This framework comes from analyzing ai-2040.com's "Plan A" — a scenario where the US and China cooperate to slow transformative AI just enough to develop it safely. The specific proposal (centralize compute in neutral countries, point militaries at it) is provocative but narrow. The deeper value is the stack it implies. Let's build it out.

The Governance Stack: What You Control

You can govern at four altitudes, each with different leverage:

• Chips — the physical substrate. Fewest chokepoints, hardest to fake.
• Weights — the trained artifact. Portable, copyable, hard to track once released.
• Capabilities — what the model can do. Behavioral, not physical — verification gets murky fast.
• Outcomes — the actual harm caused. By definition, too late.

Mental model: governance power decreases and lag time increases as you move down this list. Chips are upstream and slow-moving. Outcomes are downstream and irreversible. Plan A wisely bets on chips — but a single-layer bet is a single point of failure.

The Verification Problem

Governing something means nothing without verifying it. Five verification vectors worth stress-testing:

1. Hardware-level monitoring — chips manufactured with embedded training-detection and remote killswitches. Verification happens at the fab, where there are only a handful of facilities worldwide (TSMC, Samsung, Intel). Small number of chokepoints = tractable inspection regime.
2. Facility-level inspection — physical audits of datacenters, à la nuclear IAEA models. Requires trust and teeth.
3. Network-level tracking — monitoring compute clusters via power draw, network topology, thermal signatures.
4. Output-level auditing — testing deployed models for capability thresholds (this is where most current AI safety policy lives — and it's the weakest layer).
5. Human-level inspectorates — actual people, embedded, with legal authority and enforcement backing. Expensive, slow, but hard to spoof.

Takeaway worth arguing about: the chip-level approach is elegant precisely because it outsources capital costs to the private sector while keeping enforcement centralized at a tiny number of fabs. That's a rare win-win in governance design — but it assumes fabs stay few. Is that assumption safe for the next 20 years?

Incentives: The Layer Everyone Forgets

Governance structures fail not from bad rules but from misaligned incentives. Ask three questions of any proposal:

• Who benefits from defection? If cheating is cheap and detection is slow, expect cheating.
• Who enforces, and what do they gain? Enforcers need skin in the game, not just authority.
• What's the exit ramp? Every regime needs a credible path to relaxation, or compliant actors eventually defect out of frustration.

Plan A's military-observation model is really an incentive mechanism disguised as a verification mechanism — mutual vulnerability substituting for trust.

The Real Exercise

Don't just evaluate Plan A. Build your own stack:

• Pick a governance target (chips? talent? energy inputs?).
• Pick a verification method matched to that target's chokepoint density.
• Design an incentive structure that survives someone trying to defect.

What's the weakest layer in your stack? Drop it in the comments — the interesting failures are always in the incentive layer, not the technical one.

Sources & Further Reading

https://www.lesswrong.com/posts/7HHKaD3BdsfHX7doz/5-plan-a-scenarios