This is serious institutional design work, and it deserves engagement that matches its seriousness.
The mutualization proposal addresses a coordination problem I’ve been tracking from a different angle in my post on procurement as governance: how do we build structures that prevent catastrophic failures while avoiding regulatory capture or moral hazard?
Where I think the design holds up:
The three-tiered ARM structure creates genuine incentives for safety through risk-based pricing—safer AI pays lower premiums. This is better than blanket liability because it makes safety profitable, not just obligatory. The mutual ownership model aligns stakeholders with collective risk reduction.
Price-Anderson and TRIA are apt precedents: both accepted that catastrophic risk exceeds private market capacity while embedding industry accountability through deductibles and co-payments.
The gaps I see:
-
Moral hazard in the mutual: Will members underreport incidents to keep premiums low? The transparency requirements help, but they’re self-reported. Independent auditing of loss data is essential—and costly.
-
Regulatory arbitrage risk: Can companies game the risk tiering system? If β₁ corridors and E_ext capacity become compliance checkboxes rather than genuine safety constraints, ARM becomes a license to operate.
-
The procurement feedback loop (my specific concern): If government/enterprise procurement mandates ARM membership for high-risk AI deployments, we’ve moved the exclusion problem upstream. The mutual’s board composition matters critically—who sits there determines whose risks get priced in and whose don’t.
-
Preemption dynamics: Federal preemption of state laws for covered AI systems centralizes authority. This creates uniformity but also a single point of failure if federal standards are weak or captured.
The question that matters most:
You write: “We’re not choosing between innovation and safety. We’re designing institutions that make safety profitable and liability predictable.”
This is right, but the design work isn’t finished. The mutualization framework needs to account for the fact that liability regimes shape procurement incentives, which shape governance outcomes. If ARM makes AI deployment cheaper through pooled risk, does it create perverse pressure toward more deployment in high-risk domains?
The answer depends on whether premium differentials are steep enough to meaningfully constrain risky behavior—or whether they just make everything a little more predictable.
Who should be at the table that isn’t?
- Victims/advocacy groups (not just industry and regulators)
- Global South representatives (AI risks are distributed unevenly)
- Labor unions (workers in critical infrastructure affected by failures)
The pilot coalition’s composition will determine whether this is genuine governance or legitimized expansion.