The coordination gap in frontier AI safety policies

The paper argues that frontier AI safety policies currently suffer from a structural "coordination gap" by overemphasizing prevention while neglecting the institutional capacity needed to coordinate responses when failures occur, and proposes adapting mechanisms from nuclear safety and pandemic preparedness to address this underinvestment in ecosystem robustness.

The Gist

Here is an explanation of the paper "The Coordination Gap in Frontier AI Safety Policies" using simple language and everyday analogies.

The Big Picture: Building a Better Lifeboat

Imagine humanity is building a massive, incredibly fast spaceship (Frontier AI). We are all excited about the destination, but we know the ship might crash.

Right now, the people in charge (governments and AI labs) are obsessed with prevention. They are building better shields, stronger engines, and strict rules to make sure the ship never crashes. They are checking the fuel, testing the navigation, and locking the doors.

The Problem: The author, Isaak Mengesha, argues that we are ignoring the lifeboats.

Even with the best shields, a spaceship might still crash. When that happens, the current plan is: "Oh no, the ship crashed! Let's panic and figure out who to call, who has a lifeboat, and how to share the survivors."

The paper says this is a disaster waiting to happen. We need to agree on the lifeboat plan before the crash, not after.

---

The Core Problem: The "Coordination Gap"

The paper calls this the Coordination Gap.

The Analogy: The Neighborhood Fire
Imagine a neighborhood where every house has a great fire extinguisher (Prevention). But if a fire gets too big and the extinguishers fail, the neighbors need to work together to save the street.

• The Gap: Right now, every neighbor is focused on buying a better extinguisher. But no one has agreed on:
  • Who calls the fire department?
  • Who opens the gates for the trucks?
  • Which houses get evacuated first?
  • How do we share water if the main pipe breaks?

If a fire starts, everyone will run around shouting, "I thought you were calling!" or "I didn't know you had a hose!" By the time they figure it out, the whole street is gone.

Why is this happening?
It's a "Public Good" problem.

• The Cost: It's expensive and annoying to plan the lifeboat. You have to slow down your work, spend money on standby teams, and share secrets.
• The Benefit: If the plan works, everyone in the neighborhood is safe.
• The Result: Everyone hopes someone else will do the planning. They all free-ride, hoping their neighbor will be the hero. So, nobody plans, and we are all vulnerable.

---

The Proposed Solution: The "Scenario Response Registry" (SRR)

The author suggests a new tool called a Scenario Response Registry (SRR). Think of this as a Shared Playbook or a Digital Drill Hall.

How it works (The Analogy):
Imagine a central computer system where every AI company and government agency has to write down a "If/Then" plan for specific disasters.

1. The Library of Scenarios: An independent group creates a list of scary "What Ifs."
   • Example: "What if an AI starts hacking power grids?"
   • Example: "What if an AI spreads fake news that causes a riot?"
2. The Filing: Every company must file a plan:
   • IF the AI starts hacking power grids...
   • THEN we will immediately shut down our servers and alert the government.
   • WITH these resources (our emergency team, our backup data).
3. The Review: A central authority looks at all these plans. They check for gaps.
   • Critique: "Company A said they will shut down, but Company B didn't say they will help Company A. You need to coordinate!"
4. The Drills: They run "tabletop exercises" (like a fire drill) to see if the plans actually work when people are stressed.

Why is this better?
Instead of waiting for a real crash to see who talks to whom, we practice the conversation now. We find out that Company A and Company B have incompatible plans before the disaster happens.

---

Why Don't We Do This Already?

The paper explains that companies are hesitant because:

1. It's expensive: It costs money to have a team ready to act instantly.
2. It slows them down: You can't move as fast if you have to stop and check your safety plan every time.
3. Fear of competition: Companies don't want to show their secrets or admit their weaknesses to rivals.

However, the author argues that if we don't do this, the cost of a single catastrophic failure will be infinitely higher than the cost of planning.

The Takeaway

Current Policy: "Let's build a perfect shield so we never crash."
The Paper's Advice: "Shields aren't perfect. We need to build a coordinated rescue team and practice the rescue now, before the crash happens."

The goal isn't to stop innovation; it's to make sure that if things go wrong, we don't panic. We want a system where, when the alarm rings, everyone knows exactly what to do, who to call, and how to help each other.

In short: We are great at building the car, but we are terrible at agreeing on the evacuation plan. It's time to write the plan.

Technical Summary

Here is a detailed technical summary of the paper "The Coordination Gap in Frontier AI Safety Policies" by Isaak Mengesha.

1. Problem Statement

The paper identifies a critical structural flaw in current Frontier AI Safety Policies (FASPs): a "coordination gap" regarding institutional capacity to respond when prevention fails.

• The Prevention Bias: Current governance focuses almost exclusively on upstream preventative measures (capability evaluations, deployment gates, usage constraints). While necessary, these measures assume a "predict-then-act" model that fails under deep uncertainty.
• The Resilience Deficit: Even with high-reliability prevention (e.g., 99%), complex socio-technical systems inevitably experience failures, near-misses, or "warning shots." Current policies lack the institutional architecture to manage these failures adaptively.
• The Collective Action Problem: Building robustness (resilience) is a public good with diffuse benefits (societal safety) but concentrated costs (slowed development, investment in standby capacity). This creates a systematic underinvestment in robustness, as private actors (labs, cloud providers) prioritize speed and market share over systemic resilience.
• The Gap: When a crisis occurs, actors lack pre-committed, interoperable protocols for coordination. Responses are improvised, siloed, or delayed, leading to catastrophic outcomes that could have been mitigated through pre-existing alignment.

2. Methodology

The paper employs a comparative policy analysis and conceptual framework design approach:

• Cross-Domain Analogy: The author draws on risk management regimes from nuclear safety, pandemic preparedness, and critical infrastructure. These fields share characteristics with frontier AI (catastrophic tail risk, socio-technical complexity) and have developed specific coordination primitives (e.g., precommitment, shared protocols, standing venues).
• Gap Analysis: The paper analyzes existing FASPs (e.g., Anthropic's ASL, OpenAI's risk gates) to demonstrate that while they contain internal triggers, they lack cross-domain triggers and externalized decision logic.
• Institutional Design Proposal: The paper proposes a new institutional mechanism, the Scenario Response Registry (SRR), designed to operationalize "Robust Decision Making" (RDM) principles. This involves stress-testing strategies across plausible futures rather than relying on single-point forecasts.

3. Key Contributions

A. Theoretical Contribution: Reframing Robustness

The paper argues that robustness is not merely a technical feature but a coordination architecture. It posits that the absence of resilience forces governance to rely on prevention as the only strategy, which is insufficient. The core argument is that learning and coordination must be moved upstream (before crises) rather than relying on post-hoc institutional development after "focusing events."

B. The Scenario Response Registry (SRR)

The primary technical contribution is the proposal of the SRR, an institutional mechanism to make coordination capacity visible, comparable, and stress-testable.

• Mechanism: A public authority maintains a library of technical, socio-economic, and security scenarios.
• Process: Relevant actors (labs, cloud providers, governments) file standardized "if-then" plans ex ante. These plans specify:
  • Triggers: Specific thresholds (technical or societal) that activate a response.
  • Actions: Concrete steps to be taken.
  • Resources: Commitments of compute, personnel, or capital.
• Harmonization: The registry reviews filings to identify gaps, overlaps, and incompatibilities, creating "common knowledge" among actors.
• Incentives: The SRR links disclosure quality to tangible consequences (e.g., regulatory forbearance, access to public compute, or forfeiture of bonds/insurance upon failure to act).

C. Diagnostic Framework: The Robustness Gap

The paper introduces a visual and conceptual framework (Figure 1) distinguishing between:

• Upstream Prevention: Alignment, testing, compliance audits (High focus).
• Downstream Mitigation: Incident response, cross-sector coordination, crisis management (Low focus/underdeveloped).
  The paper argues that the "robustness gap" lies in the lack of mitigation and preparedness capacity relative to the scale of potential systemic harm.

4. Results and Findings

• Current Policies are Siloed: Existing FASPs are effective at internal gating but fail to address cross-domain cascading risks (e.g., a cybersecurity breach leading to compromised biosafety protocols).
• Uncertainty Undermines Cooperation: Drawing on game theory (climate threshold-public-goods games), the paper finds that uncertainty about others' response thresholds causes rational actors to underinvest in coordination, leading to systemically disastrous outcomes.
• Feasibility of Polycentric Governance: Global, binding treaties are currently unrealistic due to geopolitical mistrust and dual-use concerns. However, polycentric governance (local SRRs, bilateral agreements, lab consortia) is viable. The key design principle is interoperability (compatible procedures) rather than immediate uniformity.
• Transparency as a Catalyst: The paper suggests that transparency regarding decision logic (not just triggers) can substitute for legal commitment in the short term, allowing for "skin-in-the-game" mechanisms like dynamic scoring and insurance instruments.

5. Significance

• Paradigm Shift: The paper moves the AI safety discourse from a singular focus on prevention (stopping bad things from happening) to resilience (managing the inevitable when they do happen).
• Operationalizing "Learning": It provides a concrete mechanism (SRR) to institutionalize learning and adaptation before a crisis, addressing the critique that AI governance is reactive.
• Addressing the Public Good Problem: By proposing incentive structures (forbearance, scoring, bonds), the SRR attempts to solve the economic misalignment where private costs of robustness outweigh private benefits.
• Policy Roadmap: It offers a realistic, incremental path forward (starting with tabletop exercises and partial participation) that acknowledges geopolitical constraints while building the "institutional muscle" required for future global coordination.

In summary, Mengesha argues that without a dedicated architecture for pre-committed, cross-actor coordination, frontier AI governance remains brittle. The Scenario Response Registry is proposed as the technical and institutional solution to bridge the gap between isolated safety policies and a resilient, adaptive global ecosystem.