The Coordination Gap in Frontier AI Safety Policies

Isaak Mengesha · 2026

Read The Coordination Gap in Frontier AI Safety Policies by Isaak Mengesha (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-10015.txt.

Introduction

Mengesha argues that top AI safety rules focus on stop gaps and neglect joint action when stop gaps fail. Tests of skill, gates on use, and limits on use dominate, while shared response skill gets little cash or care. The gap is built in, he says, since wide gains from strong shared response spread far while costs land on single firms. Borrowing from nuclear safety and plague plans, he urges set pledges, shared rules and advance swaps of if then response plans.

Catalog record

CREATORIsaak Mengesha
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30077
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.10015

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

The Coordination Gap in Frontier AI Safety Policies Isaak Mengesha Abstract Frontier AI Safety Policies concentrate on prevention—capability evaluations, deployment gates, usage constraints—while neglecting institutional capacity to coordinate responses when prevention fails. We argue this coordination gap is structural: investments in ecosystem robustness yield diffuse benefits but concentrated costs, generating systematic underinvestment. Drawing on risk regimes in nuclear safety, pandemic preparedness, and critical infrastructure, we propose that similar mechanisms—precommitment, shared protocols, standing coordination venues—could be adapted to frontier AI governance. Without such architecture, institutions cannot learn from failures at the pace of relevance. 1 Introduction The central challenge of AI governance is to harness the economic and social benefits of widespread deployment while preventing the potential harms that such systems may generate. Yet the societal impacts of advanced AI systems remain difficult to anticipate 5 , creating fundamental tensions in policy design. Regulators seek to hold developers accountable without driving innovation offshore, and to impose sufficient safety requirements without stifling the very technological progress they aim to govern. This is a delicate balancing act, and most research has naturally focused on optimizing preventative measures for development and deployment 70 . Given the scale of potential harms, societies are likely underinvesting in AI safety, and additional preventive effort plausibly yields positive marginal returns 35 . Yet even prevention with extremely high reliability—on the order of 99%—is unlikely to guarantee full containment in complex socio-technical systems 57 . Recognizing this, many researchers anticipate that harmful incidents or near-misses (“warning shots” or “focusing events”) will be necessary to shift political feasibility toward more stringent and costly interventions 11 . In effect, this requires taking seriously the possibility of failures that generate large or even catastrophic societal harm. As AI capabilities and deployment accelerate, such failures become more likely. Importantly, this does not weaken the case for prevention: preventive investment continues to dominate expected-value calculations for existential risk. The question is instead whether accounting for inevitable failures changes where additional effort has the highest marginal payoff. We argue that it does. When institutions lack capacity for shared learning, coordinated response, and adaptive adjustment, governance is forced to rely almost exclusively and increasingly on preventive control—even when prevention cannot eliminate residual risk. Under these conditions, the absence of resilience is what makes prevention the only available, yet ultimately insufficient, strategy. Broader frameworks for societal adaptability have been proposed 10 , yet they leave open concrete implementation pathways. Meanwhile, a large body of work has identified specific harms: cognitive overload at the societal level 39 and shifts in the offense-defense balance in cybersecurity 29 , among others. More broadly, the literature identifies a “context gap” between technical prevention protocols and holistic societal impact assessments 70 . What matters for frontier-AI governance is whether systems and institutions can maintain acceptable performance under unexpected conditions and continue functioning when shocks occur. Current policies focus on prevention, but they do not address this broader requirement for adaptive (learning) capacity 14 . Most proposals contribute frameworks for monitoring and reflection around such events 28 ; 64 ; 69 . Monitoring is indeed a necessary condition for effective mitigation. Yet the underlying challenge is a collective action problem: benefits are diffuse while costs are concentrated, creating incentives to free ride and a high risk of coordination failure. These dilemmas are well studied, and established strategies exist to address them; one such mechanism is precommitment 55 . The relevant actors span government regulators, critical infrastructure operators, private labs and cloud providers, and civil society. No single actor possesses authority or control across all domains, yet optimal responses require coordination among them. During high-pressure crisis events, this coordination capacity becomes critical—a point recognized by calls for interagency task forces, public-private partnerships, and treaty-like mechanisms. From other high-risk fields, we can extract best practices for such coordination efforts 67 . While multiple actors must account for robustness to improve coordination, labs as primary developers and operators carry particular responsibility. Labs are already producing Frontier AI Safety Policies (FASPs), yet these policies implicitly assume that preventative measures will succeed. Taking seriously the possibility that they will not—and planning accordingly—requires integrating robustness standards into these policies, making them the natural anchor point for multi-actor coordination. 2 The ’robustness gap’ in the broader Ecosystem Robustness in frontier-AI governance concerns whether policies and institutions continue to function when underlying assumptions fail, signals are ambiguous, or external conditions shift in ways developers did not anticipate 46 . The problem is that current approaches operate under the assumption of control and fail to take failure seriously. 2.1 Consequences of ignoring robustness Traditional “predict-then-act” approaches assume accurate forecasts. Under conditions of deep uncertainty, this produces brittle policies that lead to one of two failure modes: gridlock, where actors cannot agree on underlying assumptions, or overconfidence, where plans fail when futures diverge from expectations 61 ; 40 . This brittleness is particularly consequential when actors operate under differing assumptions about baseline risk and responsibility 25 ; 24 . While actors may not be able to agree on underlying probability distributions—and therefore on adequate degrees of preventative measures—they can nevertheless disclose their post-hoc responses in light of impending harm or novel evidence of potential harm 20 ; 31 . The myopic focus on monitoring compliance of individual developers neglects socio-technical feedback loops. System-level harms such as cyber-risks or disinformation cascades cannot be mitigated by individual actors alone 49 . When policies meet internal compliance standards yet still fail under real-world shocks, trust in governance architectures declines. Risk management in low-trust environments poses major challenges, as documented in other high-risk contexts 51 ; 34 ; 50 . As Figure 1 illustrates, existing AI governance efforts are heavily concentrated on prevention, with substantial work on alignment and testing 4 , usage constraints 12 , compliance auditing 13 , and regulatory norms 23 ; 71 . By contrast, mitigation and preparedness remain far less developed. Although monitoring and incident reporting have begun to receive attention 52 , policy proposals addressing coordination and crisis management are sparse, leaving coordination capacity the most underexplored element of the governance landscape. Robust Decision Making (RDM) offers an alternative framework. Rather than relying on contested forecasts, it stress-tests candidate strategies across hundreds of plausible futures to reveal vulnerabilities and adaptive levers. This shifts attention from prediction to preparation, and from “what will happen?” to “what actions hold across many futures?” and “what tradeoffs are we willing to make?” 61 ; 40 . Neglecting such robustness thinking in frontier AI governance means that Responsible Scaling Policies (RSPs) remain strong at local gating—controlling deployment decisions within a single organization—but weak at global adaptability. Without cross-domain triggers, external signposts, and pre-committed adaptation pathways, policies risk either paralysis or ad hoc responses once crises unfold. Risk Prevention Mitigation Model capabilities Deployment scale Systemic exposure Societal harm Alignment, testing Usage policies Compliance audits Regulation, norms Rollback, patches Incident response Cross-sector coordination Crisis management Developer/ Lab Deployer/ Platform Integrators/ Institutions End Users/ Society ad hoc / underdeveloped current focus Figure 1: The robustness gap in frontier AI governance. Risk of harm accumulates downstream (top row). Opportunities for prevention concentrate upstream. Broader mitigation capacity remains underdeveloped. However, this determines society’s capacity to adopt and learn. 2.2 Tentative steps

Read the complete text