National AI Safety Institutes and International Coordination
Governments are building dedicated institutions to evaluate AI risk and develop safety standards. Their effectiveness will depend on whether they can coordinate globally on a technology that respects no borders.
Collective Intelligence
Research & Analysis
Purpose of AI Safety Institutes
As AI capabilities expand, governments are establishing dedicated institutions to study, evaluate, and mitigate AI risk. AI safety institutes occupy a distinctive role in the governance landscape: they are neither standard-setting bodies nor enforcement agencies, but the research and evaluation infrastructure that makes evidence-based governance possible. Their core functions include empirical risk assessment, model evaluation before and after deployment, research coordination across institutions, and the development of policy guidance grounded in technical understanding.
These institutions represent a recognition that effective AI governance cannot rely solely on government officials and regulators assessing AI systems from the outside. The technical complexity of frontier AI systems requires dedicated expert capacity that can engage with AI developers on equal technical terms, conduct independent evaluations, and develop the methodological standards that give regulatory requirements operational meaning.
The UK Model
The UK AI Safety Institute, established in 2023, provided an early model for what dedicated government AI evaluation capacity can look like. The institute developed evaluation protocols for frontier AI systems, conducted pre-deployment evaluations of major new models in collaboration with leading AI laboratories, and published research on evaluation methodology and model capabilities. Its work on eliciting dangerous capabilities from models — including cybersecurity offence, biosecurity risks, and deceptive alignment — demonstrated that independent government evaluation can identify risks that companies' own safety assessments may not surface.
The UK model has been influential internationally. Multiple governments established safety institutes following the 2023 Bletchley AI Safety Summit, and a broader network of institutes has been coordinating through the International Network of AI Safety Institutes. This network provides a mechanism for sharing evaluation protocols, pooling technical expertise, and developing common methodological standards — reducing the risk that each national institute must independently develop evaluation capacity that could be shared.
Model Evaluation and Safety Research
Evaluating frontier AI systems is methodologically complex in ways that distinguish it from conventional software testing or product safety assessment. The capabilities and risks that matter most are often emergent properties of large models that do not appear in smaller versions — meaning that evaluations conducted during development may not predict the behaviour of deployed systems. Traditional safety testing assumes that exhaustive enumeration of failure modes is possible; AI systems operate in open-ended environments where exhaustive testing is definitionally impossible.
Safety institutes are developing evaluation approaches that address these challenges: adversarial testing designed to elicit harmful capabilities that systems have been trained to avoid; robustness analysis that assesses performance degradation under distribution shift; interpretability research that seeks to understand the internal representations underlying model behaviour. These techniques are still developing, and the field lacks the standardised evaluation frameworks that more mature safety-critical industries take for granted.
International Coordination
AI safety challenges do not respect national borders. Frontier models developed in one jurisdiction are deployed globally; safety properties — or their absence — affect users everywhere. International coordination is therefore not merely desirable but necessary if safety institutes are to be effective. A model that passes one nation's safety evaluation while failing another's creates market confusion and may create incentives for deployment in permissive jurisdictions.
The International Network of AI Safety Institutes, the OECD's AI governance work, and the technical discussions emerging from the Global Partnership on AI all provide frameworks for coordination, but progress is constrained by the political sensitivity of AI governance and the competitive dynamics among major AI powers. The most tractable areas for international coordination — sharing of evaluation methodologies, protocols for incident reporting, and minimum disclosure standards for frontier model deployment — are those where the shared interest is clearest and political friction is lowest.
Related Articles
China's Generative AI Governance Framework
China's governance model for generative AI combines technological ambition with centralised regulatory oversight. Understanding it is essential for any organisation navigating the global AI landscape.
Copyright, Creativity, and Generative Models
Generative AI challenges the assumptions that underpin copyright law. Who owns AI-generated output? How should training data be treated? Courts and regulators are still writing the answers.
Defence AI and Autonomous Systems Doctrine
Autonomous systems with AI-assisted decision-making are entering defence strategy. The ethical, legal, and geopolitical implications of machines that act without continuous human oversight are profound.
Read the full intelligence feed
Signals, analysis, and strategic context from across the global AI landscape — curated for leaders.
Back to Research →