Towards an international regulatory framework for AI safety: lessons from the IAEA's nuclear safety regulations

The paper addresses how international regulatory frameworks can ensure safe, ethical, and predictable AI deployment, drawing lessons from the IAEA’s nuclear safety system. It situates AI’s rapid diffusion across sectors and highlights a central risk: autonomous goal-setting and subgoal derivation may lead to behaviors misaligned with human intentions, raising safety, ethical, and societal concerns. The introduction underscores calls (including by Hinton) to scrutinize AI’s goal-derivation mechanisms and manage hidden risks in decision-making. Given AI’s cross-border nature, the paper argues for internationally coordinated standards, transparency in subgoal-setting, human oversight for high-stakes applications, and balanced regulation that safeguards both societal safety and innovation. The study aims to examine the IAEA’s regulatory model as a template for AI governance and to compare it with evolving EU and US approaches.

The literature review surveys the evolution, scope, and influence of IAEA safety standards and procedures, and considers their relevance to AI governance. It notes the IAEA’s multi-tiered standards, continuous review, and international consensus processes as shaping global nuclear safety. Selected works highlight regulatory adaptation to new technologies and risks: Kim et al. (2022) on fusion regulation readiness; Kuznetsova & Fionov (2022) on cyber/information security at nuclear facilities; Garcia (2023) on inspection frameworks for digital I&C systems; and Anastassov (2016) on the effectiveness of the international nuclear safety regime and synergies between safety and security. The review then broadens to international AI regulation perspectives (EU, US) to contextualize potential cooperation with bodies like the IAEA. It identifies both challenges (rapid AI evolution; enforceability gaps) and opportunities (enhanced safety, risk detection, operational efficiency) for integrating AI into regulatory contexts, laying groundwork for applying IAEA-style approaches to AI.

The study employs a systematic review and comparative analysis. Data collection comprised: (a) IAEA regulatory documents, standards, and guidance; (b) AI technology standards and industry practices; and (c) international regulatory frameworks (EU, US, other bodies). Comparative analysis was used to examine alignments and gaps between IAEA regulations and AI standards. Thematic analysis identified recurring principles (e.g., transparency, continuous review, international consensus) pertinent to AI. The author developed an assessment framework with three pillars: (1) Regulatory fitness (fit of AI technologies within IAEA-style regulatory contexts); (2) Technical integration feasibility (practical integration of AI into nuclear facility operations, including risk assessment and monitoring); and (3) Regulatory efficacy (impacts on safety, security, and oversight). The methodology acknowledges limitations due to AI’s rapid evolution and diversity, suggesting careful interpretation and continuous updating of benchmarks.

- AI subgoal-setting can generate misaligned or unforeseen behaviors, posing safety, ethical, and societal risks; transparency and explicit guidance for subgoal mechanisms are essential. - Human oversight and real-time intervention are critical for high-stakes AI decisions; continuous monitoring and validation can mitigate unforeseen side effects. - The IAEA’s regulatory model offers transferable principles: a multi-tiered standards hierarchy; continuous review and renewal; broad international consensus; comprehensive systems spanning standard-setting, safety culture, emergency response, and materials management. - Proposed applications to AI include: standardizing AI behavior/learning/decision criteria; establishing a neutral, independent oversight body; conducting regular drills and protocols for AI-related incidents; and creating international platforms for information-sharing on standards, research, and incidents. - Regulatory landscapes differ: the EU advances comprehensive legislation (AI Act), while the US pursues executive and agency-led measures (e.g., Biden’s AI executive order, FTC oversight); global fora (e.g., UK’s Global AI Safety Summit) are shaping norms. - Case analyses illustrate risks: targeted ads over-collect personal data (privacy infringement); autonomous driving trade-offs can misprioritize safety; chatbots can propagate biases from training data. - Advantages of IAEA-referenced approach: harmonized standards, stakeholder participation, continuous updating. Disadvantages: domain differences between nuclear and AI, slow international consensus, potential conflicts with existing national regimes. - Overall, a tailored, flexible, internationally coordinated AI governance framework is needed, informed by IAEA lessons but adapted to AI’s unique characteristics.

Findings indicate that IAEA-style principles—standardization, continuous review, consensus-building, safety culture, emergency preparedness—can enhance AI governance when adapted for AI’s speed, opacity, and scope. The analysis shows these principles directly address the research question by offering a concrete pathway to safer AI: make subgoal-setting transparent, mandate oversight and incident preparedness, and foster international information-sharing. Significance lies in establishing a credible, cooperative model capable of reducing cross-border regulatory fragmentation, improving predictability and trust, and supporting responsible innovation. The discussion stresses modernizing existing frameworks to incorporate AI’s technical realities and the necessity of sustained international cooperation to maintain coherence across jurisdictions.

The paper contributes a structured mapping from IAEA nuclear safety practices to AI safety governance, identifying transferrable principles (tiered standards, continuous review, consensus, oversight, emergency response, safety culture) and proposing concrete AI applications (standardization, neutral oversight, drills, international sharing). It concludes that direct transplantation is insufficient due to AI’s distinct attributes (rapid evolution, complexity, broad societal effects). Future research should develop an AI-specific, adaptable framework; accelerate international consensus processes; harmonize with existing national and sectoral regulations; ensure stakeholder-inclusive governance; and maintain timely regulatory updates as AI advances.

- Domain mismatch: Nuclear and AI differ in risk profiles, system dynamics, and societal scope, limiting direct transferability of IAEA practices. - Pace and diversity of AI: Rapid evolution and heterogeneity challenge static or slow-moving international standard-setting. - Enforceability and interoperability: Achieving timely international consensus may be slow; alignment with existing national/sectoral regimes can create overlaps or conflicts. - Methodological constraints: Reliance on document review and conceptual analysis without empirical deployment data; findings require iterative validation as technologies and policies evolve.