AI Diplomacy: an infrastructural framework

AI Diplomacy: an infrastructural framework Mark Daley Draft for policy discussion December 8, 2025 Abstract Artificial intelligence (AI) is rapidly becoming a global infrastructure for cognition, coordination, and coercion. Yet the policy tools we are using to govern its cross-border impacts: science diplomacy, tech diplomacy, and classical arms control, were designed for a different technological era. They treat AI either as a sectoral topic, a bundle of discrete risks, or a generic “emerging technology” to be slotted into existing regimes. This paper argues that such inherited tools are necessary but insufficient. It develops a framework for AI diplomacy that is explicitly tailored to AI’s infrastructural character. The core contribution is a four-by-four framework organised around four layers of AI governance targets (resources, systems, applications, epistemics) and four diplomatic functions (coordination, containment, composition, contestation). The framework surfaces three hard problems that define the frontier of AI diplomacy: (a) epistemic governance and the challenge of building something like an “IPCC for AI” in a fast-moving, private-sector-led domain; (b) the status of AI systems as quasi-actors and the problem of legitimate delegation of diplomatic and coercive discretion; and (c) justice and composition for the Global South in a world where compute, data, and talent are highly concentrated. These are not simply descriptive gaps; they constitute a research and policy agenda. The final section briefly applies the framework to Canada as an illustrative case. As a middle power with strong AI research hubs, a bilingual and multicultural society, and a tradition of functional multilateralism, Canada is well placed to pursue an AI diplomacy agenda that leverages coordination and composition rather than relying on hard power. The Canadian case illustrates how the framework can guide the strategic positioning of actors that are neither AI superpowers nor rule-takers. 1 1 Introduction: Building AI Diplomacy with the Wrong Tools Most of the emerging practice labelled “AI diplomacy” is being built with inherited tools designed for different kinds of problems. Science diplomacy offers the now-canonical triad of “science in diplomacy, diplomacy for science, and science for diplomacy.” Tech diplomacy extends foreign policy outward to technology companies and “critical technologies.” Arms control and export-control regimes supply templates for constraining dangerous capabilities. All three are visible in contemporary initiatives around artificial intelligence: the Bletchley Declaration on AI safety, the Seoul Declaration for safe, innovative and inclusive AI, UNESCO’s Recommendation on the Ethics of AI, the OECD AI Principles, and the Council of Europe’s Framework Convention on AI, Human Rights, Democracy and the Rule of Law.1 These tools have real value. They have unlocked political attention, generated soft-law norms, and created club-style spaces where like-minded governments gather frontier labs, civil society, and international organisations to talk about safety and governance. They have also allowed states such as Canada, with strong research communities but limited hard power in chips and cloud infrastructure, to exercise influence in global debates. But the fit between these inherited tools and the object they are meant to govern is increasingly poor. Science diplomacy presumes that the main task is to mobilise and move scientific knowledge and instruments across borders, either in service of diplomatic goals (science in diplomacy), in support of scientific cooperation (diplomacy for science), or as a channel for building broader political relationships (science for diplomacy). Tech diplomacy and cyber diplomacy extend this logic to digital infrastructures and companies, treating platforms and network operators as quasi-sovereign actors and bastions of critical infrastructure. Arms control and export regimes focus on identifiable artefacts (e.g., missiles, enrichment facilities, dual-use chemicals, advanced chips) and the flows between them. Modern AI systems do not sit comfortably in any of these categories. They are not a bounded scientific field, a discrete sector, or a single class of weapons. Foundation models and the infrastructures that support them behave much more like thick, layered infrastructures of cognition and coordination. They mediate how individuals and institutions know, decide, 1For the Bletchley Declaration and subsequent AI safety summitry, see UK Government, “The Bletchley Declaration by Countries Attending the AI Safety Summit,” 2023. For the Seoul Declaration, see AI Seoul Summit outcome documents, 2024. For UNESCO and OECD instruments, see UNESCO (2021) Recommendation on the Ethics of Artificial Intelligence and OECD (2019, updated 2024) Recommendation on Artificial Intelligence. For the Council of Europe treaty, see the Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law, adopted 2024. 2 and act. They reconfigure labour and organisational processes. They amplify or dampen coercive capabilities in domains ranging from cyber operations to autonomous weapons and surveillance. In other words, AI is not just something diplomacy talks about; it is increasingly part of the infrastructure through which diplomacy itself is done, and part of the infrastructure that shapes the wider geopolitical environment diplomacy must navigate. Yet most accounts of “AI diplomacy” tend to graft AI onto existing categories. AI becomes one item on a list of emerging technologies, a line in a cyber-diplomacy toolbox, or a topic for science advisers in foreign ministries. Alternatively, it is narrowed to a handful of salient risks: lethal autonomous weapons systems, deepfakes, AI-enabled information operations. Here, we take a different starting point. We treat AI as a socio-technical infrastructure with at least three roles: •atopic of diplomatic negotiation (e.g., treaties, codes of conduct, export controls, standards); •atool in diplomatic practice (e.g., translation, analysis, foresight, generative drafting); and •a potential quasi-actor embedded in political and coercive systems (e.g., autonomous decision support in crises, agentic systems negotiating contracts or resource allocations). From that starting point, the question becomes: what would a field of AI diplomacy look like if it were designed for the infrastructural, multi-role reality of AI, rather than stretched around a conceptual template meant for the “big science” of the mid-twentieth century or for classical arms control? The answer offered here is a framework with two simple axes. Along one axis sit four layers of AI governance targets: the resource layer (compute, chips, data centres, energy, datasets), the system layer (models and training pipelines), the application layer (domain-specific uses), and the epistemic layer (how AI systems structure knowledge and public reasoning). Along the other axis sit four diplomatic functions: coordination, containment, composition, and contestation. The intersection creates a map of AI diplomacy as it is emerging and as it could be reshaped. Rather than offering a landscape review followed by a framework, the paper folds existing initiatives into the framework itself. The Bletchley Declaration and Seoul Declaration, for example, are read as coordination efforts focused primarily on the system layer, with some 3 nods to applications and little explicit treatment of resources or epistemic governance. Export controls on advanced semiconductors are treated as containment at the resource layer, largely decoupled from any reciprocal composition agenda for expanding access to safe compute in the Global South. UNESCO’s ethics recommendation and the Council of Europe AI Convention are seen as early experiments in epistemic and application-layer governance, still loosely connected to resource and system-layer bargains. The opening two sections develop this framework and show how it can organise, critique, and guide contemporary AI diplomacy. The third section uses the framework to identify three hard problems that define a forward-looking research and policy agenda. The final section applies the framework to Canada, a state whose AI diplomacy options are shaped by middle-power status, strong research assets, and a tradition of multilateral engagement. 2 A Four-by-Four Framework for AI Diplomacy 2.1 AI as infrastructural cognition, coordination, and coercion The case for treating AI as infrastructure rather than as a sectoral technology can be made in straightforward terms. First, large-scale AI systems are increasingly embedded in the basic operations of states, firms, and platforms. Natural-language models and recommender systems mediate communication, classification, and information retrieval. Vision systems classify and track objects and people. Planning and optimisation models structure logistics, finance, and resource allocation. These systems form part of a layered stack that resembles an infrastructural system more than a discrete product. Second, the marginal cost of deploying AI-mediated cognition is falling rapidly once the underlying models and infrastructure are in place. This enables new forms of centralisation and coordination: a small number of actors that control key models and platforms can shape the cognitive environment of very large populations. At the same time, open models and open-source tooling lower barriers for smaller actors to build or adapt systems that once required large corporate or state resources. Third, AI systems increasingly mediate coercive capabilities. They enable more efficient and scalable surveillance, more targeted information operations, and more autonomous or semiautonomous weapons systems. They are infused into cyber operations and strategic early warning. This creates new stability problems: learning systems may behave unpredictably 4 under distributional shifts or adversarial manipulation, and they may interact in complex ways across borders. An AI diplomacy worthy of the name must therefore treat AI as a layered infrastructure whose political and ethical properties emerge from interactions between resources, systems, applications, and epistemic effects. It must also recognise that AI is not merely an object of negotiation but an increasingly active mediator of the conditions under which negotiation occurs. 2.2 Four layers of AI governance targets The first axis of the framework identifies four analytically distinct but tightly coupled layers at which AI diplomacy operates: resources, systems, applications, and epistemics. 2.2.1 The resource layer: compute, chips, data, energy The resource layer covers the material and informational inputs that make large-scale AI possible: •advanced semiconductors and manufacturing tools; •cloud and data-centre capacity, including energy and cooling infrastructure; •large, high-quality datasets, often acquired via platform-scale data collection; •skilled labour and research talent. In current practice, this layer is governed largely through trade policy, export control, industrial policy, and investment screening. The United States and its allies have introduced export controls on high-end chips, fabrication equipment, and AI accelerators; China and others have responded with their own controls and industrial strategies. Countries with strong semiconductor industries negotiate supply-chain resilience and diversification. Data flows are contested through data protection laws, localisation requirements, and cross-border transfer arrangements. These are deeply diplomatic issues; they involve bargaining over access, reciprocity, compensation, and security assurances. They are not yet, however, framed as part of a coherent AI diplomacy agenda. They are treated as trade or security issues in which AI appears only as one justification among others. 5 Summitry such as Bletchley and Seoul touches the resource layer indirectly. Leaders emphasise capacity building and the risks of divergence between AI-rich and AI-poor countries, but concrete provisions for resource sharing (for example, shared compute facilities or subsidised access to cloud infrastructure) remain thin. The emerging Council of Europe convention on AI does not directly address compute access or semiconductor governance, even though adherence will be easier for states with strong infrastructures and harder for those without. An infrastructural AI diplomacy would bring the resource layer into the foreground. It would treat chips, compute, data, and energy as strategic resources whose allocation and denial are subject to explicit norms, not just unilateral control policies. 2.2.2 The system layer: models and training pipelines The system layer consists of models, training pipelines, and the software and organisational processes that surround them: data curation, training runs, evaluation, deployment, and post-deployment monitoring. This is where much of the current global activity around AI safety and governance is concentrated. The OECD AI Principles and their 2024 update articulate high-level values for “trustworthy AI” and give governments a vocabulary for talking about risk, robustness, accountability, and transparency. The Bletchley Declaration focuses specifically on “frontier” systems and commits states to identifying and mitigating catastrophic risks, including through evaluation, monitoring, and incident sharing. The Seoul Declaration builds on Bletchley and the G7 Hiroshima Process, emphasising interoperable governance, safety science, and cooperation with the private sector. At the same time, many of the concrete mechanisms at the system layer are being developed in club-like settings and in direct negotiation with companies. Governments and labs are creating safety institutes, pledging to publish risk management frameworks, and experimenting with voluntary commitments on evaluations and deployment thresholds. These are elements of coordination at the system layer, enabled by a mix of diplomatic and para-diplomatic channels. Yet gaps remain. Evaluation standards are fragmented. There is no shared understanding of what constitutes an unacceptable system-level capability or risk profile across different political communities. System-layer governance is mostly decoupled from resource-layer bargaining: commitments about how models are evaluated and deployed are not systematically linked to bargains over who gets access to training compute, Nvidia’s latest GPUs, or highquality models. 6 2.2.3 The application layer: domain-specific uses The application layer encompasses sectoral uses of AI in domains such as defence, security, health, finance, education, migration, and border management. At this layer, AI intersects with existing regulatory regimes and diplomatic processes. In defence and security, debates around lethal autonomous weapons systems (LAWS), military AI, and AI-enabled cyber operations are the most visible application-layer topics. States disagree over whether to pursue legally binding bans on fully autonomous weapons, how to define “meaningful human control,” and how to integrate AI into command and control without eroding strategic stability. In other domains, application-layer governance appears in more fragmented form: cross-border medical AI approvals, financial supervision using AI tools, risk-based frameworks for AI in migration and asylum processing. Much of this governance is domestic or regional, but it has clear cross-border implications. Countries must decide how to recognise each other’s approvals, how to share information on incidents, and how to handle differential access to AI-enabled services. International organisations, from NATO to the World Health Organization, are beginning to develop guidance, but an integrated picture of application-layer AI diplomacy is still emerging. 2.2.4 The epistemic layer: knowledge, narrative, and representation The epistemic layer refers to the ways in which AI systems structure knowledge, information flows, and public reasoning. It includes search and recommendation engines, generative models used for information retrieval and content creation, and AI tools deployed in education, media, and cultural production. Diplomatically, this layer is still under-theorised. Yet it is already central to international concern. States worry about AI-generated disinformation and deepfakes, about the erosion of shared factual baselines, and about the marginalisation of languages and cultures in global AI infrastructures. UNESCO’s Recommendation on the Ethics of AI is one of the few instruments that explicitly addresses cultural and linguistic diversity, media pluralism, and the risk of epistemic injustice, but implementation is uneven. The epistemic layer raises uncomfortable questions about who controls the models that increasingly mediate historical memory, news, cultural expression, and which thoughts your AI cognitive assistant encourages you to think. It exposes tensions between free expression, information sovereignty, and human-rights commitments. It also highlights the danger that a small number of models, trained primarily on data from certain regions and languages, 7 subtly encoding strong non-universal cultural priors, will become de facto arbiters of what counts as normal, plausible, or authoritative. An AI diplomacy that takes the epistemic layer seriously would treat these questions not as downstream side-effects but as core governance targets. 2.3 Four diplomatic functions: coordination, containment, composition, contestation Across these four layers, diplomacy performs at least four distinct functions. These functions cut across the science-diplomacy triad and re-orient attention from the roles of science toward the tasks of governance. 2.3.1 Coordination Coordination refers to efforts to make socio-technical systems interoperable, predictable, and mutually intelligible. In the AI context, it includes agreeing on: •shared definitions of risk categories and capability thresholds; •compatible evaluation, monitoring, and reporting practices; •interoperable technical and legal standards for deployment; •common reference points for concepts such as “trustworthy,” “human-centric,” or “highrisk” AI. Most of the high-profile summitry around AI (e.g., the Bletchley and Seoul Declarations, the G7 Hiroshima process) can be read as coordination at the system and application layers. They seek to align governments’ vocabularies, norms, and basic expectations without imposing detailed, binding obligations. They also begin to set expectations for the role of private labs and platforms in providing transparency, evaluation data, and early warning of dangerous developments. Coordination is necessary but not sufficient. If it is not connected to the resource and epistemic layers, it risks producing a club of states and firms that can comply with sophisticated coordination regimes while leaving others behind. 8 2.3.2 Containment Containment encompasses attempts to limit the propagation and use of dangerous capabilities. At the resource layer, this includes export controls on advanced semiconductors, chipmaking equipment, and large-scale compute infrastructure, as well as investment screening and cybersecurity measures to prevent theft of models or training data. At the system and application layers, containment appears as restrictions on training or deploying models that enable weapons development, cyber offence, or mass surveillance, and as attempts to keep AI out of the most sensitive parts of nuclear command and control. Containment logics are prominent in AI diplomacy. They are legible in the focus on “frontier” risks in safety summitry, in debates over LAWS, and in national security strategies that treat AI as a critical technology whose diffusion must be managed. They are also driving fractures in global technology ecosystems, as states seek to secure control over key resources and to deny them to adversaries. Containment is a familiar function for arms control and export-control bureaucracies. What is novel in AI is the coupling between resource and system-layer containment: restricting access to compute and chips becomes a central tool for controlling access to system capabilities. This blurs the line between economic coercion and security policy, raising questions about the conditions under which containment is legitimate and about the collateral impacts on third countries. 2.3.3 Composition Composition refers to the deliberate construction of shared infrastructures, institutions, and capacities. It is the least developed function in current AI diplomacy, yet it may be the most important if we want to avoid a future in which AI simply amplifies existing inequalities. In the AI context, composition could mean building: •shared public-interest models that serve as multilingual, culturally diverse epistemic resources; •common evaluation infrastructures and open benchmarks; •cooperative compute facilities accessible to countries and communities that cannot afford large national investments; •joint safety institutes and research programmes that pool expertise across borders; 9 Without such connections, there is a risk that AI diplomacy will resemble earlier episodes in global technology governance: a thin layer of inclusive rhetoric on top of entrenched asymmetries in infrastructure and rule-setting power. 4 Implications for Canada’s AI Diplomacy Agenda For Canada, the four-by-four framework is not an abstract exercise. It can be used to structure concrete strategic choices. Canada enters the AI diplomacy space with several distinctive assets and constraints. It was the first country to adopt a national AI strategy, anchored in the Pan-Canadian AI Strategy and centred on world-class research hubs in Montreal, Toronto–Waterloo, and Edmonton. It hosts leading academic and non-profit institutions in machine learning and AI governance. It is a G7 and OECD member with a track record in multilateral diplomacy but without the hard-power levers of a superpower. Canada has already participated in key AI diplomacy moments. It is a signatory to the Bletchley Declaration and the Seoul Declaration; it has championed voluntary codes of conduct for generative AI domestically; Global Affairs Canada is experimenting with AI tools for data analysis and foresight; and Canada has endorsed instruments such as UNESCO’s AI ethics recommendation. More recently, Canada has positioned itself as a convener on issues at the intersection of AI, critical minerals, quantum technologies, and science in general in G7 fora, and has launched digital and security partnerships that explicitly reference AI governance. Through the lens of the framework, these activities could be factored into an explicit AI diplomacy strategy. 4.1 Resource-layer strategy: from critical minerals to shared compute At the resource layer, Canada’s main levers are indirect. It does not host leading-edge semiconductor foundries, but it has significant deposits of critical minerals, strong renewable energy resources, and a trusted regulatory environment. It can use these assets in several ways. First, Canada can position itself as a partner in sustainable AI infrastructure: promoting low-carbon data centres, advancing standards for energy-efficient AI, and linking AI governance to climate and energy diplomacy. Second, it can integrate AI considerations into 16 critical-mineral diplomacy, recognising that batteries, data centres, and chips are part of an interconnected resource web. Third, Canada can champion and co-fund shared compute facilities and cloud-access schemes aimed at researchers and public institutions in the Global South, perhaps in partnership with multilateral development banks and regional organisations. Such initiatives would align containment and composition. Canada participates in allied export-control regimes and will continue to support measures to prevent diversion of advanced chips to hostile uses. At the same time, it can argue that legitimate containment requires complementary composition: mechanisms that ensure that safety-driven controls do not translate into permanent exclusion for entire regions. 4.2 System-layer strategy: safety, evaluation, and lab diplomacy At the system layer, Canada has both domestic and diplomatic opportunities. Domestically, it is in the process of shaping regulatory instruments which will define requirements for highimpact AI systems. Internationally, it sits in clubs where safety and evaluation practices are being articulated, such as the G7 Hiroshima process, OECD working parties, and AI safety summits. Canada can use this position to pursue three goals. First, it can invest in public and academic capacity for AI evaluation and red-teaming, ensuring that its regulators and researchers can independently assess claims made by powerful firms. Initiatives like the CIFAR Canadian AI Safety Institute Research Program are important steps in this direction. Second, it can support the development of common evaluation frameworks in multilateral settings, with special attention to including perspectives from outside the usual transatlantic circle. Third, it can practice “lab diplomacy” by building structured relationships with frontier AI labs that include expectations around transparency, incident reporting, and participation in safety research. Canadian diplomatic missions could, for example, include specialised AI attaches in key hubs, tasked with monitoring technical developments and feeding back into Ottawa’s risk assessments. NSERC and Global Affairs Canada could jointly fund fellowships that place technical experts in diplomatic teams and place policy experts in AI research labs, knitting together epistemic communities. 17 4.3 Application-layer strategy: sectoral leadership and restraint At the application layer, Canada’s choices include both positive and negative commitments. On the positive side, Canada can pilot responsible uses of AI in sectors where it has strong public institutions and regulatory capacity, such as health, education, and public administration. It can share lessons learned with partners and contribute to building templates for AI-enabled service delivery that respect privacy, equity, and human rights. On the restraint side, Canada can articulate clear red lines for AI use in defence and security, aligning with like-minded states on issues such as LAWS, autonomous targeting, and AI in nuclear command and control. It can push for transparency and confidence-building measures in military AI, including incident reporting and doctrinal exchanges. These positive and negative commitments are not just domestic choices; they are bargaining chips in AI diplomacy. They allow Canada to enter application-layer debates with credibility and to argue for standards that reflect both security concerns and human-rights commitments. 4.4 Epistemic-layer strategy: bilingual pluralism and public-interest models Canada’s bilingual and multicultural character, combined with its strong media and education institutions, gives it distinctive stakes at the epistemic layer. It has an interest in ensuring that global AI models properly support French and English, but also in promoting representation of Indigenous languages and perspectives. A Canadian AI diplomacy agenda could therefore prioritise support for public-interest models and datasets that reflect linguistic and cultural diversity, including Indigenous knowledge systems, developed under governance arrangements that respect Indigenous data sovereignty. It could contribute to UNESCO and other multilateral efforts to address disinformation, historical denial, and cultural homogenisation in AI-mediated media. Domestically, investments in public-service AI (e.g., publicly governed conversational agents that provide access to government services) could serve as demonstrations of an epistemic infrastructure that is not entirely dependent on private platforms. Internationally, Canada could lead coalitions calling for transparency from large platforms about how generative models treat content in different languages and from different cultural contexts. 18 5 Conclusion Artificial intelligence is forcing a rethinking of diplomatic practice not because it is an exotic new technology, but because it is becoming a pervasive infrastructure of cognition, coordination, and coercion. The inherited categories of science diplomacy, tech diplomacy, and arms control provide useful starting points, but they are not tailored to this infrastructural reality. The four-by-four framework developed in this paper — four layers of governance targets (resources, systems, applications, epistemics) crossed with four diplomatic functions (coordination, containment, composition, contestation) — offers one way to reframe AI diplomacy around what AI actually is. It helps make sense of current initiatives, identifies neglected domains, and generates concrete research and policy questions. The three hard problems highlighted here: epistemic governance, quasi-actors and delegation, and justice and composition for the Global South, are not the only ones, but they illustrate the kind of work that a serious AI diplomacy agenda must undertake. They require institutional innovation, imagination, and sustained investment, not just incremental tweaks to existing regimes. For Canada, and for other middle powers, the framework suggests that AI diplomacy is not a luxury. It is a domain in which they have comparative advantages in research, in multilateral convening, in normative entrepreneurship, and in which early moves can shape the infrastructures and norms that will govern AI for decades. More broadly, the framework is offered as an invitation to move beyond a narrow understanding of AI as a topic for expert panels and voluntary codes. If AI is an infrastructure of thought and action, then AI diplomacy is, in part, the diplomacy of the conditions under which we think and act together. Designing it well is not only a technical challenge; it is a constitutional one for the emerging machine-cognitive order. 19 Figure 1: An infographic summarizing the framework, courtesy of Google Gemini. 20