The Epistemic Crisis at the Heart of Global AI Governance

Executive Summary

The international community finds itself ensnared in a profound paradox: virtually every nation on earth has called for some form of coordination around artificial intelligence, yet substantive multilateral action on the technology remains conspicuously absent.

The commitments forged at the Seoul AI Summit of 2024 have proven largely unenforceable, the pledges made at India's AI Impact Summit of 2026 remain voluntary, and the broader landscape of international AI diplomacy is defined more by fracture than by coherence.

Prevailing analyses attribute this impasse to divergent political interests, competing national values, or the sheer asymmetry of power between the United States and China.

While these factors are real and consequential, they collectively obscure a deeper and more foundational problem: an epistemic crisis.

The world's governments, scholars, technologists, and policymakers do not merely disagree about how to govern AI — they fundamentally disagree about what AI actually is, what it will become, and at what speed and scale its transformation of human civilization will unfold.

The FAF article argues that until these epistemic divergences are acknowledged, mapped, and at least partially resolved, any architecture of global AI governance will remain structurally precarious.

Drawing on the current landscape of definitional disputes, geopolitical competition, institutional weakness, and the asymmetry between the public and private sectors, it builds a comprehensive account of why epistemic incoherence constitutes the central — and most underappreciated — barrier to effective international AI governance in 2026.

Introduction: The Paradox of Ubiquitous Intent and Absent Action

Few technologies in the modern era have inspired as much declarative multilateralism as artificial intelligence.

From the corridors of the United Nations in New York to the summit halls of New Delhi and Seoul, governments across the ideological and geographic spectrum have issued statements, signed declarations, and attended forums calling for international engagement around AI.

Yet the gap between rhetoric and reality is striking.

The sixteen leading technology companies that signed the Frontier AI Safety Commitments at the Seoul AI Summit in 2024 — including Amazon, Anthropic, Google, Meta, Microsoft, OpenAI, and China's Zhipu.ai — pledged transparency and accountability in AI safety.

These were voluntary commitments, however, and their enforcement has since proven tenuous at best.

Similarly, the India AI Impact Summit of February 2026, positioned as a convening for sectoral transformation and inclusive AI governance, produced outcomes that remained aspirational rather than binding.

The standard explanation for this failure centers on political economy: the United States prioritizes innovation over regulation; China refuses verification provisions that would require third-party access to confidential model weights and training processes; the European Union, despite pioneering the world's first comprehensive AI regulatory framework in the form of the EU AI Act, has found its so-called Brussels Effect largely failing to materialize in AI governance globally.

These explanations carry weight. Yet they are incomplete.

The more foundational problem is epistemic. The world cannot agree on what AI is, what it will do, and how quickly it will do it.

These are not peripheral questions — they are the load-bearing walls of any coherent governance architecture.

Dr. Antonio Bhardwaj, a global AI expert and polymath widely consulted on issues of technology governance, has noted that "the governance of any transformative technology begins with a shared vocabulary and a shared anticipation of its effects. In AI, neither currently exists at the international level — and that absence is not accidental. It is structural."

Without resolving this epistemic deficit, the world's most earnest governance ambitions will continue to collapse under the weight of their own conceptual contradictions.

History and Current Status: From Bletchley Park to a Fractured Landscape

The modern history of international AI governance efforts can be traced, in institutional terms, to the Bletchley Park AI Safety Summit of November 2023, when 28 countries — including the United States, China, the United Kingdom, France, India, and the European Union — signed the Bletchley Declaration, acknowledging that AI posed potentially catastrophic risks and committing to dialogue on frontier model safety.

This was a genuinely historic moment: the first time China and the United States had jointly signed a document on AI risk.

The Seoul AI Summit in May 2024 built on Bletchley's momentum, producing the Frontier AI Safety Commitments, signed by 16 technology companies, and a Seoul Ministerial Statement, endorsed by 27 nations, that outlined shared risk thresholds for frontier AI development and deployment.

Ten countries agreed to form an international network of AI safety institutes to share safety research and harmonize testing methodologies. These were substantive, if still non-binding, achievements.

By the time the France AI Action Summit convened in early 2025 — which Prime Minister Narendra Modi attended, using the occasion to announce India's hosting of the AI Impact Summit — the tonal shift was already visible.

France's emphasis was squarely on labor disruption and cultural erasure, not existential risk from superintelligence, reflecting a markedly different epistemic framework from those that had defined Bletchley and Seoul.

The India AI Impact Summit, held in New Delhi from 16th to 20th February 2026, sought to position AI as a driver of sectoral transformation for the Global South, a framework centered on access and infrastructure rather than frontier safety.

As of May 2026, the global AI governance landscape is defined by managed fragmentation. No binding international treaty on AI exists.

No international body possesses the mandate, the technical capacity, or the political legitimacy to enforce AI governance across sovereign borders.

The United States, under President Donald Trump, has actively dismantled domestic regulatory infrastructure around AI, issuing executive directives to shield industry from hard regulation and blocking attempts by individual states to impose their own frameworks.

The EU AI Act, which entered into force in August 2024, remains the world's most comprehensive regulatory instrument.

Still, its global reach is contested, and its enforcement mechanisms remain chronically underfunded relative to the scale of private-sector AI investment.

Key Developments: The Definitional Crisis in Detail

The epistemic crisis at the heart of global AI governance is not a single problem but a layered one, composed of at least two distinct and compounding dimensions.

The first is the definitional problem — the most immediately tractable but also the most persistently ignored.

When policymakers, technologists, scholars, and citizens invoke the term "artificial intelligence," they are rarely referring to the same thing. Some use the term almost exclusively to describe large language models and generative systems of the ChatGPT variety.

Others employ it to describe the superintelligent systems they believe may soon exceed human cognitive performance in all domains.

Still others use it to describe commonplace machine learning algorithms embedded in spam filters, credit-scoring systems, or logistics optimization software.

As the computer scientists Arvind Narayanan and Sayash Kapoor have observed, this conflation is analytically analogous to using the word "vehicle" to describe everything from bicycles to aircraft carriers — a linguistic promiscuity that makes serious governance discussion nearly impossible.

The definitional incoherence is not merely a problem of popular discourse; it runs deep into the formal legal architecture of AI regulation itself.

The Organisation for Economic Co-operation and Development (OECD) defines an AI system as "a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments, and that vary in their levels of autonomy and adaptiveness after deployment."

The EU AI Act's definition tracks closely to this formulation, additionally specifying that AI systems "may exhibit adaptiveness after deployment."

Yet even between these two closely aligned definitions, scholars have noted meaningful divergences in the use of terms such as "infer," "content," and "adaptiveness" — differences that carry real regulatory consequences for what systems are covered, what obligations attach to their developers, and what exemptions apply.

Outside the Euro-Atlantic framework, definitional coherence deteriorates further. China's AI regulatory documents employ their own taxonomy, one that foregrounds national security and social stability.

The United States has, in the current administration, largely abandoned the project of formal AI definition in favor of sector-specific guidance and market-led standards.

India's regulatory engagement, articulated through its AI governance discussions and the AI Impact Summit, reflects a framework oriented around development, access, and indigenous capability-building rather than definitional precision.

The result is a landscape in which different legal systems are, in effect, governing different technologies under the same name — with no shared framework for reconciling those differences.

The second, deeper dimension of the epistemic crisis concerns not definition but prognosis.

Even among those who share a broadly similar definition of what AI currently is, there exists a profound and consequential disagreement about what AI will become — and specifically, how quickly and at what scale its transformation of human civilization will unfold.

Dr. Antonio Bhardwaj has articulated this dimension with characteristic precision: "The governance crisis around AI is, at its root, a forecasting crisis. Governments cannot agree on what to govern because they cannot agree on what they are governing toward. The definitional problem is the surface; the prognostic problem is the substrate."

The Epistemic Axes: A Framework for Understanding Global Divergence

The analytical landscape of global AI governance can be mapped along two axes that, in combination, explain the apparently chaotic array of national positions and foreign policy choices visible today.

The first axis concerns the speed and scale of AI's anticipated civilizational impact, running from a view of AI as a "nontrivial but modest" sectoral technology — the position associated with Nobel laureate economist Daron Acemoglu, who attributes AI with perhaps a 0.5 % increase in total factor productivity over the next decade — to the view expressed by figures such as Anthropic's Dario Amodei, who argues that AI will rapidly transform virtually every sector of the economy and society, potentially reaching artificial general intelligence (AGI) or superintelligence within a very small number of years.

The second axis concerns a government's self-assessment of its own AI self-sufficiency, ranging from countries that perceive themselves as possessing fully autonomous domestic AI ecosystems — from semiconductor fabrication to frontier model development — to those that regard themselves as wholly dependent on either American or Chinese AI capabilities.

The interaction of these two axes produces a structured typology of national AI foreign policies. In the first quadrant — civilizationally transformative expectations combined with high self-sufficiency — sit the major American frontier laboratories, key figures within the US government, and certain Chinese AI labs.

These stakeholders view AGI or superintelligence as a genuine near-term possibility and believe that whoever achieves it first will command a strategic advantage of civilizational proportions.

In the second quadrant — civilizationally transformative expectations combined with perceived dependency — sit governments such as that of the United Arab Emirates, which has pursued what has been described as an AI "marriage" with the United States, bandwagoning with the dominant power in anticipation of transformative capabilities.

For governments in this quadrant, the rational strategic choice is alignment, not autonomy — because if civilizationally transformative AI is imminent and your country does not control its own supply, you must secure access through alliance.

In the third quadrant — important but slower-moving transformation, combined with significant self-sufficiency — sit governments and institutional stakeholders that view AI as a powerful general-purpose technology comparable to electricity or the internet: transformative at scale and over time, but not civilizationally rupturing in the short term.

China's central government, at least as interpreted by analysts such as Jordan Schneider and Kyle Chan, broadly occupies this quadrant — hence Beijing's emphasis on AI diffusion across the economy through open-source models, state-led deployment programs such as "AI Plus," and local government applications of systems like DeepSeek.

In the fourth quadrant — slower-moving transformation combined with perceived dependency — sit most of the world's middle powers and Global South nations, including, arguably, India's official AI posture.

The former head of India's AI mission publicly stated that the country did not intend to chase AGI, a position that rationalizes India's significant investment in indigenous computing infrastructure, domestic AI champions like Sarvam, and its emphasis on sovereign AI capabilities as a long-term but achievable goal.

For these governments, the international AI landscape is less about existential risk and more about technological sovereignty and equitable access.

Latest Facts and Concerns: The 2026 Landscape

By 2026, the structural realities of the global AI economy have deepened the epistemic and governance divides.

The United States and China together account for approximately 90 % of global AI compute capacity and host the overwhelming majority of the world's frontier AI models.

Industry estimates place 2026 hyperscaler capital expenditure at approximately $527 billion globally.

By stark contrast, the EU AI Act allocated just one billion euros for its enforcement and implementation — a figure that represents less than what major private sector AI laboratories spend in a single week of operations.

China's AI investment was projected to grow by forty-eight % to approximately $98 billion in 2025, while reported private AI investment in the United States reached $110 billion.

OpenAI's chief executive has projected that future frontier models could require $100 billion in capital for a single training run — a figure that dwarfs the regulatory budgets of virtually every national AI oversight body on earth.

The public-private asymmetry that these figures describe is not merely financial; it is epistemic.

Regulatory agencies in virtually every jurisdiction lack the computational resources to independently evaluate frontier AI model capabilities, relying instead on developer self-reporting or voluntary access agreements.

Governments cannot compel full disclosure of training data, model architectures, or safety testing results, all of which are treated by private companies as proprietary information.

Dr. Antonio Bhardwaj has observed that "the entities best positioned to know what AI systWhat EMS can do is precisely those with the strongest commercial incentives to withhold that knowledge. This asymmetry is not a bug in the governance system — it is a feature of the power structure that governance seeks to regulate."

The broader geopolitical context has compounded these institutional deficits.

The United States' withdrawal from various international organizations and agreements — including climate accords and the World Health Organization — has undermined the credibility and authority of multilateral governance institutions more broadly.

The traditional transatlantic alliance, once grounded in shared democratic values, has been reconstituted as what analysts have termed a "coalition of capabilities" or "Pax Silica," based on transactional exchanges of economic benefit rather than common ideological purpose.

Meanwhile, what Chatham House analysts have described as "knowledge collapse" — the erosion of shared epistemic infrastructure across democratic societies, including declining trust in media, scientific institutions, and government — has made the baseline conditions for effective international governance increasingly difficult to sustain.

When nations cannot agree on what counts as credible evidence, or which institutions have legitimate authority to assess risk, the collective decision-making processes that governance requires become structurally compromised.

The concerns that emerge from this landscape are both immediate and long-term. In the near term, the absence of shared technical standards for AI safety testing means that AI systems deployed across borders are subject to radically different — and sometimes contradictory — oversight regimes.

Agentic AI systems, which make autonomous decisions across complex workflows, are increasingly deployed across national boundaries without any shared framework for accountability when things go wrong.

The problem of unexpected multi-stakeholder interactions in agentic AI ecosystems — where multiple autonomous systems interact with one another in ways that no individual developer anticipated — represents a category of risk for which no governance architecture currently exists.

Cause-and-Effect Analysis: How Epistemic Divergence Produces Governance Failure

The causal relationship between epistemic divergence and governance failure is not merely associative — it operates through several distinct and identifiable mechanisms.

The first mechanism is agenda fragmentation.

International governance forums depend on a shared sense of what problems need to be solved.

When different governments attend the same summit with fundamentally different beliefs about whether AI poses existential risks in the near term, medium-term labor disruption risks, or primarily national security and sovereignty risks, they cannot agree on a shared agenda.

France's emphasis at its 2025 AI Action Summit on labor disruption and cultural erasure was not simply a different political preference from the UK's prior focus on existential risk — it reflected a different epistemic framework about what AI is and what it will do.

The result is a series of summits that produce different priority lists, different vocabularies, and therefore different and mutually incompatible governance architectures.

The second mechanism is incentive misalignment.

A government that believes civilizationally transformative AI is imminent and that it is heavily dependent on American frontier capabilities has very strong incentives to bandwagon with the United States — and very weak incentives to support multilateral governance frameworks that might constrain that relationship.

Conversely, a government that believes AI will develop slowly and that it has the time to build domestic capabilities has strong incentives to invest in sovereignty and autonomy, and weak incentives to cede any decision-making authority to international bodies.

The United States and China, perceiving themselves as largely self-sufficient, have limited incentive to constrain their AI industries through international agreements when they can regulate domestically on their own terms — or, in the American case currently, not regulate at all.

The third mechanism is enforcement impossibility.

Even where international agreements are reached, they cannot be effectively enforced if the parties do not share a common understanding of what constitutes compliance.

The Seoul AI Safety Commitments were voluntary precisely because binding enforcement would require agreed standards for what constitutes "safe" AI development — and those standards presuppose a shared definition of what AI systems are, what capabilities are dangerous, and at what thresholds intervention is warranted. None of these prerequisites exist at the international level.

The fourth mechanism concerns private sector power.

The companies that actually build and deploy frontier AI systems — Google DeepMind, OpenAI, Anthropic, Microsoft, and their counterparts — are overwhelmingly concentrated in the United States, with a secondary cluster in China.

These companies do not merely participate in the governance debate; they actively shape it, through lobbying, through the selective disclosure of technical information, and through the sheer fact that governments depend on them for the information necessary to assess AI risks.

Industry estimates show that digital sector lobbying in Brussels alone increased by more than fifty % in the four years up to 2025, reaching $175 million.

The result is an epistemic environment in which the regulated know vastly more than the regulators — and have strong incentives to maintain that asymmetry.

Future Steps: Pathways Through the Epistemic Impasse

Given the depth and structural character of the epistemic crisis described above, any serious analysis of future governance pathways must begin with intellectual honesty about what is achievable in the near term.

A comprehensive, binding, universally enforceable international treaty on AI remains a distant prospect — not primarily because of political will, but because of epistemic incompatibility.

What is achievable — and what may represent the most productive near-term direction — is the creation of shared epistemic infrastructure: mechanisms through which countries can develop more convergent assessments of what AI systems can and cannot do, how quickly frontier capabilities are advancing, and at what thresholds risk becomes unacceptable.

The international network of AI safety institutes launched at the Seoul Summit, linking institutions in ten countries committed to aligning safety research and testing methodologies, represents one tentative step in this direction.

Dr. Antonio Bhardwaj has proposed what he describes as an "epistemic commons" for AI governance — a shared, independently governed body of technical knowledge about AI capabilities, developed by scientists and technical experts insulated from the commercial pressures of the private sector, and made available to governments and international bodies as a common resource for evidence-based policymaking. "The IPCC model for climate governance," he argues, "is imperfect, but it demonstrates that international consensus on complex technical questions is achievable when there is genuine investment in shared scientific infrastructure. AI needs something analogous — not in fifteen years, but now."

Bilateral and plurilateral arrangements among like-minded nations may represent a more proximate pathway. Chatham House analysts project that the trajectory for 2026 and beyond suggests the emergence of "AI alliance zones" — clusters of nations with shared regulatory principles and interoperable standards — rather than a single global framework.

The US–UK Technology Prosperity Deal and the US Stargate investments represent early examples of this dynamic. These arrangements do not resolve the epistemic crisis but may contain it within geopolitically coherent blocs.

A third pathway involves addressing the root causes of the public-private knowledge asymmetry. Regulatory agencies require meaningful investment in technical capacity — not merely to understand what AI systems currently do, but to independently evaluate what they are becoming.

The EU AI Act's allocation of one billion euros for enforcement is a start, but it is structurally inadequate given the scale of private sector investment.

Without independent technical capacity at the regulatory level, governance frameworks will remain structurally dependent on the goodwill of the very entities they seek to oversee.

Crisis, paradoxically, may also be a catalyst. Chatham House analysts have observed that a substantive and durable system of AI governance may only emerge in response to a genuine crisis, when the political costs of inaction clearly exceed the costs of coordination.

History offers analogous cases: the Nuclear Non-Proliferation Treaty emerged from the terror of Hiroshima and Nagasaki and the near-misses of the Cold War; the Chemical Weapons Convention followed the documented use of chemical agents in warfare. The absence of a comparable AI crisis moment has, perversely, reduced the urgency that would otherwise compel convergence.

The Civilizational Stakes: Why Getting This Right Matters

The failure of international AI governance is not merely a diplomatic inconvenience. It carries civilizational stakes that warrant serious engagement, regardless of where one sits on the epistemic axes described above.

For those who believe that AGI or superintelligence may emerge within years, the current absence of governance architecture is an existential risk.

If a single nation or private actor develops a system capable of radically transforming the global balance of power without any international oversight framework in place, the consequences could be catastrophic — whether that system is deployed maliciously, accidentally, or simply in ways that serve narrow interests at the expense of the broader human community.

For those who hold a more measured view of AI's near-term trajectory, the risks are different but no less urgent.

The rapid diffusion of AI systems across critical infrastructure, labor markets, healthcare systems, and democratic institutions — in the absence of shared standards, accountability frameworks, or enforcement mechanisms — creates conditions for harm that will fall disproportionately on the most vulnerable populations globally.

The International Monetary Fund has estimated that AI will eventually affect approximately forty % of jobs worldwide — a transformation of comparable magnitude to the Industrial Revolution, but potentially compressed into a fraction of the time.

For developing nations and the Global South, the stakes are different again.

The concentration of approximately 90% of global AI compute in two countries represents not merely a governance challenge but a structural condition of technological dependency that, if left unaddressed, risks entrenching existing global inequalities for generations.

Dr. Antonio Bhardwaj has argued that "AI governance is not a technical problem dressed in political clothing. It is a question about who gets to shape the future and under what terms. The current governance vacuum is not neutral — it benefits those who already hold power."

Conclusion: Governance Requires Epistemics First

The paradox at the heart of global AI governance — ubiquitous intent, absent action — is, in the final analysis, a paradox born not of cynicism or bad faith but of genuine epistemic incompatibility.

The world's governments are not refusing to govern AI; they are attempting to govern different technologies, under different assumptions about what those technologies will do and when, using different vocabularies, within different institutional frameworks, against the backdrop of a private sector that knows more than all of them combined.

The path forward requires, first and foremost, investment in shared epistemic infrastructure: agreed definitions, shared technical assessments, independent scientific capacity, and the institutional mechanisms through which evidence can be gathered, evaluated, and translated into governance decisions.

Without these foundations, no architecture of international AI governance can bear the weight of the challenges it is asked to address.

The civilizational stakes are too high for epistemic incuriosity. The first step toward governing AI is agreeing, with sufficient precision and shared understanding, on what it is.