Redistributed

Between November 2023 and February 2025, three international summits attempted to establish the foundations of global AI governance. Bletchley Park, Seoul, Paris — the circuit traced a familiar arc of diplomatic ambition meeting institutional reality.

The summits produced declarations, voluntary commitments, and new safety institutes. They did not produce binding treaties, enforcement mechanisms, or any transfer of power from those building AI systems to those affected by them.

This is partly a function of how international diplomacy works — consensus-building precedes binding commitments, and the summits are early stages in what may become a more substantive governance architecture. But it is also partly a function of design: the summit circuit was structured in ways that made substantive governance unlikely. Understanding why requires examining not just what happened at each summit, but who was in the room, what "safety" was taken to mean, and whose interests the entire exercise served.

Three Summits, One Pattern

Bletchley Park: The Opening Move

The UK AI Safety Summit at Bletchley Park in November 2023 was orchestrated as a geopolitical statement. Prime Minister Rishi Sunak positioned Britain as a neutral convener — neither the regulatory assertiveness of the EU nor the market deference of the US, but a third way anchored in technical expertise.

Twenty-eight countries signed the Bletchley Declaration, acknowledging that AI poses risks requiring international coordination. The declaration was deliberately vague — a lowest-common-denominator document designed to get signatures rather than to compel action. China signed, which was presented as a diplomatic achievement. But signing a non-binding statement that AI "should" be safe is costless. The declaration committed no one to anything specific.

The summit's more concrete output was the announcement of the UK AI Safety Institute (AISI), tasked with evaluating frontier AI models. This was a genuinely significant institutional innovation — the first government body dedicated to evaluating frontier AI systems, and it has since attracted serious technical talent and produced substantive evaluations. But its mandate was narrow: testing models for catastrophic risks, not examining the broader social, economic, or distributional consequences of AI deployment. Safety was defined from the start as a technical property of frontier models, not a question about who AI systems harm and how.

The guest list told its own story. Frontier AI companies — OpenAI, Google DeepMind, Anthropic, Meta — had prominent seats. Civil society organizations were present but peripheral. Labor unions were largely absent. Representatives from the Global South were few and had minimal influence on the agenda. The people most affected by AI systems — workers facing displacement, communities subjected to algorithmic decision-making, populations in countries where AI is deployed but not developed — were not meaningfully represented.

Seoul: Voluntary Pledges

The Seoul AI Safety Summit in May 2024 narrowed the focus further. Where Bletchley Park at least gestured toward broad risk categories, Seoul concentrated on frontier AI safety commitments — voluntary pledges from the companies building the most powerful models.

Sixteen leading AI companies signed the "Frontier AI Safety Commitments," promising to assess risks before deploying frontier models, to share safety information with governments and other companies, and to develop internal governance structures for managing catastrophic risks. The commitments were voluntary. There was no verification mechanism, no consequences for non-compliance, and no independent audit process.

This is a familiar pattern in technology governance. Industry, facing the prospect of regulation, offers self-regulation instead. The promises are genuine in the moment — companies do establish safety teams, do conduct evaluations, do publish responsible use policies. But voluntary commitments are structurally fragile. They hold when they are costless and erode when they conflict with competitive pressure or shareholder expectations. As Virginia Eubanks has documented in the context of automated public services, the gap between what institutions promise and what they deliver is widest for those with the least power to hold them accountable (Eubanks, 2018).

Seoul also expanded the AI Safety Institute concept, with South Korea and several other countries announcing their own safety institutes, modeled on the UK's AISI. The proliferation of safety institutes is presented as progress. But a network of institutes focused on testing frontier models for catastrophic capabilities does not constitute governance. It constitutes an evaluation infrastructure — useful, but fundamentally limited to the questions it is designed to ask.

Paris: Broader Ambitions, Familiar Limits

The Paris AI Action Summit in February 2025 attempted a course correction. French President Emmanuel Macron positioned Paris as a counterweight to the narrow safety focus of previous summits, emphasizing AI for public good, inclusive development, and Global South participation.

The summit was larger and more diverse. Over sixty countries participated, alongside international organizations, civil society groups, and representatives from developing nations. The agenda expanded beyond frontier risk to include AI and sustainability, open-source AI, cultural diversity, and equitable access to AI benefits.

But expansion created its own tensions. The broader agenda meant less depth on any single issue. Countries disagreed about scope — the US and UK pushed to maintain the frontier safety focus, while France, developing nations, and many civil society groups argued for addressing present harms and economic concentration. The resulting summit declaration papered over these disagreements with diplomatic language that committed participants to "pursuing" various goals without specifying mechanisms, timelines, or enforcement.

The Paris summit also revealed the limits of the summit format itself. Diplomatic events are designed to produce consensus statements. Consensus, in contexts where participants have fundamentally different interests, means ambiguity. And ambiguity means that after the summit ends, each party interprets the commitments to align with what they were already doing.

What "Safety" Means — And for Whom

The most consequential feature of the summit circuit is not what it decided but what it defined. The summits established "AI safety" as the organizing concept for international AI governance. That framing is not neutral.

"Safety," as deployed in the summit context, primarily means the risk that frontier AI models develop dangerous capabilities — bioweapon synthesis, autonomous cyberattacks, manipulation at scale, or the speculative scenario of misaligned artificial general intelligence. This is the x-risk framing, heavily influenced by the Effective Altruism movement and organizations like the Centre for AI Safety, the Future of Life Institute, and the AI safety teams within frontier labs themselves.

The x-risk frame is not wrong — the possibility that increasingly capable AI systems could cause catastrophic harm is worth taking seriously. But it is radically incomplete. As AI ethics researchers including Timnit Gebru, Emily Bender, and Kate Crawford have argued, the most urgent AI harms are not speculative. They are happening now: algorithmic discrimination in hiring, lending, and criminal justice; mass surveillance enabled by facial recognition and predictive analytics; labor displacement and wage suppression; the extraction of training data from creators without consent or compensation; and the concentration of economic power in a handful of companies (Bender et al., 2021; Crawford, 2021).

The framing battle matters because it determines resource allocation and political attention. Every hour of summit time spent on superintelligence scenarios is an hour not spent on the algorithmic systems currently denying people welfare benefits, jobs, or parole. Every safety institute focused on frontier model capabilities is an institution not focused on auditing the AI systems already deployed in workplaces, welfare offices, and police departments.

This is not an accident. The x-risk frame serves the interests of frontier AI companies in at least two ways. First, it focuses governance attention on a small number of models built by a small number of companies — making governance tractable but also ensuring that the governed are the ones defining the terms. Second, it implicitly argues for concentration: if the primary risk is that powerful AI escapes human control, then the solution is to ensure that only well-resourced, responsible organizations build powerful AI. Safety becomes an argument for barriers to entry. As Gebru has noted, the "AI safety" discourse can function as a form of regulatory capture, where the companies most in need of oversight are the ones shaping how oversight works (Gebru, 2024).

The Table and Its Absences

International summits are exercises in inclusion and exclusion. Who gets invited determines what gets discussed.

The AI safety summit circuit has been dominated by three groups: governments of wealthy nations, frontier AI companies, and a select group of academics and researchers — predominantly from institutions in the US and UK. This is not a representative sample of those affected by AI.

Civil society organizations have been present but marginal. They attend side events, issue statements, and occasionally speak on panels. But they do not set the agenda, draft the declarations, or shape the commitments. Their role remains largely advisory — and the trend from Bletchley to Paris, while moving in the direction of broader inclusion, has not yet translated into genuine agenda-setting power for non-industry voices.

The Global South has been progressively more included — Paris was better than Seoul, which was better than Bletchley Park. But inclusion in a room is not the same as influence over outcomes. Developing nations face AI systems built elsewhere, trained on data extracted from their populations, deployed by multinational companies with minimal local accountability. Their primary concern is not whether GPT-5 might develop dangerous capabilities but whether algorithmic systems are making consequential decisions about their citizens without transparency, consent, or recourse.

Labor representatives have been almost entirely absent. This is striking given that labor displacement is one of the most concrete and immediate effects of AI deployment. Workers whose jobs are being automated, whose wages are being suppressed by AI-enabled management systems, whose working conditions are being shaped by algorithmic scheduling — these are direct AI harms. But the summit circuit treats labor as an economic side effect, not a governance question.

Affected communities — people subject to algorithmic decision-making in welfare, criminal justice, immigration, and employment — have no representation at all. The summits are about those who build AI, not those who live under it.

Langdon Winner's foundational question — "do artifacts have politics?" — applies to the summits themselves (Winner, 1980). The structure of the room embeds assumptions about whose knowledge counts, whose interests matter, and what governance is for. A summit organized by and for frontier AI companies will naturally produce governance frameworks that serve frontier AI companies.

The Overton Window

Perhaps the most lasting effect of the summit circuit is its influence on what counts as a legitimate AI governance concern.

The summits have established a hierarchy of issues. At the top: existential risk, frontier model safety, international coordination on catastrophic capabilities. In the middle: AI ethics, bias, and responsible development — acknowledged but not the focus of binding commitments. At the bottom, barely visible: labor displacement, economic concentration, data rights, algorithmic accountability for systems already deployed, and the redistribution of AI-generated value.

This hierarchy shapes national policy. Governments that participate in summits allocate resources accordingly. The UK invested heavily in the AI Safety Institute while doing comparatively little about algorithmic decision-making in public services. The pattern is repeated across participating countries: frontier safety gets institutions and funding; present harms get reports and recommendations.

The hierarchy also shapes public discourse. Media coverage of AI governance focuses on the summit circuit's preferred topics — will AI destroy humanity? How do we ensure advanced AI is aligned with human values? — while receiving less coverage are the questions that affect people now: Why did the algorithm deny my loan? Why was I not hired? Why is my rent increasing because of algorithmic pricing?

What Would Governance Actually Require

The summit circuit is not governance. It is the diplomatic precursor to governance — the stage where norms are discussed, relationships are built, and the parameters of future agreements are explored. This is not useless. But it should be recognized for what it is and what it is not.

Meaningful international AI governance would require several things the summits have not provided.

Binding obligations. Declarations and voluntary commitments are not governance instruments. Governance requires rules with consequences for violation. The closest analogy is international environmental law — not the Paris climate pledges (which are voluntary and insufficient) but frameworks like the Montreal Protocol, which imposed binding obligations, established compliance mechanisms, and actually worked. An international AI governance regime would need similar features: specific obligations, monitoring mechanisms, and consequences for non-compliance.

Democratic accountability. The summits are diplomatic events — governments negotiating with governments, with industry at the table. Democratic AI governance would require public participation: citizens' assemblies deliberating on AI policy, as Ireland used for constitutional questions; participatory processes that give affected communities standing, not just a seat in the audience; and transparency requirements that enable informed public debate rather than expert-only discussion.

Institutional infrastructure. Governance requires institutions with authority, expertise, and resources. Not just safety institutes that test frontier models, but regulatory bodies that can audit deployed systems, investigate complaints, compel disclosure, and impose penalties. The EU AI Act moves in this direction, but it remains to be seen whether implementation will match ambition.

Scope expansion. Governance that addresses only frontier model capabilities while ignoring the AI systems already shaping lives is governance of the future at the expense of the present. A comprehensive framework would encompass algorithmic accountability, data rights, labor protections, competition policy, and the distribution of AI-generated economic value.

Who Benefits from the Status Quo

The summit circuit, in its current form, serves specific interests.

Frontier AI companies benefit from a governance conversation focused on their products, conducted on their terms, producing voluntary commitments they define and self-assess. They benefit from the implicit argument that safety requires concentration — that only large, well-funded labs can build AI "safely." And they benefit from the Overton window that keeps labor displacement, data extraction, and economic concentration off the main stage.

Wealthy nations benefit from a framework that positions them as conveners and standard-setters while the Global South participates on terms set by others. The summit circuit reinforces existing geopolitical hierarchies in technology governance.

The diplomatic class benefits from a new issue area that generates summits, declarations, and institutions — the infrastructure of international engagement, regardless of outcomes.

Who does not benefit: workers facing displacement without voice in the process. Communities subjected to algorithmic systems without accountability mechanisms. Nations where AI is deployed but not developed. Researchers and advocates pushing for attention to present harms rather than speculative futures.

The redistribution question is stark. The AI safety summit circuit is shaping the rules that will govern one of the most consequential technologies in human history. Those rules are being written by and for the most powerful actors in the AI ecosystem. Unless the process changes — unless the table expands, the agenda broadens, and the commitments bind — the result will be governance that protects the interests of those who least need protecting.

The summits are not irrelevant. They have created communication channels between governments and AI developers that did not exist before, seeded institutions like the AI Safety Institutes, and normalized the idea that AI governance is a legitimate international concern. These are real achievements. But they are the wrong tool presented as the only tool. Real governance requires more: more voices, more accountability, more binding force, and a definition of "safety" that includes the people AI systems are already harming.

Sources

Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 2021. ACM
Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Publisher
Eubanks, Virginia. Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin's Press, 2018. Publisher
Winner, Langdon. "Do Artifacts Have Politics?" Daedalus 109, no. 1 (1980): 121-136. JSTOR
UK Government. "The Bletchley Declaration by Countries Attending the AI Safety Summit." November 2023. GOV.UK
Republic of Korea. "Frontier AI Safety Commitments, AI Seoul Summit 2024." May 2024. GOV.UK
Gebru, Timnit. "Effective Altruism Is Pushing a Dangerous Brand of 'AI Safety.'" Wired, 2024. Wired

General

The AI Safety Summit Circuit: Diplomacy or Theater?