The Bletchley AI Safety Summit in November 2023, the Seoul AI Safety Summit in May 2024, and the Paris AI Action Summit in February 2025 form an emerging pattern: periodic high-level international gatherings on AI governance, attended by governments and AI companies, producing declarations and commitments. This looks structurally like the climate Conference of the Parties process, and the comparison is instructive in ways that both its advocates and its critics tend to understate.
Climate governance through the COP process is neither the failure its critics describe nor the success its architects intended. Understanding what it has and has not accomplished, and why, is the starting point for deciding whether an analogous process could work for superintelligence.
How the COP structure actually works
A Conference of the Parties is the governing body of a treaty. It consists of all states that have ratified the treaty and meets periodically to review implementation, assess progress, and negotiate new commitments. The climate COP traces its origin to the UN Framework Convention on Climate Change, adopted in 1992, which established the basic structure of periodic review and progressive commitment. The Kyoto Protocol (1997) and Paris Agreement (2015) are both products of the COP process.
The Paris Agreement's central innovation was replacing top-down internationally determined emissions targets, which Kyoto had used and which produced significant non-compliance and withdrawal, with nationally determined contributions. Each country sets its own emissions reduction target, submits it to the UNFCCC secretariat, and is then assessed against its own stated ambition rather than against an internationally imposed standard. The Paris COP reviews aggregate progress against the global temperature goal of limiting warming to 1.5 degrees Celsius above pre-industrial levels, and each review cycle is meant to produce increased national ambition.
The structural appeal of this design is that it is achievable. A process that allows countries to set their own targets, assessed against their own ambitions, with transparency mechanisms rather than penalties for shortfalls, can secure broad participation. All the world's major economies are Paris Agreement parties. Kyoto, which had binding national targets with defined consequences for non-compliance, lost the United States and eventually Canada, Japan, and Russia.
The structural problem with this design is that participation and effectiveness are not the same thing. The nationally determined contributions submitted under Paris have, in aggregate, been significantly short of what the IPCC assesses is needed to meet the 1.5-degree goal. The compliance mechanism is transparency and political pressure, not formal penalties. Countries that miss their own targets face public scrutiny in the next review cycle but no legal consequences.
Three structural weaknesses in the climate model
Three features of the climate COP that have limited its effectiveness would transfer directly to an AI governance COP that adopted the same design.
Consensus requirements. Most COP decisions require agreement of all parties, giving any single country effective veto power over the outcome. This produces language negotiated to the level of ambition that the most resistant party will accept, which in successive climate COPs has meant language about "phasing down" rather than "phasing out" fossil fuels, about "pathways" toward net zero rather than binding schedules, and about "noting" scientific findings rather than formally adopting their implications. Consensus-based international negotiations systematically produce outputs at the lowest common denominator of political feasibility among the attending parties.
Self-determination of commitments. The nationally determined contribution model solved the participation problem Kyoto created, but it created an ambition problem. When countries set their own targets with no binding floor, the targets they set reflect domestic political constraints as much as the technical requirements of the underlying goal. Countries with powerful fossil fuel industries set targets their industry can meet. Countries facing near-term elections set targets that are politically defensible in the near term. The aggregate result is a gap between stated ambitions and the aggregate reduction needed, which the COP reviews document but cannot close without a mechanism to mandate higher ambition.
The Paris Agreement solved the participation problem of the Kyoto Protocol by allowing countries to set their own targets. Kyoto's binding national targets led to withdrawal by the United States, Canada, Japan, and Russia. Paris achieved near-universal participation. But near-universal participation with nationally determined and non-binding commitments produces a governance gap at exactly the point governance is most needed: when the gap between stated ambition and required action is largest. AI governance faces the same trade-off.
The enforcement gap. The Paris Agreement's compliance mechanism, established under Article 15, is explicitly non-adversarial and non-punitive. It operates through transparency: countries report on their progress, the COP reviews the reports, and the political and reputational pressure of public assessment is the primary consequence for shortfalls. For climate, this has produced decades of documented shortfalls with limited response. For AI, where a single year of unmonitored frontier development could produce capabilities that outpace any governance response, a transparency-only enforcement mechanism would be structurally insufficient.
Where the COP structure does work
The climate COP's weaknesses are well-documented. Its strengths are less often acknowledged and are genuinely relevant for AI governance design.
Regular convening is itself a governance mechanism. Annual COP meetings create a political calendar that concentrates attention on the underlying governance problem each year. Civil society, journalists, and policymakers treat COP meetings as accountability moments regardless of whether formal enforcement exists. The political cost of visible non-compliance, documented in the global press at an annual COP review, has motivated some governments to strengthen their national commitments. Regular convening also provides a venue for side agreements, technical working groups, and bilateral negotiations that are harder to organize without a recurring global gathering.
Progressive commitment structures can work over long timescales. The ratchet mechanism built into the Paris Agreement, which requires that each successive nationally determined contribution be no less ambitious than the previous one, creates a legal obligation of directional improvement even without a binding floor. Over decades, this can produce significant aggregate movement, as the political baseline for "current policy" becomes progressively more ambitious. For a governance problem like climate, where the consequences of policy failure accumulate over decades, this slow ratchet may be sufficient. For AI governance, where catastrophic outcomes could arrive on a timescale of years rather than decades, it is too slow.
The COP process has been most effective in establishing the normative framework within which domestic climate politics operates globally. The Paris Agreement's 1.5-degree goal is now the reference point against which national climate policies are publicly assessed in virtually every country. This normative function, creating a shared standard that domestic politics must justify itself against, is something the AI safety summit process has begun to accomplish and should continue.
What a stronger AI governance COP would need
An AI governance process modeled on the COP structure but designed to avoid the climate COP's enforcement gap would need three features that the Paris Agreement deliberately omits.
Mandatory minimum commitments. Rather than nationally determined contributions with no binding floor, an AI governance treaty could establish mandatory minimum obligations that all parties must meet, combined with a nationally determined contribution structure for measures above that floor. The mandatory minimum would cover the highest-risk activities: frontier model training above a defined compute threshold, deployment of AI systems in specified high-risk domains, and mandatory safety evaluations before deployment. Countries could go further, but not less far.
An independent verification body. The climate COP relies on self-reported national inventories, with independent review conducted by technical expert teams under the UNFCCC. An AI governance verification body with direct access to relevant information, analogous to the OPCW's inspectorate, would produce more reliable compliance assessment than self-reporting. This is the single structural difference between the CWC and the Biological Weapons Convention that most clearly explains their different effectiveness records.
Consequences with teeth. Consequences for non-compliance that go beyond transparency would require binding the treaty's market access benefits to compliance with its safety obligations. States that fail to meet mandatory minimum requirements would lose the preferential market access and technology sharing arrangements that treaty membership provides. This creates a consequence structure that does not require Security Council authorization and that is directly relevant to the competitive interests that motivate AI development.
"The climate COP shows that regular international gatherings can build a normative framework and a political calendar around a governance problem. It also shows that this is not enough if the underlying commitments lack binding floors and the enforcement mechanism lacks formal consequences. AI governance can use the COP structure and learn from these specific design gaps."
The AI safety summit process that began at Bletchley is worth building on, not abandoning. Annual or biennial summits create exactly the kind of regular international attention that governance requires. The question is whether the summits remain a venue for voluntary commitment ceremonies or become the governing body of a treaty with binding obligations and a functional verification and enforcement architecture. That question will be answered not by the summits themselves but by what governments decide to negotiate between them.
Common questions.
What is a Conference of the Parties and how does it work?
A Conference of the Parties (COP) is the governing body of an international treaty, consisting of all ratifying states. It meets periodically to review implementation, assess progress, and negotiate updates. The Paris Agreement on climate uses this structure: countries submit nationally determined contributions to emissions reduction, the COP reviews aggregate progress against the global temperature goal, and countries are expected to increase ambition over successive review cycles. The COP has no direct enforcement power over national policies; political pressure through public assessment is the primary consequence mechanism.
Why has the climate COP process been slow to produce results?
Three structural factors: consensus requirements give any single country veto power over outcomes, producing language at the lowest common denominator of political feasibility; nationally determined contributions allow countries to set their own targets with no binding floor, producing aggregate ambition gaps; and the enforcement mechanism is a transparency mechanism rather than a penalty mechanism, so countries that miss their targets face scrutiny but no formal consequences. Each factor that made the Paris Agreement politically achievable also limits its governance effectiveness.
What would an AI governance COP need to do differently than the climate COP?
An effective AI governance COP would need mandatory minimum commitments rather than nationally determined contributions with no floor; an independent verification body with inspection authority rather than self-reported compliance; and consequences for non-compliance that go beyond transparency, such as market access restrictions for non-compliant AI systems. The climate COP's greatest structural weakness is the gap between its goals and its enforcement tools. AI governance faces a faster-moving problem where a transparency-only compliance mechanism is insufficient.
Are there COP-style processes that work better than the climate model?
Yes. The Chemical Weapons Convention's governing body, the Conference of the States Parties, has stronger compliance tools: the OPCW has its own inspection and verification function, can attribute violations without Security Council authorization, and can suspend the rights and privileges of violating states. The lesson from the CWC is that the COP structure is not the limiting factor. The binding character of the underlying commitments and the independence of the verification body are what determine whether a COP process produces effective governance.