International governance is no longer empty—but it is fragmented
As of September 2026, it is inaccurate to say that no binding international AI instrument exists. The Council of Europe Framework Convention on Artificial Intelligence is a legally binding treaty for parties once its entry-into-force conditions are met. It is also inaccurate to imply that a global, verifiable treaty now limits frontier-model development. No such universal regime exists.
The landscape contains political declarations, voluntary company commitments, regional law, standards, and one framework convention focused on human rights, democracy, and rule of law. Their legal force and subject matter differ.
Bletchley established a forum
The November 2023 Bletchley Declaration was endorsed by 28 countries and the European Union at the first AI Safety Summit. It recognized potentially serious harm from frontier AI and supported cooperation on measurement, safety testing, and public-sector capacity (UK Government, November 2023).
It is a political declaration, not a treaty. It created no inspectorate, compulsory evaluation, or penalty. Its significance was diplomatic: the United States, China, EU members, and other states accepted common language about frontier risk. Participation was geographically incomplete, and attendance does not establish agreement on enforcement.
Seoul in 2024 continued the summit process and produced voluntary frontier-safety commitments from companies. Paris in 2025 broadened the agenda toward AI action, innovation, inclusion, and public interest, exposing disagreement over emphasis and text. Summit communiqués are evidence of stated positions; implementation requires separate verification.
The International AI Safety Report
The summit process helped sponsor an independent scientific synthesis. The 2026 International AI Safety Report distinguishes documented harms from uncertain severe scenarios and records areas of expert disagreement (International AI Safety Report 2026). It does not create legal obligations, and scientific consensus about evidence is not political agreement about policy.
This separation is valuable. Negotiators can share a baseline on capabilities, incidents, and uncertainty without agreeing on regulation. Future reports should retain editorial independence, publish methods and conflicts, and include expertise beyond leading AI states and laboratories.
The Council of Europe convention
The Framework Convention opened for signature on September 5, 2024. It requires parties to address AI throughout its lifecycle consistently with human rights, democracy, and rule of law, using principles such as transparency, accountability, equality, privacy, reliability, and remedies. It is technology-neutral and allows implementation through domestic measures (Council of Europe treaty page).
Its official status page should be checked for signatures, ratifications, and entry into force; as of September 11, 2026, the treaty page still listed the requirement of five ratifications including at least three Council of Europe members and did not list an entry-into-force date (Treaty Office, CETS 225). Signature indicates intent to consider becoming bound; ratification is the stronger legal act. The convention permits reservations and flexible implementation, and national-security scope has drawn criticism. It is not a compute or AGI non-proliferation treaty.
The EU is binding but regional
The EU AI Act is directly consequential for providers placing covered systems or models on the EU market. General-purpose-model duties began applying in August 2025, and the Commission and national authorities began enforcing major provisions in August 2026. Systemic-risk model providers face evaluation, mitigation, incident-reporting, and cybersecurity obligations (European Commission enforcement framework).
The Act has extraterritorial market effects, but it is EU law, not a multilateral treaty. Its high-risk-system timeline has also changed through later EU legislation, so summaries must be dated. Official consolidated law and Commission implementation pages should replace unofficial explainers.
The United States changed direction
U.S. Executive Order 14110 was revoked in January 2025. The replacement Executive Order 14179 and the July 2025 AI Action Plan emphasize innovation, infrastructure, national security, and global leadership (White House, January 2025). Some agency authorities, sectoral laws, procurement rules, export controls, and voluntary laboratory frameworks remain, but the former order’s reporting scheme should not be presented as current.
States filled part of the gap. California’s SB 53, signed September 29, 2025 and effective in 2026, requires covered frontier developers to publish safety frameworks, report specified critical incidents, and protects covered whistleblowers (California Governor). New York’s RAISE Act was signed December 19, 2025 and imposes framework and reporting requirements within its definitions (New York Governor). These are state laws, not U.S. treaty commitments, and coverage thresholds matter.
What a frontier-risk agreement would need
A meaningful agreement must define covered activity without freezing technology. Inputs might include training compute, capability evaluations, deployment scale, and access to dangerous tools. Obligations could include notice, independent testing, security, incident reporting, and restrictions on specified military or autonomous uses.
Verification is the hardest element. Options include secure reporting of large clusters, chip-supply records, on-site inspection, privacy-preserving attestations, and challenge procedures. Each can expose commercial, personal, or national-security information. Verification needs access rules, data minimization, protection against espionage, and consequences for false declarations.
Participation matters as much as design. A regime that excludes key developers creates leakage and fairness problems. A consensus rule can allow obstruction; majority rules can drive nonparticipation. Assistance with evaluation and infrastructure can make obligations feasible for lower-capacity states.
Lessons from arms control—with limits
Arms-control experience shows that adversaries can cooperate on shared catastrophic risk while remaining rivals. Hotlines, inspectors, notifications, test bans, and incident agreements offer institutional patterns. AI differs because civilian uses dominate, software diffuses quickly, private firms hold key assets, and capability cannot be counted like deployed weapons.
The best near-term targets may be specific and verifiable: keeping autonomous systems out of nuclear launch decisions, reporting severe model incidents, protecting safety research, securing weights above defined capability levels, and agreeing on evaluation methods. Narrow agreements can build trust without waiting for a comprehensive AGI treaty.
How to judge progress
Count ratifications, enacted duties, budgeted regulators, completed audits, disclosed incidents, and verified compliance—not conferences alone. Ask which jurisdictions and models are covered, who can investigate, whether affected people have remedies, and what happens after violation. Track corporate commitments separately from government law.
International governance has advanced beyond statements of concern, especially in Europe and in several U.S. states. It remains patchy, weakly harmonized, and far from a global regime capable of verifying mutual restraint at the frontier. That mixed assessment is more accurate than either “nothing exists” or “the world has regulated AI.”