
AI Safety, Frontier Development Pace, and Regulation
International regulation
Comparing regulatory systems without false equivalence
GLOBAL REGULATION As of 19 November 2026 · Regulatory architecture · Regulated actor and object · Status of the instrument · Compare the control points
Unit of regulation
The useful question is not whether each jurisdiction has an “AI law.” It is which actor must do what, for which technology or service, at what point in the lifecycle, and with what power to verify or intervene. The European Union regulates both general-purpose models and distinct categories of AI systems. China’s interim generative-AI measures focus on services offered to the domestic public. The United Kingdom combines pre-existing sector-specific oversight with government-backed evaluation capacity. Those different units explain why a checkbox comparison of “strictness” can produce the wrong answer. [67] [81] [74]
A second distinction concerns instrument strength. A binding regulation can be reviewed against its operative articles and dates. Administrative guidance describes a regulator’s interpretation but does not rewrite the statute. A technical code can clarify what evidence may demonstrate compliance. A voluntary commitment can still shape behavior, but whether someone can compel disclosure or impose a remedy depends on separate authority. Treat an institute’s test, a company’s commitment, and a regulator’s decision as distinct evidence classes.
Reading a comparison register
For a concrete example, consider one supplier releasing a broad model, a hosted general-purpose service, and a third party’s high-risk application based on that service. A jurisdiction may govern the first, the second, the third, or several. Marking the whole supplier “covered” misses where responsibilities and evidentiary burdens sit. Likewise, finding an exemption for one product does not establish that copyright, safety, privacy, procurement, or sector-specific obligations disappear elsewhere.
A comparison register should therefore preserve status and dates: proposed, adopted, entered into force, applicable, enforceable, or superseded. These are separate labels. Add the official source, article or section, supervised actor, trigger, decision maker, examination process, and available remedy. If the source cannot answer a field, mark it unknown rather than inferring a power. The resulting register is not a league table. It is a map of control points and the gaps between them.
EU AI Act | Scope and classification
GLOBAL REGULATION As of 19 November 2026 · European Union · Model is not the same as downstream system · Classification before launch · Map entities and intended purpose
Two sets of obligations
The EU AI Act does not establish one uniform permission layer for every algorithm. Its architecture combines prohibited practices, obligations for certain high-risk systems, transparency rules in specified settings, and a separate framework for general-purpose AI models. Which obligations matter depends on the regulated actor, where and how the technology is made available or used, the statutory definitions, and any applicable exceptions. The consolidated legal text, not shorthand from the original enactment, is the starting point. [67] [69]
The unit of classification matters. A general-purpose model can be integrated into many products. One integration may be a routine assistant; another may be deployed in a sensitive process listed by the Act. The model provider’s duties and a downstream system provider’s duties are not interchangeable. An enterprise that incorporates a third-party capability should define its own intended purpose and check whether its modifications, branding, or deployment change its role under the Act rather than assuming supplier assurances settle the issue.
Classifying a deployment
Practical classification starts with a system inventory, not a benchmark leaderboard. Record the model and version, the output relied on, the affected people, the jurisdiction of use, the deployment pathway, and whether a human can disregard or reverse the output. Identify entities acting as providers, deployers, importers, or distributors as applicable. Retain the rationale for any exemption; where classification is unsettled, seek a review rather than silently selecting the less burdensome category.
The EU approach can make obligations more predictable across its market, but uncertainty can move to interpretation, standards, classification, and supervisory capacity. The policy tension is not simply safety versus speed. It is whether a company can classify the same configuration consistently enough to act before launch, while regulators can revisit the classification as use or capability changes. The regulatory record should document what changed, the legal trigger reviewed, and the owner accountable for updating that record.
EU timetable | What changed in July 2026
GLOBAL REGULATION As of 19 November 2026 · European Union · 2026 Omnibus is enacted · Separate applicability from enforcement · Maintain a cohort-specific calendar
Operative dates
The EU’s AI Act entered into force in August 2024. Most of its initial prohibited-practice and AI literacy provisions began applying on 2 February 2025. General-purpose AI model obligations became applicable on 2 August 2025, and the Commission’s enforcement powers over those model obligations became applicable on 2 August 2026. The legal timetable is staggered; referring to a single “start date” loses information that is material to an enforcement or launch decision. [67]
Regulation (EU) 2026/1744, the Digital Omnibus on AI, entered into force on 27 July 2026. As reflected in the consolidated Act, relevant Chapter III provisions for Annex III high-risk systems now apply from 2 December 2027, while the applicable provisions for Annex I systems under Article 6(1) apply from 2 August 2028. The amendments are enacted law, not merely the Commission’s 2025 proposal. [68]
Transitions and implications
Additional milestones remain important. The Commission identifies 2 December 2026 as the application date for newly added prohibitions involving certain uses that generate or manipulate non-consensual intimate material and child sexual abuse material. Under Article 111, general-purpose models placed on the market before 2 August 2025 have a transition until 2 August 2027; some synthetic-content systems placed on the market before 2 August 2026 have a separate Article 50(2) transition until 2 December 2026. [67] [70]
For the pace debate, the July amendment is a real example of policymakers extending some downstream compliance windows while leaving other model and system obligations in place. It does not show that the entire regulatory regime was delayed. A sensible implementation calendar has separate rows for each obligation, release cohort, and system classification. When deadlines move, update the controlling source and the affected workflow rather than replacing all dates in a presentation with the most prominent headline.
EU model obligations | Documentation as a control
GLOBAL REGULATION As of 19 November 2026 · European Union · Transparency and copyright · Provider and downstream evidence · Keep release records versioned
Article 53 evidence
Article 53 of the EU AI Act sets obligations for providers of general-purpose AI models, including technical documentation, information for downstream providers, a policy for compliance with EU copyright law, and publication of a sufficiently detailed summary of training content using the AI Office’s template. Information prepared for regulators, information needed by downstream integrators, and information published for public transparency do not serve identical purposes. [67] [71]
The documentation pathway is more than a filing exercise. A downstream provider may need the model’s intended characteristics, constraints, and integration requirements to decide whether a product can be released safely. A governing body may need provenance and evaluation evidence to reconstruct a release decision. Different audiences need different levels of detail, and disclosure should be handled with attention to lawful confidentiality, intellectual property, and safety-sensitive information.
Scope and artefacts
The Commission’s related guidelines describe its practical interpretation and timelines. A qualifying free and open-source licence can affect specific documentation duties under Article 53, subject to statutory conditions and exceptions; it is not a blanket exclusion from the regime. In particular, the relevant exemption does not apply to general-purpose models classified as posing systemic risk. The copyright policy and training-content summary obligations should not be assumed away by labelling the release open. [67] [72]
A useful internal ledger pairs each requirement with the owner, evidence artefact, intended recipient and escalation channel. Examples of artefacts include a model identifier, release manifest, documentation revision, source governance policy, and contact for downstream questions. Version control matters because a release can differ from what earlier documentation describes. That is also why an impressive model card, by itself, cannot establish that obligations were fulfilled for the launch under review.
EU systemic risk | Threshold and evidence
GLOBAL REGULATION As of 19 November 2026 · European Union · Compute presumption, not outcome prediction · Systemic-risk duties · Separate classification and risk judgement
Article 51 classification
The Act presumes that a general-purpose AI model has high-impact capabilities when the cumulative computation used for its training exceeds 10^25 floating-point operations. That presumption is a classification mechanism. Article 51 also provides for Commission designation based on equivalent capability or impact criteria; the legal analysis does not collapse into counting FLOP. Providers must examine the current legal text and guidance rather than treating a rounded internal compute estimate as a final legal conclusion. [67] [71]
A compute proxy offers officials a relatively legible screen before deployment. Its weakness is that useful or potentially dangerous capability can also change through new architectures, improved data, system integration, or later updates. Conversely, a high training compute value is not proof of a specific accident pathway. The policy value of the proxy comes from triggering a deeper evidence and reporting obligation, not from treating it as a measurement of expected loss.
Article 55 provider controls
Article 55 assigns additional duties to the providers of systemic-risk general-purpose models. These include model evaluation, assessment and mitigation of systemic risks, documentation and reporting of relevant serious incidents, and adequate cybersecurity for both the model and its physical infrastructure. The responsibilities apply at the model layer and should not be confused with a downstream application’s conformity assessment. [67]
For an accountable release decision, preserve both classifications: the statutory finding and the internal safety judgement. A legal threshold crossing can require new notifications or documentation even if the team expects strong mitigation. A below-threshold result does not invalidate internal concerns supported by reliable evidence. Record the threshold assumption, model and release configuration, decision owner, residual risk and re-evaluation trigger. This architecture avoids turning a screening number into a false assurance label.
EU enforcement | Where responsibility lands
GLOBAL REGULATION As of 19 November 2026 · European Union · Two-tiered oversight · A complaint is not a finding · Record the path from claim to decision
Enforcement roles
The Commission’s AI Office enforces obligations for providers of general-purpose AI models and has functions for some AI systems. Member States’ national competent authorities have major supervisory and enforcement duties regarding AI systems. The AI Board helps coordinate implementation, while scientific and stakeholder bodies provide input. Separating these functions is essential: a technical advisory opinion is not a sanction, and a national complaint channel is not automatically the competent route for a model provider’s specific duty. [70] [69]
The Commission describes authority to request model documentation, undertake evaluations and seek corrective measures, alongside applicable fines. It also identifies reporting channels for alleged violations, whistleblowers, and certain downstream model-integration complaints. A channel’s existence does not establish that a complaint was upheld, a violation found, or a remedy imposed. Each stage should be logged separately. [70]
Evidence and supervisory records
A monitoring plan should identify the first official document at each escalation stage: complaint acknowledgement, request for records, formal assessment, proposed remedy, final decision, or appeal. Agency web pages can explain process, but an enforcement claim needs evidence tied to the actual action. Date-stamp every fact and distinguish public disclosures from private contacts that cannot be independently checked.
The General-Purpose AI Code of Practice is a voluntary route for helping demonstrate compliance; it is not the law itself. A non-signatory remains subject to applicable duties and may need to demonstrate how its alternative measures satisfy them. A signatory should not advertise membership as official model approval. The decision-useful question is whether the evidence, scope and timing of the compliance route satisfy the relevant obligation. [73] [72]
United Kingdom | Sector oversight and coordination
GLOBAL REGULATION As of 19 November 2026 · United Kingdom · Sector rules plus coordination · Technical evaluation is not regulatory approval · Follow the actual deployment
Cross-sector policy
The UK’s published cross-sector approach asks existing regulators to interpret AI governance principles within their remits rather than treating every AI use as subject to the same standalone regime. Government guidance describes safety and robustness, transparency, fairness, accountability, and contestability as cross-cutting objectives intended to supplement existing legislation. Those objectives do not displace the statutory framework for an affected industry or use. [74]
This approach can use sector expertise earlier than a new central regulator could develop it. A regulator already understands the decision processes and harms that matter in its sector. But boundaries can be difficult when a general-purpose system is shared by several sectors or no single actor can see its full lifecycle. Firms need a map of applicable regulators and cross-cutting duties, not just a list of principles drawn from policy guidance.
Deployment-specific implementation
A useful test case is a hosted model adopted under separate contracts by a public body and a regulated private firm. The two deployments can require different controls and evidence even though they depend on the same upstream model. The release record should distinguish changes to the underlying model, changes in how the organisation uses it, and changes to sector or statutory expectations. A successful pilot in one setting is not transferable certification for another.
Technical institutions matter in this model. The UK AI Security Institute undertakes research and testing and informs policymaking. It is distinct from a regulator and from an organisation’s decision-maker for release. This separation can create useful checks, but also requires a documented route from important test findings to someone with authority to act. [75] [76]
United Kingdom | Readiness is not enactment
GLOBAL REGULATION As of 19 November 2026 · United Kingdom · Proposal versus actual duty · Agentic resilience · Prepare without claiming enactment
The October package
In October 2026 the UK government announced a package addressing advanced AI risks, including evaluation, defensive security work, and preparedness for incidents involving autonomous systems. Those measures are not all of one legal kind. An announced evaluation programme can operate as research or government activity, while a proposal to make new critical-infrastructure resilience duties requires the relevant legislation, delegated authority and implementation steps. [77]
The government’s November 2026 factsheet on agentic AI resilience described a plan to consult on using delegated powers under the Cyber Security and Resilience Bill to help regulated entities respond to agentic loss-of-control incidents affecting networks and information systems. It is evidence of intended policy direction. It is not evidence that the contemplated secondary legislation was already in force, or that equivalent reporting duties apply to every frontier developer. At this report’s cutoff, Parliament showed Lords amendments awaiting scheduled Commons consideration on 24 November 2026, not Royal Assent. [78] [79]
Readiness and escalation
For an operations team, separate existing duties from readiness work. An organisation can develop clearer incident ownership, preserve evidence and exercise a rollback process while proposed rules move through the legal process. These practical steps reduce future adaptation costs whether or not the final rules match the announced design. Scope changes should affect the legal register, not retrospectively turn preparatory work into proof of compliance.
The policy risk of a dispersed approach is delayed escalation if operators and regulators disagree about what constitutes an incident, which body has jurisdiction, or whether a near miss qualifies. Its advantage is the ability to align responses with established infrastructure, incident-response frameworks and industry expertise. Measure the quality of this system by how quickly a validated warning reaches an accountable decision-maker, not by how many government documents use the term “agentic AI.”
United Kingdom | What evaluation can establish
GLOBAL REGULATION As of 19 November 2026 · United Kingdom · Public aggregate is not ranking · Controlled evaluation caveats · Document the tested configuration
What testing can support
The AI Security Institute describes its evaluations as a supplementary layer of oversight rather than a comprehensive certification of safety. It tests advanced systems before and after release and develops technical research that can inform governments. This role is particularly valuable where capability evidence is changing faster than new statutory duties can be designed. It does not transfer release responsibility away from the parties making that decision. [75] [76]
AISI’s frontier trends report describes work across more than 30 frontier systems using several methods, including controlled capability tasks and human-impact research. Among other limits, the report distinguishes task performance from real-world effectiveness, describes its aggregation and sampling, and does not use the public aggregate as a named developer ranking. Its evidence cuts against both ignoring capability change and treating an isolated result as a forecast of realised harm. [80]
How to use the evidence
For decision use, request the evaluated system version, the relevant evaluation context, the coverage of the tasks, and the conditions under which the result might fail to generalise. A score can move because of changes to the model, configuration, evaluator, sample, or available compute. Clarify what was held fixed before treating a difference as an improvement. Where public data intentionally omit sensitive details, state the resulting limits on external reproduction rather than supplying missing operational information.
The independent-testing model faces an access problem. A voluntary agreement can permit urgent tests and protect confidential material, but may not guarantee that the evaluator can test all configurations or publish all findings. Mandatory access rules may broaden coverage but still require funding, security controls, version traceability and qualified staff. The institutional design question is how to connect technical evidence to a lawful remedy while preserving the conditions needed to obtain that evidence reliably.
China | Regulating the public generative service
GLOBAL REGULATION As of 19 November 2026 · China · Domestic public-facing service scope · Filing trigger depends on service attributes · Compare supervised activities
Service scope
China’s Interim Measures for the Management of Generative Artificial Intelligence Services took effect on 15 August 2023. Article 2 applies the measures to services using generative AI to supply specified kinds of content to the public in mainland China. It excludes certain research and use by organisations that do not offer such services to the domestic public. That stated exclusion is not a conclusion that other Chinese laws cannot apply to internal research or processing. [81]
The measures require attention to lawful sources for training data and base models, intellectual property, personal-information rules, quality of training data, service practices and prohibited content. They address service providers and also define the service pathway to include provision through programmable interfaces. A cross-border provider should not assume that a foreign address or the absence of a local storefront decides the public-service question.
Administrative triggers
Article 17 requires a security assessment and compliance with the applicable algorithm-filing provisions for services with public-opinion attributes or social-mobilisation capacity. This is a defined trigger in the measures. It should not be rephrased as an assertion that every model-training project follows an identical filing pathway or that filing is proof of general safety. The measures also describe complaint handling and powers for supervisory examination. [81]
This framework differs structurally from the EU’s threshold-based systemic-risk layer. It pays particular attention to public-facing services, content governance and administrative oversight, without relying on a single frontier compute number in this instrument. One global operator may need separate evidence for the public service in one country and for the underlying model’s classification in another. A compliance comparison should trace the product and operator across jurisdictions, rather than translating legal categories into apparent equivalents.
China | Labels and distribution responsibility
GLOBAL REGULATION As of 19 November 2026 · China · Visible labels and metadata · Separate provider and platform roles · Treat disclosure as evidence, not truth
Two kinds of label
China’s Measures for Labeling AI-Generated and Synthetic Content entered into force on 1 September 2025. For covered services, the measures distinguish explicit indications a person can perceive from implicit markings in file metadata. The rules address multiple formats and create responsibilities not only for generators but also for providers that distribute or make content available. The underlying Chinese text governs; an English description is a research paraphrase. [82]
Distribution responsibilities include checking for implicit labels and handling disclosures or indicators in specific circumstances. The rules also address application distribution services and user declarations. Different actors may hold different evidence about content origin. A notice visible to a human and an implicit marker available to a platform help answer separate questions; the presence of one does not by itself guarantee the preservation or visibility of the other throughout distribution.
A publication pathway
One traceability scenario is an organisation that generates an image using a third party, stores a copy internally, and publishes it on a platform. Evaluate who generated the file, who can identify its origin, how the distribution platform confirms or displays a label, and whether the publishing workflow preserves relevant information. These are governance and usability questions, not a substitute for legal advice about a specific publication. [82]
AI-content labeling is a signal about provenance or disclosure, not a determination that material is true, lawful, or harmless. False claims can remain false when accurately labelled as synthetic, and unlabelled material is not necessarily authentic. A coherent policy pairs technical provenance with meaningful disclosures, channels to report mislabeling, explicit responsibility, and monitoring of label performance. Treating the label alone as a safety outcome confuses a mechanism meant to improve transparency with the downstream results it is meant to support.
Japan | Promotion statute and public procurement
GLOBAL REGULATION As of 19 November 2026 · Japan · Promotion statute and planning instruments · Public sector standards · Assess the procurement control point
Statute and public-use standards
Japan’s Act on Promotion of Research and Development, and Utilization of Artificial Intelligence-related Technology is Act No. 53 of 2025. The Japanese Digital Agency’s November 2026 provisional English government standards identify a first AI Basic Plan adopted in December 2025 and a second adopted on 14 July 2026. These dates describe distinct government actions: a statute, Cabinet planning decisions, and administrative standards for use in government. [83]
The November standards are specifically written for government generative AI governance, security and safety. They illustrate how a government can influence implementation through purchasing, system design and deployment rules even apart from a comprehensive product licensing regime. Government-use requirements should not be promoted as universal obligations for every private company unless the underlying instrument actually reaches it.
The purchasing control point
Public procurement changes incentives at a specific control point. A purchasing agency may require information about the provider, intended use, update practices, privacy protections, and responsibility for monitoring as conditions of deployment. Documented ability to meet those conditions may matter more to the agency than a broad benchmark score. Providers can develop repeatable evidence packages for this purpose without treating a procurement award as a universal certification.
The trade-off is reach against adaptability. Procurement rules can create detailed and testable practices in government-supported deployments and build institutional competence quickly. They do not automatically govern all private deployments, exported services, or the earliest model development decisions. Comparisons should state which part of the market is reached, who confirms compliance, and what happens if a provider’s system materially changes after initial selection.
Republic of Korea | A trust-and-innovation statute
GLOBAL REGULATION As of 19 November 2026 · Republic of Korea · Effective 22 January 2026 · Foundational act and decrees · Do not borrow labels across regimes
Enacted framework
The Republic of Korea’s AI Basic Act entered into force on 22 January 2026, subject to specific timing provisions in the amended statute. The law creates a national architecture for AI development and trustworthiness rather than operating solely as a prohibition instrument. It includes provisions on transparency, safety and high-impact AI, alongside institutions and programmes intended to support AI development. The current Korean text and applicable decrees govern; official English materials are useful aids, not substitutes for the operative Korean law. [84] [85]
The law’s categories are not equivalent to an EU classification merely because they describe elevated risks. Compare the definitions, affected actor, deployment context, and evidentiary obligation article by article. A system may be classified differently under the two frameworks even when deployed for broadly similar purposes. Use a detailed obligation matrix that preserves these differences instead of importing a label such as “high risk” from one jurisdiction unchanged.
Reading the implementing record
A second reading discipline concerns implementation. Korean public notices may describe enacted provisions, implementing decrees, or legislative notices for amendments. A notice opening a proposal for comment is not by itself proof that the final text entered into force. The Ministry of Science and ICT’s May 2026 notice describes the distinction between the January effective date and amendments taking effect at a later stage. [86]
Korea’s framework is useful to this study because it combines industrial development and trust mechanisms in one national architecture. Firms may gain support and standardisation while also encountering new duties on qualifying systems. The public policy test is whether implementation translates the statute into obligations that can be verified and updated, without confusing government industrial leadership with an exemption from safety or consumer accountability.
Singapore | Accountable deployment of agents
GLOBAL REGULATION As of 19 November 2026 · Singapore · Guidance for deployers · Permissions determine consequences · Test approvals in the deployed workflow
Governance framework
Singapore’s Infocomm Media Development Authority published a Model AI Governance Framework for Agentic AI in 2026. The document is a practical governance framework for deploying agentic systems, including both internally developed agents and third-party services. Its four areas span risk assessment and boundaries, meaningful human accountability, technical and process controls, and the user dimension. It is guidance rather than a universal statutory licensing requirement for frontier training. [87]
The framework is particularly relevant when agents can act across connected business tools and workflows. Rather than relying on a general statement that a human is “in the loop,” it asks whether responsibility is assigned, oversight is usable, and agents operate with suitable limits. A deployment should establish which actions may be automated, what needs meaningful approval, how people can stop the process, and what evidence the deployer retains.
Illustrative deployment
Consider a procurement agent that can propose a supplier and also initiate external action. The same underlying model can be low impact in a recommendation-only mode and more consequential when granted execution access. Define the authorised scope first. Then test the deployed configuration, including whether approval actually blocks action and whether the operator can detect and reverse an unintended outcome. This is an operational control example, not a recommendation to launch without applicable legal and security review.
Singapore’s guidance is most transferable as a method for aligning controls with actual permissions and actions. Its limits arise when readers substitute a general framework for evidence about a specific device, workflow or jurisdiction. A robust release record links agreed controls to an observed test result and an identified decision-maker. Where the process cannot be paused, reversed or adequately monitored, the framework invites reconsidering the deployment design rather than claiming that documentation alone reduced risk.
G7 Hiroshima Process | Voluntary reporting
GLOBAL REGULATION As of 19 November 2026 · International cooperation · Voluntary organisational reporting · Verification and remedy are separate · Ask whether evidence is traceable
Instrument and reporting
The G7 Hiroshima process developed guidance and a code of conduct for organisations developing advanced AI. In February 2025, the OECD launched a reporting framework to structure how participating organisations describe their implementation. This can make otherwise scattered safety policies more comparable and establishes a recurring basis for disclosure. It is not a global regulator, a universal filing requirement, or proof that self-reported controls work under live conditions. [88] [89]
Distinguish commitment, disclosure, verification and remedy. A commitment is an announced policy. A disclosure describes practices or outcomes. Verification requires sufficient evidence or independent examination. A remedy requires authority, ability and a trigger. A voluntary framework can improve the first two and support the third, but does not itself create the fourth unless a separate statute, contract or authority supplies it.
A workable comparison method
An analyst reading Hiroshima reports should compare like with like: the same period, the same model or model family, publication versus internal practices, and comparable meanings of safety evaluation and incident. An omitted detail may indicate legitimate confidentiality, non-applicability or material information gaps. Without verification it cannot be treated as proof either that a control is absent or that it was performed effectively.
The regulatory benefit is interoperability. Common questions may lower duplicative reporting burdens and help smaller oversight teams identify what evidence is available. Common questions do not imply one outcome is acceptable in every jurisdiction. A business can use one evidence spine and different jurisdictional overlays. The reasonable test is whether a reported mitigation is traceable to the deployment it covers and can support a follow-up decision when conditions change. [89]
Seoul commitments | What a pause promise means
GLOBAL REGULATION As of 19 November 2026 · International cooperation · Commitment is not a court order · Threshold, control, pause · Use reviewable decision rules
Voluntary commitments
The Seoul Frontier AI Safety Commitments are voluntary commitments announced by the UK and Republic of Korea, with a published list of participating organisations. They call for public frameworks that define thresholds for severe risks, specify mitigations and set out what happens if the thresholds are met or exceeded. In the extreme, the commitments contemplate not developing or deploying a model or system when mitigations cannot keep residual risk below the threshold. [90]
This is stronger in language than a generic pledge to prioritise safety. However, it is not an automatic legal moratorium triggered by a single company statement. The practical meaning depends on how thresholds are defined, the independence of evaluation, who may assess and override evidence, which stages of development or deployment the commitment covers, and who can learn whether the decision rule was used.
Testing a threshold decision
A clear internal threshold protocol separates the measurement rule, evidence submission deadline, decision authority and disposition. When evidence is inconclusive, follow a previously defined procedure rather than letting a deadline decide by default. Preserve a record of why a mitigation is expected to reduce the residual risk, what result would show that expectation was wrong, and who can pause or restart work. The same discipline applies to decisions to continue or to halt: each needs a reviewable rationale.
Signatory names must also be read as published, according to the source and as of the relevant date. A listed laboratory name is not alone proof that a certain parent, affiliate or individual is legally bound in some separate dispute. Corporate identity, legal responsibility and voluntary policy adherence are independent factual questions, even where the same technology or personnel appear across several contexts. [90]
Council of Europe | A rights-based treaty
GLOBAL REGULATION As of 19 November 2026 · International cooperation · Council of Europe is not the EU · Signature and ratification differ · Verify the current treaty chart
Treaty and parties
The Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law opened for signature in September 2024. Its focus is the integrity of rights, democratic institutions and the rule of law across relevant AI activities. This is distinct from EU rules classifying particular models or products for obligations. The Council of Europe and the European Union are also distinct institutions; one should not attribute EU-style market supervision to the Convention. [91]
The Treaty Office lists signatures and ratifications separately. Its entry-into-force condition is five ratifications, including at least three Council of Europe member States. The European Union ratified on 15 May 2026, according to the Council of Europe’s October 2026 committee record. A ratification by one party is not, on its own, evidence that the aggregate treaty threshold has been met. The official status chart should be rechecked when asserting applicability on a specific date. [92] [93]
Rights oversight in practice
For risk analysis, the Convention highlights a different set of questions: are affected people able to challenge decisions, do safeguards protect rights, and can the oversight system identify and remedy harms? A frontier model evaluation geared toward severe technical risks cannot by itself answer those questions. Conversely, a rights-governance process may need external technical evidence that it cannot produce on its own.
Treaty participation can encourage convergence through domestic implementation and cooperation, but translation into effective practice depends on the party, institutional capacity, and the specific implementation measures. Do not infer an identical duty or remedy in every jurisdiction from participation alone. The existence of a legal framework can be meaningful even where its international enforcement architecture is unlike a regulator who can investigate one supplier directly.
OECD and United Nations | Principles to shared evidence
GLOBAL REGULATION As of 19 November 2026 · International cooperation · OECD and UN mandates differ · Assessment is not product approval · Link shared evidence to authority
Complementary mandates
The OECD AI Principles include responsible stewardship, transparency and explainability, robustness and safety, and accountability. They offer governments and organisations a shared vocabulary for governance across the AI lifecycle. The recommendation is conceptually influential; it should not be misread as a law that directly requires a particular frontier model to halt in every participating jurisdiction. [94]
The UN Global Digital Compact made wider commitments on digital cooperation, including work toward inclusive AI governance and capacity building. In August 2025, General Assembly Resolution A/RES/79/325 established an independent scientific panel and a Global Dialogue on AI Governance. The first Dialogue occurred in Geneva on 6–7 July 2026 and included presentation of the panel’s preliminary report. The institutional functions are assessment and deliberation rather than a global product-approval system. [95] [96] [97]
What cooperation can support
Shared assessment can flag threats that reach more than one country, including contexts with little independent evaluation capacity. It can also help ensure that evidence reflects languages, institutions, and social conditions outside the largest model-producing markets. The scientific and policy processes should distinguish how often an outcome was directly observed from how often a potential concern was raised by a participant. Participation is not evidence of agreement on probability or on policy remedy.
An international comparison can use these institutions as a connective layer: the OECD for common terminology and reporting questions, the UN for scientific exchange and broader representation, and domestic authorities for legal action within their mandates. This architecture is less centralised than a global regulator, but can still improve decisions if it supplies coherent evidence and reliable routes for governments to respond.
Cross-border cooperation | Testing without one global enforcer
GLOBAL REGULATION As of 19 November 2026 · International cooperation · Network and scientific synthesis · Technical testing is not enforcement · Define the route for escalation
Research and evaluation
At the Seoul summit in 2024, governments agreed to launch a network for collaboration among publicly backed AI safety institutes. Common work on research, testing, and evaluation capacity can improve the quality of evidence. The network’s existence does not establish authority to compel access, sanction a provider, or require a cross-border pause. Cooperation and enforcement are distinct institutional layers. [98]
The International AI Safety Report is another connective device. It synthesises evidence and describes uncertainty across domains rather than certifying particular models or recommending specific domestic laws. The 2026 edition’s scope focuses on emerging risks from general-purpose AI and draws on evidence published before December 2025. Even an extensively reviewed report cannot substitute for checking the current model configuration and more recent primary observations. [99]
Evidence handoff
A modest cross-border data standard could record model version, evaluation time, test scope, indicator definition, quality of evidence, and what follow-up is pending. Information that cannot be shared publicly could go to authorised evaluators under a secure process. Summaries released to the public should describe enough about the result to distinguish a capability signal from an actual incident without disclosing unsafe or proprietary details.
Interoperability is not uniformity. Common definitions help a receiving body make sense of evidence, but agencies can have different tolerances, procedures and statutory responsibilities. An escalation could therefore trigger a joint review without forcing every country to impose the same remedy. That flexibility is useful if clocks, markets and institutional capacities differ. Its weakness is that comparison without an assigned action owner can leave serious shared signals discussed widely but resolved nowhere.
Global milestones | Maintaining the watch register
GLOBAL REGULATION As of 19 November 2026 · Implementation · Historical versus future milestones · One line per event · Refresh against official sources
Date register
A reliable watch register separates what has happened from what has been announced. Historical dates include China’s August 2023 interim generative-service measures, the EU’s August 2025 general-purpose model obligations, China’s September 2025 labeling measures, Korea’s January 2026 AI Basic Act commencement, the UN Dialogue in July 2026, and EU model enforcement powers from August 2026. Future published EU milestones include December 2027 for the relevant Annex III high-risk provisions and August 2028 for Annex I provisions. [81] [82] [84] [97] [67]
A watch register should be capable of absorbing new developments without rewriting earlier history. Keep source date and effective date separate. Maintain one row per instrument and event, leave uncertainty visible and use change indicators. For example, after a proposal to extend implementation passes, record the earlier proposal, enacted amendment and entry into force as separate events rather than overwriting the public record silently.
Questions that trigger action
Five recurring review questions keep the register useful: does the instrument cover the model, the service, or the specific deployment; who must provide evidence; when does the rule apply; who can order or help remedy a problem; and what source verifies the latest status? Promised consultation dates, claimed voluntary commitments and enforceable deadlines belong in different fields.
For an organisation evaluating development pace, the watch register is now a schedule input. Route changes affecting planned releases to engineering, legal and the responsible business owner together. Treat unfamiliar jurisdictions or hard-to-classify product changes as issues to investigate, not as inferred permission. A governance system is more useful when it preserves both when the organisation knew of an obligation and when it decided what to do about it.
Exhibit 10 · How regulatory instruments connect
Authorities can define duties and read records; metrics and documentation improve oversight but cannot substitute for proof that controls work. This is not a map of any individual jurisdiction’s legal powers.
48PEGASUS Research conceptual exhibit; not observed data.
Browse: White Papers · Reports · Downloads
