Statutes are local. Frameworks travel.
When a Colorado court asks what a reasonable deployer would have done, when a Monetary Authority of Singapore examiner reviews a bank's model governance, when a procurement officer writes an AI clause into a federal contract, they tend to reach for the same small set of references. Four of them carry most of the weight. This is our reading of them.
NIST AI Risk Management Framework 1.0. Voluntary by design, widely relied on.
Published in January 2023 by the US National Institute of Standards and Technology as NIST AI 100-1.
The AI RMF is organised around four functions: Govern, Map, Measure and Manage. Each breaks down into categories and subcategories, a taxonomy of roughly seventy outcomes an organisation is expected to demonstrate. It is iterative rather than a checklist. It asks operators to run the cycle continuously over the AI lifecycle, to document decisions, and to calibrate risk tolerance to context.
Its weight comes from diffusion more than legal status. The Colorado AI Act treats adoption of a recognised framework as an affirmative defence, Texas TRAIGA gives a defence for substantial compliance with it, federal procurement draws on it, and ISO/IEC 42001 maps to it. When a plaintiff in a US negligence action asks whether a defendant acted reasonably, the RMF is the document closest to hand.
NIST released the Generative AI Profile (NIST AI 600-1) in July 2024 as a companion to the framework. It names twelve risks, among them confabulation, human and AI configuration, and value chain integration, that operators of autonomous agents should map and manage before deployment.
| Function | Core outcome | Applied to agents |
|---|---|---|
| Govern | Culture, policies, accountability, risk tolerance | Board oversight of agent authority |
| Map | Context, use, stakeholders, intended impact | Define agent scope and user reliance |
| Measure | Evaluate, track, monitor, benchmark | Evaluate hallucination, drift, misuse |
| Manage | Prioritise, respond, recover, communicate | Kill switch, incident protocol, rollback |
ISO/IEC 42001. The AI management system standard.
Published in December 2023 by the International Organization for Standardization and the International Electrotechnical Commission. The first certifiable management system standard for artificial intelligence.
ISO/IEC 42001 sits in a different place from NIST. It is certifiable: an organisation can be audited against it by an accredited certification body and hold a certificate recognised across its supply chain. Where statutory obligations are ambiguous, certification gives an operator a defensible baseline and a document procurement teams can point to.
The standard uses the management system architecture familiar from ISO 27001 and ISO 9001: context of the organisation, leadership, planning, support, operation, performance evaluation and improvement. Annex A lists 38 controls specific to AI, covering governance, lifecycle, impact assessment, data quality, system documentation and human oversight. Annex B gives guidance, Annex C maps AI related organisational objectives, and Annex D links to sector considerations.
Its practical value is the documentation scaffold. An organisation that runs a real AI management system produces the records an EU deployer needs for Article 26, a Colorado deployer needs under SB 24-205, and a Singapore bank needs under the MAS FEAT Principles. One system, several jurisdictions.
| Annex A | Control category | Controls |
|---|---|---|
| A.2 | Policies and organisational governance | 6 |
| A.3 | Internal organisation | 3 |
| A.4 | Resources for AI systems | 6 |
| A.5 | Impact assessment | 4 |
| A.6 | AI system lifecycle | 10 |
| A.7 | Data for AI systems | 6 |
| A.8 | Information for interested parties | 3 |
AI Verify and the Model Framework. Testing over prescription.
Developed by Singapore's Infocomm Media Development Authority and the AI Verify Foundation.
Singapore began with the Model AI Governance Framework in 2019 and built AI Verify, an open source testing toolkit that puts eleven internationally recognised governance principles into practice. Rather than codify duties in statute, Singapore offers principles, self-assessment instruments and a toolkit that produces a testing report.
The Model Framework's generative AI edition (2024) proposes nine governance dimensions: accountability, data, trusted development and deployment, incident reporting, testing and assurance, security, content provenance, safety and alignment research, and AI for public good. Deployers are expected to treat them as an operational agenda.
For cross-border operators the signal is strategic. The framework is narrower than the EU Act and less prescriptive than the Colorado statute, yet financial, health and platform regulators in the region refer to it, so meeting it in Singapore usually covers the minimum expected across much of Southeast Asia and supports a global posture.
| Dimension | Our reading |
|---|---|
| Accountability | Core |
| Data | Core |
| Trusted development and deployment | Core |
| Incident reporting | Core |
| Testing and assurance | Core |
| Security | Core |
| Content provenance | Developing |
| Safety and alignment research | Emerging |
| AI for public good | Horizon |
The OECD AI Principles and the treaty layer.
Adopted in 2019 and updated in 2024. The text most national frameworks cite when they explain themselves.
The OECD AI Principles are a recommendation, without force of law. They are the vocabulary the G7 Hiroshima Code, the Council of Europe Framework Convention, the UNESCO Recommendation on the Ethics of AI and most national strategies borrow from: five values and five policy recommendations for governments. They were updated in May 2024 to address foundation models and misinformation.
The Council of Europe Framework Convention on Artificial Intelligence, opened for signature in September 2024, turns several of these principles into binding obligations among the states that ratify it. It is the first international treaty on AI. Its operative articles impose duties on equality, privacy, accountability, safe development and remedies. It is not yet in force, and ratification continues.
For a global deployer these instruments matter less for immediate compliance than for interpretation. When a Colorado court or a Singapore regulator looks for the normative content behind a statutory term, it is likely to find it here.
| Principle (2024) | Field |
|---|---|
| Inclusive growth, sustainable development and well-being | Social |
| Human-centred values and fairness | Rights |
| Transparency and explainability | Procedural |
| Robustness, security and safety | Technical |
| Accountability | Governance |
The crosswalk. Where the four meet the EU AI Act.
A reduced mapping between the core duties of a high-risk deployer under the EU AI Act and the parallel outcomes in the four frameworks. Read it as reference, without legal equivalence. For Annex III systems these EU duties apply from 2 December 2027.
| EU AI Act duty | NIST AI RMF | ISO/IEC 42001 | AI Verify, generative AI | OECD |
|---|---|---|---|---|
| Risk management (Art. 9) | Map, Measure, Manage | A.5 Impact assessment | Testing and assurance | Robustness |
| Data governance (Art. 10) | Map 2.3 | A.7 Data for AI | Data | Transparency |
| Transparency (Art. 13) | Map 5, Manage 4 | A.8 Interested parties | Content provenance | Transparency |
| Human oversight (Art. 14) | Govern 4, Measure 2 | A.6.2 Lifecycle | Accountability | Accountability |
| Operator logs (Art. 26(6)) | Measure 4 | A.6.2.8 Logging | Incident reporting | Accountability |
| Fundamental rights impact assessment (Art. 27) | Map 1.1, 3.1 | A.5.2 Impact | AI Verify test report | Human-centred values |
The mapping is analytical. Equivalence is decided in each jurisdiction and does not replace local legal review.
How we read the field. Four duties every serious regime asks for.
Three features organise our coverage across borders, and the same four duties appear in almost every text we track.
Liability is accruing before any regime is mature. Operators deploying systems today carry duties that will only be litigated later, and the evidence they keep now is what a court will read then.
The texts converge more than they diverge. Governance, risk management, transparency to affected parties, human oversight, logging and incident response appear in almost every serious draft. A deployer that builds to the EU AI Act will meet much of what the Colorado AI Act, the Singapore Model Framework and the UK sector regulators expect. The frameworks differ at the edges; the centre is shared.
The insurance market follows the regulation. Specialty AI cover exists. It is fragmented and underwritten against the same operator duties the statutes describe. The sound posture is to harden the deployment and insure what remains.
| Duty | Requirement | Example |
|---|---|---|
| Risk management | A programme maintained over the system lifecycle | C.R.S. § 6-1-1703; AI Act Art. 9 |
| Human oversight | Competent people with authority to intervene | AI Act Art. 14 and 26(2) |
| Transparency | Notice to affected parties and traceable documentation | C.R.S. § 6-1-1703(5); NIST Map 5 |
| Logging | Monitoring and incident reporting procedures | AI Act Art. 26(6); NIST Measure |