The NIST AI RMF, formally the NIST AI Risk Management Framework, is a voluntary, sector-neutral framework published by the US National Institute of Standards and Technology that helps organizations identify, assess and reduce the risks that artificial intelligence systems pose to people, organizations and society. In the landscape of AI regulation, it is not a law: it creates no legal obligations, penalties or certification, yet it has become a common reference point that legislators, regulators and auditors use to judge whether an AI risk program is reasonable.
Version 1.0 was released in January 2023 as document NIST AI 100-1, following a mandate in the National Artificial Intelligence Initiative Act of 2020. The framework is deliberately use-case agnostic and rights-preserving, so it can be applied by a startup building a chatbot, a hospital deploying a diagnostic model or a bank buying a credit scoring tool.
How the AI RMF is structured
The AI RMF has two parts: a foundational part that frames AI risk and defines trustworthy AI, and the AI RMF Core, which turns those ideas into concrete outcomes. The Core is organized into four functions, each split into categories and subcategories that describe what a mature organization should achieve, without prescribing how.
Across the four functions there are 19 categories and 72 subcategories. The framework treats Govern as a cross-cutting function that informs the other three, while Map, Measure and Manage are applied iteratively throughout the AI lifecycle, from design and data collection to deployment, monitoring and decommissioning.
| Function | Purpose | Typical outcomes |
|---|---|---|
| Govern | Build a culture and structure for risk management | Policies, roles, accountability, third-party oversight |
| Map | Establish context and identify risks | Intended use, affected people, system categorization |
| Measure | Analyze and track identified risks | Test metrics, evaluations, independent review |
| Manage | Prioritize and act on measured risks | Treatment plans, incident response, decommissioning |
Characteristics of trustworthy AI
The AI RMF defines trustworthiness through seven characteristics that an AI system should balance. The framework stresses that these are socio-technical qualities: they depend on the deployment context and on human behavior around the system, not only on model properties.
Valid and reliable is described as the necessary base for all the others, while accountable and transparent spans the remaining characteristics. Improving one characteristic often costs another, for example explainability versus predictive accuracy or privacy versus fairness testing, and the framework expects organizations to document such trade-offs explicitly rather than claim to maximize all of them.
| Characteristic | Core question | Example evidence |
|---|---|---|
| Valid and reliable | Does the system perform as intended over time | Accuracy and robustness testing on representative data |
| Safe | Can it endanger life, health or property | Hazard analysis, fail-safe design, kill switches |
| Secure and resilient | Can it withstand attacks and disruption | Adversarial testing, access controls, recovery plans |
| Accountable and transparent | Is information about the system available | Model documentation, decision logs, clear ownership |
| Explainable and interpretable | Can outputs and mechanisms be understood | Feature attributions, user-facing explanations |
| Privacy-enhanced | Is personal autonomy and data protected | Data minimization, de-identification, privacy testing |
| Fair with harmful bias managed | Are disparate impacts identified and reduced | Subgroup performance analysis, bias mitigation records |
Profiles, the Playbook and the generative AI profile
The Core is intentionally generic, so NIST supplements it with companion resources. The AI RMF Playbook offers suggested actions, documentation ideas and references for each subcategory, and it is updated more often than the framework itself. Organizations are encouraged to build profiles: a current profile describes what they do today, a target profile describes the desired state, and the gap between them becomes a prioritized roadmap.
In July 2024 NIST published NIST AI 600-1, the Generative AI Profile, which applies the Core to large language models and image generators. It names twelve risks that are unique to or amplified by generative AI, including confabulation, information integrity, intellectual property and value chain and component integration, and maps suggested actions to the existing subcategories.
How the AI RMF compares with binding rules and standards
The AI RMF is best understood as a management toolkit, not a compliance regime. Its relationship to neighboring instruments in AI regulation matters because organizations often use it as the operational backbone for meeting obligations that come from elsewhere.
The EU AI Act imposes legally binding duties on providers and deployers of high-risk AI systems, enforced with fines. ISO/IEC 42001 specifies an AI management system that can be independently certified. The Colorado AI Act goes further in linking the two worlds: it treats a risk management program aligned with the NIST AI RMF or ISO/IEC 42001 as a benchmark, and recognized framework compliance supports an affirmative defense in enforcement.
| Instrument | Legal force | Scope | Verification |
|---|---|---|---|
| NIST AI RMF | Voluntary guidance | Any AI system, any sector | Self-assessment, no certificate |
| EU AI Act | Binding regulation | Risk-tiered, EU market | Conformity assessment, market surveillance |
| ISO/IEC 42001 | Voluntary standard | Organizational AI management system | Third-party certification audit |
| Colorado AI Act | Binding state law | High-risk consequential decisions | Attorney General enforcement |
NIST publishes crosswalks that map its subcategories to ISO/IEC 42001 and other instruments, so a single control can serve as evidence for several regimes.
Putting the AI RMF into practice
Adoption usually starts with Govern: assigning an accountable executive, setting a risk tolerance, and building an AI inventory that records every system, its owner, purpose and supplier. Without an inventory, shadow AI tools remain invisible.
Effort should scale with potential impact. The framework does not define risk tiers itself, so most organizations create their own, often borrowing categories such as consequential decisions about employment, credit, housing or health.
| Situation | Recommended approach | Reason |
|---|---|---|
| Internal productivity tool, low impact | Lightweight Map and periodic review | Limited harm potential |
| Customer-facing chatbot | Generative AI Profile actions, red-teaming | Confabulation and content risks |
| Hiring or lending model | Full Measure with subgroup testing | Discrimination and legal exposure |
| Vendor-supplied AI component | Govern third-party controls, contract terms | Limited visibility into the model |
Limitations and common misunderstandings
The flexibility that makes the AI RMF widely usable is also its main weakness. It provides no thresholds, no mandatory metrics and no pass or fail criteria, so two organizations can both claim alignment while operating very different levels of control.
- Alignment is not certification. No body certifies conformity with the AI RMF, so claims of compliance should be backed by documented evidence.
- It does not replace legal analysis. Sector laws on discrimination, privacy and consumer protection still apply in full.
- Measurement science is immature. Many trustworthiness properties, especially fairness and explainability, lack agreed metrics, as the framework itself acknowledges.
- Govern is often skipped. Teams that jump straight to model testing without clear ownership tend to produce reports that no one acts on.
In plain words
The NIST AI RMF is a free checklist of good habits for anyone who builds or uses artificial intelligence. Nobody fines you for ignoring it. Instead, it gives a shared vocabulary and a sequence of questions: who is responsible, what could go wrong, how do we test it, and what do we do when something breaks.
Think of it like a safety management handbook for a factory. The handbook does not build the machines, but it makes sure someone owns each risk, checks it regularly and fixes problems before people get hurt. Because so many laws now point to it, following it is often the easiest way to show you acted responsibly.
A real-world example
A regional health insurer in the United States plans to use a machine learning model to flag claims for manual review. Its compliance team uses the NIST AI RMF to structure the project. Under Govern, the chief risk officer becomes the accountable owner and the model is entered in the AI inventory as high impact, because wrong flags can delay care for patients.
Under Map, the team documents that the model affects members of different ages, regions and plan types. Under Measure, data scientists compare false positive rates across those groups and discover that claims from rural clinics are flagged about twice as often because of coding differences. Under Manage, the insurer retrains the model with normalized codes and adds a monthly drift report.
When a state regulator later asks how the insurer controls algorithmic bias, the team hands over the AI RMF profile, test results and remediation log. The documented process satisfies the inquiry without further action.