NIST AI RMF

What is NIST AI RMF?

The NIST AI RMF is the NIST AI Risk Management Framework, a voluntary framework from the US National Institute of Standards and Technology for managing risks that AI creates for individuals, organisations and society. It applies to AI actors across the AI lifecycle: organisations that design, develop, deploy or use AI systems.

The current version is AI RMF 1.0, published as NIST AI 100-1 in January 2023 (NIST's page gives 26 January 2023). NIST says the AI RMF 1.0 is being revised as part of the White House AI Action Plan, with no revised version number or date given. It also calls the framework a living document and expects a formal review, with input from the AI community, no later than 2028.

Two companions sit alongside it. The NIST AI RMF Playbook, also voluntary, gives suggested actions. NIST also published a Generative AI Profile (NIST AI 600-1) on 26 July 2024.

Why it matters

A customer, a board member or a regulator asks how you manage AI risk. You have a policy and good intentions, but nothing that shows the structure behind them.

The NIST AI RMF supplies that structure. It gives you a shared vocabulary for trustworthy AI and a way to describe where you are and where you intend to be. NIST expects AI risk management to be folded into enterprise risk management, so it is meant to sit inside the risk process you already run, not beside it as a separate programme.

An AI policy with no named owner is a statement of intent, not a control.

Who it applies to

The NIST AI RMF is voluntary. NIST describes it as intended for voluntary use, to improve how trustworthiness is built into the design, development, use and evaluation of AI products, services and systems. It is not a law, and nobody is required to adopt it.

Its audience is AI actors across the AI lifecycle. That reaches beyond model developers to organisations that deploy or use AI built by someone else. A firm with no AI in its products or operations has little reason to pick it up.

The nearest framework most UK firms hold is ISO 27001, a management system for information security that an accredited certification body can assess. The NIST AI RMF is a different kind of document: an outcome catalogue covering wider AI concerns such as harmful bias and explainability. NIST points to its Cybersecurity Framework, Privacy Framework, Risk Management Framework and Secure Software Development Framework as existing guidance on security and privacy risk that shares features with it.

The UK position

No UK law names the NIST AI RMF. Because NIST built it to be law- and regulation-agnostic, a UK firm can use it alongside UK and EU obligations, but using it does not itself meet any UK or EU legal requirement.

A UK firm meets it by choice: as a structure for its own AI governance, or because a customer asks about it. Firms weighing a certifiable alternative can read the ISO 42001 page.

Key requirements

The NIST AI RMF sets no requirements in the legal sense. It has two parts, and the second is where the working structure sits.

1. Part 1: framing. It covers how to frame AI risk, who the audience is, and the characteristics of trustworthy AI.
2. Part 2: the Core and Profiles. The Core is the set of outcomes you work towards, and Profiles describe your own use of it.

The full text is NIST's AI RMF 1.0 publication.

Characteristics of trustworthy AI

NIST lists seven characteristics that Part 1 sets out as the target:

• valid and reliable;
• safe;
• secure and resilient;
• accountable and transparent;
• explainable and interpretable;
• privacy-enhanced;
• fair, with harmful bias managed.

The four functions of the Core

The Core is made up of four functions, each divided into categories and subcategories:

• GOVERN sets the policies, accountability and culture for AI risk management. NIST describes it as cross-cutting: it runs through the other three functions and enables them.
• MAP establishes the context in which an AI system operates and identifies its risks.
• MEASURE analyses and tracks those risks.
• MANAGE prioritises risks and acts on them.

The Core is not a checklist. NIST says its actions are not necessarily an ordered set of steps either.

A framework with no pass mark can only be used well or badly.

Profiles

A Profile describes how an organisation applies the Core. NIST describes three kinds: use-case Profiles, temporal Profiles (current and target) and cross-sectoral Profiles. Comparing a Current Profile with a Target Profile shows the gaps to close. NIST does not prescribe profile templates. The Generative AI Profile (NIST AI 600-1) is a published example for one class of AI system.

Frequently asked questions