The Artificial Intelligence Risk Management Framework, known as the AI RMF, is a guidance document from the National Institute of Standards and Technology (NIST), part of the U.S. Department of Commerce. According to NIST, it aims to support people who build, use and assess AI systems in managing AI risks more effectively. Those risks could affect individuals, organizations, society or the environment. Organizations are not required to use it.
Why AI gets its own framework
NIST says AI poses a number of risks that differ from those of traditional software. AI systems learn from data that may shift over time, at times by a large amount and without warning, which can affect the systems in ways that are hard to understand. NIST also calls these systems “socio-technical,” meaning they are influenced by societal dynamics and human behavior.
According to NIST, AI risks can emerge from the interplay of these technical and societal factors. They can affect people’s lives in situations ranging from online chatbots to the results of job and loan applications.
NIST was asked why a separate framework is needed when others already cover cybersecurity, privacy and enterprise risk management. It answers that each framework focuses on a specific set of challenges, with similarities and important differences. It adds that many people who design, deploy, evaluate or are affected by AI have called for specific guidance to help ensure AI is trustworthy and its risks are properly addressed during the AI lifecycle.
How the framework is organized
NIST says the AI RMF has two parts. The first covers how organizations can think about and frame AI-related risks, and it lays out the traits of trustworthy AI systems. The second part, which NIST calls the core of the framework, describes four functions: govern, map, measure and manage.
The agency says these functions help organizations address the risks of AI systems in practice. They can be applied in context-specific use cases and at any stage of the AI life cycle. NIST describes the overall approach as a flexible, structured and measurable process.
NIST also says the framework promotes a change in institutional culture. That includes how organizations think about, communicate, measure and monitor AI risks and their potential positive and negative impacts.
The characteristics of trustworthy AI
The framework names several characteristics of trustworthiness. They include being valid and reliable, being safe, being secure and resilient, being accountable and transparent, being explainable and interpretable, being privacy-enhanced, and being fair while harmful bias is managed.
NIST says users of the framework and other AI actors should consider these characteristics across these stages: before design, during design and development, at deployment, in use, and in testing and evaluation. The agency cautions that addressing the characteristics one at a time will not ensure that a system is trustworthy. Choices between them often involve tradeoffs, the full set seldom applies in any one setting, and the weight of each differs with the situation.
According to NIST, trustworthiness ultimately depends on every one of these characteristics and on how different AI actors and affected individuals and communities perceive them.
Who it is for and whether it is required
NIST says the framework should be useful to those who design, develop, use or evaluate AI technologies. It is written in language meant to be understandable to a broad audience, including senior executives and people who are not AI professionals. NIST says it still has enough technical depth to be useful to practitioners across many domains.
The agency says the framework should work for organizations of any size, whether public or private, in any sector, and whether they operate within a single country or across borders. Use is voluntary for both private and public sector organizations. NIST says Congress directed it to work with both private and public sectors on creating an AI RMF that is voluntary.
How it was developed and kept up to date
NIST says it developed the framework through an open, transparent and collaborative process. It issued a Request for Information on July 29, 2021, and a summary of the responses on Oct. 15, 2021. A concept paper followed on Dec. 13, 2021.
A draft was released on March 17, 2022, and discussed at a workshop March 29-31, 2022. A second draft came out on Aug. 18, 2022, followed by another workshop on Oct. 18-19, 2022. AI RMF 1.0 was released on Jan. 26, 2023.
According to NIST, the work took 18 months. The final document reflects about 400 sets of formal comments from more than 240 organizations on the drafts.
NIST calls AI RMF 1.0 a living document. It says it will review the framework’s content and usefulness regularly to decide whether an update is appropriate.
Companion resources
NIST released a companion Playbook alongside the framework. It is a voluntary resource with actionable suggestions drawn from industry best practices and research insights. NIST says organizations may borrow as many or as few suggestions as fit their industry use case or interests.
The Playbook is part of the NIST Trustworthy and Responsible AI Resource Center. NIST says it welcomes comments on the Playbook at any time and reviews them on a semi-annual basis.
NIST also encourages the development of Profiles, which describe ways to tailor the framework to a specific context. The agency says the framework is part of a broader portfolio of AI-related work that includes research, measurement and evaluation, technical standards, and contributions to AI policy.