The NIST AI Risk Management Framework (AI RMF) is a voluntary guidance document published by the National Institute of Standards and Technology (NIST), a US federal agency. Released in January 2023, it provides a structured, process-based approach for organizations to identify, assess, and manage risks associated with artificial intelligence systems. The framework is designed to be applicable across sectors and organization sizes, offering a common language and set of practices for AI risk governance without imposing regulatory mandates. It complements, but does not replace, existing legal and regulatory obligations, and is intended to be used alongside other risk management standards.
The AI RMF was developed through a multi-stakeholder process initiated by NIST in response to a 2020 executive order on AI. It draws on existing risk management practices, such as those in cybersecurity and enterprise risk, and adapts them to the unique characteristics of AI, including its sociotechnical nature, data dependencies, and potential for unintended harms. The framework emphasizes that AI risks are not purely technical but also involve social, ethical, and organizational dimensions, and it encourages a proactive, continuous approach to risk management rather than a one-time compliance exercise.
Core Functions
The AI RMF is organized around four core functions: Govern, Map, Measure, and Manage. Each function represents a high-level category of activities that organizations should undertake to manage AI risks effectively. These functions are not intended to be performed in a strict linear sequence; rather, they are iterative and interrelated, with feedback loops that allow for continuous improvement.
The Govern function focuses on establishing the organizational structures, policies, and processes necessary for AI risk management. It includes activities such as defining risk management roles and responsibilities, integrating AI risk into broader enterprise risk management, and ensuring that AI systems align with organizational values and legal requirements. This function also addresses the need for transparency, accountability, and stakeholder engagement throughout the AI lifecycle.
The Map function involves understanding the context in which an AI system operates, including its intended purpose, the data it uses, and the potential impacts on individuals and communities. Mapping activities include identifying the system's capabilities and limitations, analyzing the sociotechnical context, and documenting assumptions and potential failure modes. This function helps organizations establish a baseline for risk assessment by clarifying what the system is designed to do and where it might go wrong.
The Measure function focuses on assessing and quantifying AI risks using a variety of methods. This includes developing metrics and evaluation procedures, testing the system under different conditions, and analyzing the results to determine the likelihood and severity of potential harms. Measurement can involve technical testing, such as evaluating model performance on machine learning benchmarks, as well as qualitative assessments that consider social and ethical factors. The framework encourages the use of both quantitative and qualitative approaches, recognizing that some risks are difficult to quantify.
The Manage function involves making decisions about how to respond to identified risks and implementing those decisions. This includes prioritizing risks, selecting mitigation strategies, and monitoring the effectiveness of those strategies over time. Management activities may involve redesigning the system, adding safeguards, providing human oversight, or discontinuing the system if risks are unacceptable. The framework emphasizes that risk management is an ongoing process, requiring regular review and adaptation as the system evolves or as new information becomes available.
Key Principles and Characteristics
The AI RMF is built on several key principles that guide its application. One of the most important is that AI risk management should be a sociotechnical endeavor, meaning it must consider the interplay between the technical system and the social context in which it is used. This principle recognizes that AI harms often arise from the interaction of algorithms with human behavior, institutional practices, and societal norms, and that technical fixes alone are insufficient.
Another principle is that risk management should be voluntary and flexible, allowing organizations to tailor the framework to their specific circumstances. The framework does not prescribe specific tools or techniques, but rather offers a set of practices that can be adapted to different sectors, use cases, and maturity levels. It also encourages organizations to go beyond minimum compliance and to strive for continuous improvement in their risk management capabilities.
The framework also emphasizes the importance of transparency and communication. Organizations are encouraged to document their risk management activities and to communicate with stakeholders, including users, regulators, and affected communities. This transparency helps build trust and enables external scrutiny, which can lead to better risk identification and mitigation.
Relationship to Other Standards and Regulations
The AI RMF is designed to complement existing standards and regulations, both in the US and internationally. It aligns with the general principles of the EU AI Act risk-based approach, which categorizes AI systems into different risk tiers and imposes corresponding obligations. However, the AI RMF is voluntary, whereas the EU AI Act is a binding regulation with specific requirements for high-risk AI systems. Organizations operating in multiple jurisdictions may use the AI RMF as a baseline for their risk management practices, then adapt to meet stricter legal requirements where applicable.
The framework also builds on earlier NIST work, such as the NIST Face Recognition Vendor Test, which evaluates the accuracy and bias of facial recognition algorithms. While that program focuses on specific technical performance, the AI RMF takes a broader view, addressing the full lifecycle of AI systems from design to deployment and decommissioning.
Development and Adoption
NIST initiated the AI RMF development process in 2020, following the issuance of an executive order on AI that directed the agency to create a framework for managing AI risks. The process involved a series of workshops, public comment periods, and consultations with experts from industry, academia, and civil society. The first version of the framework was released in January 2023, with an accompanying AI RMF Playbook that provides more detailed guidance and examples.
Since its release, the AI RMF has been widely adopted by organizations across various sectors, including technology companies, healthcare providers, financial institutions, and government agencies. Many organizations use it as a starting point for developing their own AI governance programs, and some have integrated it into their existing risk management frameworks. The framework has also influenced international discussions on AI governance, serving as a reference for other countries developing similar guidance.
Criticisms and Limitations
Despite its broad adoption, the AI RMF has faced criticism. Some critics argue that its voluntary nature is insufficient to address the potential harms of AI, particularly in high-stakes applications such as hiring, lending, and criminal justice. They contend that without mandatory requirements, organizations may not invest adequately in risk management, especially when there are competing business pressures.
Others have noted that the framework's generality can make it difficult to apply in practice, as it does not provide specific metrics or thresholds for acceptable risk. Organizations may struggle to interpret the framework's principles and translate them into concrete actions. Additionally, the framework has been criticized for not sufficiently addressing the power imbalances between AI developers and affected communities, and for relying too heavily on self-assessment rather than independent oversight.
Future Directions
NIST continues to evolve the AI RMF, with plans to update it based on feedback and emerging developments in AI technology. The agency has also developed related resources, such as the AI RMF Generative AI Profile, which addresses risks specific to generative AI systems like large language models. This profile provides additional guidance on managing risks such as misinformation, copyright infringement, and model hallucinations.
As AI technologies continue to advance, the AI RMF is likely to play an increasingly important role in shaping how organizations approach AI governance. Its emphasis on sociotechnical risk management and continuous improvement provides a flexible foundation that can adapt to new challenges, while its voluntary nature allows for innovation and experimentation. However, the framework's effectiveness will depend on the extent to which organizations genuinely commit to its principles and integrate them into their operations.
Conclusion
The NIST AI Risk Management Framework represents a significant effort to provide practical guidance for managing AI risks in a rapidly evolving field. By emphasizing a sociotechnical approach, continuous improvement, and stakeholder engagement, it offers a valuable resource for organizations seeking to develop responsible AI systems. While it is not without limitations, its influence on both US and international AI governance is likely to grow as AI becomes more pervasive in society.