Artificial intelligence (AI) technologies hold tremendous potential to transform our lives and society for the better. However, as with any powerful technology, AI also comes with risks that must be carefully managed. To help organizations address these challenges, the National Institute of Standards and Technology (NIST) has developed a pioneering AI Risk Management Framework (AI RMF).

This voluntary framework provides practical guidance on identifying, assessing, and mitigating the risks of AI systems across their entire lifecycle. It offers a flexible, adaptable approach that organizations of any size or sector can use to implement responsible AI practices aligned with their specific values and priorities. This article will provide an overview of the AI RMF, its purpose and key attributes, fundamental concepts around AI risks and trustworthiness, intended audiences, and how it can enable organizations to unlock AI’s benefits while minimizing harms.

AI Risk Management Framework

The AI RMF was developed by NIST in collaboration with industry, academia, and civil society to meet a growing need for consensus guidance on AI risk management. It aligns with NIST’s broader National AI Initiative, as called for by the National Artificial Intelligence Initiative Act of 2020. The goals of the AI RMF are to:

• Offer a practical resource to help organizations manage AI risks and foster trustworthy AI development and use
• Provide common language around AI risks to enable effective communication across teams and stakeholders
• Highlight relevant standards, best practices, tools and methodologies for AI risk management
• Be flexible, accessible and applicable to diverse users across technology domains and industry sectors

The AI RMF aims to equip organizations with approaches to enhance AI system trustworthiness and responsible design, development and deployment over time. It emphasizes human-centric values like fairness, transparency, accountability and safety. The voluntary framework is non-prescriptive, offering a catalog of suggested outcomes rather than one-size-fits-all requirements.

Defining AI Risk

The AI RMF focuses on identifying and minimizing potential negative impacts of AI systems. It refers to risk as a composite measure of the probability and magnitude of harm from an AI-related event. Risks can range from impacts on individuals like loss of privacy, to societal harms like perpetuating unfair bias, to organizational risks like model failures.

While risks are the focus, the framework also highlights opportunities to maximize AI’s positive impacts on people, organizations and society. Responsible risk management practices can build trust and enable innovative uses of AI.

Trustworthy AI Characteristics

A core concept in the AI RMF is trustworthy AI. This refers to AI systems that responsibly manage risks and provide results that are valid, reliable, safe, secure, transparent, accountable, fair and privacy-enhanced.

The framework outlines seven key characteristics of trustworthy AI systems:

• Valid and Reliable: Results are accurate, relevant and can be trusted. The system functions reliably in expected conditions over its lifetime.
• Safe: Does not create unacceptable risks or endanger human health, safety or rights. Fails safely even in unfamiliar conditions.
• Secure and Resilient: Protects confidentiality, integrity and availability of the system and data through cybersecurity controls. Can recover effectively from attacks or failures.
• Accountable and Transparent: Provides appropriate visibility into how and why the system operates to enable oversight. Enables identification of issues and redress.
• Explainable and Interpretable: Users can understand and make sense of system outputs and functionality relative to their context.
• Privacy-Enhanced: Manages data ethically and implements safeguards to preserve privacy.
• Fair: Addresses unfair bias, accessibility limitations, and other issues that can lead to discriminatory or unjust impacts.

Who is the AI RMF for?

The AI RMF is intended for all individuals and organizations involved in the AI system lifecycle. This includes AI researchers, developers, users and operators, risk managers, policymakers, business leaders and other relevant stakeholders.

Specific roles called out in the framework include:

• AI Designers: Responsible for planning, objectives and data collection. Examples: data scientists, domain experts, human factors engineers.
• AI Developers: Build and interpret AI models. Examples: machine learning engineers, developers.
• AI Deployers: Implement AI systems into business processes. Examples: system integrators, end users.
• Operators: Run and monitor AI systems. Examples: IT professionals, compliance experts.
• TEVV (Testing, Evaluation, Verification, Validation): Assess AI systems through audits, monitoring and red teaming.
• Domain Experts: Provide expertise on industry sectors where AI is applied.
• Impact Assessors: Evaluate AI accountability, fairness, safety and other broader impacts.
• Procurement: Acquire AI systems, products and services.
• Governance: Establish policies, standards and controls around AI risks.

The framework emphasizes that diverse perspectives across these groups are essential for holistic risk management. It also notes the importance of engaging people who may be impacted by AI systems, like local communities and consumer advocates. Their input helps surface potential issues and blind spots.

Overview of the AI RMF Functions

The core of the AI RMF sets out four functions to put AI risk management into practice:

• Govern – Defines organizational policies, accountabilities, and structures to enable AI risk management.
• Map – Identifies AI risks, impacts and objectives in a specific application context.
• Measure – Assesses risks using qualitative, quantitative, and mixed methods.
• Manage – Prioritizes risks and implements controls to treat high-priority risks.

The Govern function is foundational for integrating AI risk management into organizational culture and business processes. The Map, Measure, and Manage functions apply across the stages of the AI system lifecycle to understand, evaluate and respond to risks in context.

Each function contains categories and subcategories with specific suggested outcomes. For example, the Measure function includes categories like “Appropriate methods and metrics are identified and applied” and “Mechanisms for tracking identified AI risks over time are in place.”

The framework connects effective risk management to concepts like transparency, diversity and responsible innovation. Practices like documentation, stakeholder input and impact assessments are woven throughout.

Using the Framework

Since the AI RMF is voluntary and non-prescriptive, organizations have flexibility in how they apply it. It can be used at a high level to shape policy and strategy or more tactically to assess and upgrade specific AI projects. Users can go through the complete framework or focus only on certain functions relevant to their priorities and resources.

For those just starting with AI risk management, the Govern and Map functions provide a strong foundation for establishing organizational structures, identifying objectives and assessing AI risks in context. As practices mature, the Measure and Manage functions help select metrics, monitor systems and implement controls tailored to top risks.

The AI RMF does not prescribed fixed controls or approval processes. Rather, it guides users in determining appropriate controls based on assessed risks for a given system, use case and tolerance thresholds. Organizations in regulated sectors like healthcare can integrate the framework alongside mandatory policies like HIPAA compliance.

NIST AI Framework Core Functions

Artificial intelligence (AI) brings tremendous promise along with new risks that require proactive management. The National Institute of Standards and Technology (NIST) AI Risk Management Framework (AI RMF) provides expert guidance grounded in industry consensus to help organizations realize benefits while minimizing AI harms. This article takes a deeper look into the framework’s four core functions for putting AI risk management into practice.

The AI RMF core sets out clear categories and outcomes to implement across the AI system lifecycle. The functions work synergistically to embed responsible AI through policies, risk assessments, measurement, controls and transparency. When used together, they provide a comprehensive approach customized to an organization’s needs and priorities. Let’s explore what each entails.

Govern – Creating a Foundation for AI Risk Management

The Govern function focuses on establishing organizational policies, structures and accountabilities to enable AI risk management. This strategic groundwork provides the foundation to integrate AI risk practices into business processes. Key activities under Govern include:

• Developing policies and procedures for AI risk management aligned with ethics and values. These should cover the entire AI lifecycle and set expectations at the leadership level.
• Defining clear roles and responsibilities for AI risk management. Cross-functional teams are emphasized to enable diverse input.
• Training personnel on AI risks and ensuring competency to uphold duties around transparency, bias evaluation etc.
• Prioritizing workforce diversity and inclusion. Multidisciplinary perspectives lead to better risk analysis and oversight.
• Creating mechanisms for internal and external feedback during development and post-deployment.
• Instituting controls around third-party AI systems and data including risks of intellectual property infringement.

Essentially, the Govern function establishes organizational guardrails and culture to enable the other core functions. Leadership buy-in and oversight are critical to reinforce its focus on responsibility. Documentation and human oversight help counter AI opacity. Workforce policies can mitigate risks of fragmented perspectives on complex AI systems.

Overall, this function aims to foster an AI risk management mindset reinforced through policies, business processes and stakeholder input.

Map – Understanding AI Risks in Context

The Map function entails identifying and analyzing AI risks, impacts and objectives within the specific application context. This upfront analysis provides crucial situational awareness that informs later risk measurement and management.

Key mapping activities include:

• Documenting the intended users, uses and environment where the AI system will operate.
• Describing applicable laws, regulations, sectoral practices and community expectations.
• Categorizing the system, tasks and methods such as classifiers or generative models.
• Defining processes for human oversight over AI system outputs and decisions.
• Assessing benefits, costs and capability tradeoffs compared to benchmarks.
• Evaluating operational, safety, bias and other risks for system components.
• Estimating likelihood and magnitude of potential impacts, including through external input.

This level of contextual grounding helps anticipate downstream issues and requirements early when mitigation is most effective. Mapping informs appropriate design choices and surfaces assumptions to be tested. Impact estimation drives risk prioritization.

By scoping objectives, use cases, affected groups, regulatory issues and technical risks upfront, organizations can preemptively address challenges. This proactive posture contrasts with reactively catching problems late in development when fixes are costly.

Measure – Evaluating and Monitoring AI Risks

The Measure function involves qualitative, quantitative and mixed methods to analyze, benchmark and monitor AI risks. Measurement provides data-driven insights to gauge system and process trustworthiness. Key practices include:

• Selecting metrics to evaluate risks and trustworthiness characteristics based on intended use and impacts.
• Employing test methods representative of the user population and deployment environment.
• Demonstrating system safety, accuracy, reliability and other attributes through statistical and operational performance benchmarks.
• Verifying AI system behavior across expected and unexpected conditions. Independent auditors provide objectivity.
• Implementing ongoing monitoring of risks like data drift, model degradation and system outages during production.
• Capturing user feedback and complaints to identify emerging issues.
• Tracking metrics over time to determine effectiveness of controls and guide improvements.

Taken together, these practices yield crucial visibility into how AI systems operate under real-world conditions. Measurement provides empirical evidence to ground risk reviews and oversight in data rather than assumptions. Continuous monitoring and user input enable rapid detection of failures or harms.

However, organizations should be strategic in their approach. Useful metrics tailored to priority risks and development milestones provide far more value than large metric dashboards. Consultations with users, domain experts and impacted groups inform appropriate measures for each system and context.

Manage – Controlling and Responding to Top Risks

The Manage function focuses on implementing preventive and corrective controls customized to top risks identified through mapping and measurement. Key practices involve:

• Making go/no-go deployment decisions based on assessed risks, costs and benefits.
• Prioritizing resource allocations to highest risks based on potential impact.
• Reducing unavoidable risks through strategies like containment, redundancy and recovery planning.
• Addressing emergent risks with defined incident response and system deactivation procedures.
• Monitoring and optimizing controls to maximize effectiveness and system value delivery.
• Communicating transparently with users on incidents, issues and improvements.
• Maintaining full documentation of risks assessed, prioritizations, controls, responses and outcomes.

Effective management stems from precise targeting of controls to risk severity and mitigation potential. Rather than blanket controls, impact-focused investments timed appropriately in development provide efficiency. Risk-based prioritization also helps justify when an AI application is unviable due to trustworthiness or fairness gaps unable to be closed.

Post-deployment vigilance to tune controls, act on issues and decommission unsafe systems prevents small failures from becoming large crises. It also builds user trust through communication and accountability.

Implementing the NIST AI Risk Framework

With artificial intelligence (AI) now embedded in products and services that impact lives, effective risk management is essential. The NIST AI Risk Management Framework (AI RMF) provides expert-informed guidance for organizations seeking to implement responsible AI practices. This article explores how users across sectors and roles can apply the framework to manage AI risks in line with their specific values and objectives.

The AI RMF aims to be flexible and adaptable for diverse organizations and use cases. As a voluntary, non-prescriptive framework, it serves as a toolbox offering suggested outcomes rather than mandated controls. By walking through real world scenarios, this article provides guidance on customizing and implementing the AI RMF based on an organization’s maturity, resources and needs.

For those just beginning with AI risk management, the NIST framework offers a logical pathway to build competency:

• Educate stakeholders at all levels on AI risks and ethical concerns to foster buy-in for the importance of responsible practices.
• Establish overarching values and principles to guide AI system development and use. These provide an ethical North Star rooted in human rights and the public good.
• Develop organizational policies and procedures that encode values into concrete governance expectations for AI systems.
• Define roles and responsibilities for AI risk management across teams including development, legal, risk, procurement and leadership.
• Prioritize diversity and multidisciplinary input to enable broad participation and perspective in AI oversight.

These foundational activities centered on governance and education establish guardrails aligned with organizational values. They equip people to assess AI risks and make values-based decisions throughout system lifecycles.

Governance controls like ethics review boards and AI system inventories can be implemented incrementally where needed most. The key is instilling an AI risk management mindset that permeates processes and culture.

Assessing AI Risks in Context

With basics in place, organizations can apply the NIST framework to assess and manage risks for specific AI projects or applications. A model workflow is:

• Detail the intended users, uses and environment for the AI system based on discussions with stakeholders.
• Identify applicable laws and regulations, community expectations and internal standards.
• Map system components, tasks, risks and potential impacts through workshops with diverse experts. Engage external groups to surface blindspots.
• Estimate risk likelihood and potential harms through impact assessment methodologies to prioritize focus areas.
• Select metrics and benchmarks to evaluate priority risks and trustworthiness factors.

This contextual grounding provides crucial input to refine objectives, surface assumptions and focus design choices on responsible outcomes. Consultations with domain experts, users and community representatives help spot issues early when more mitigation options exist.

Measuring and Managing Priority Risks

With priority risks and metrics defined, organizations can implement controls and oversight processes targeted to their situation:

• Evaluate risks through techniques like red teaming, simulations and audits representative of the deployment environment. Measurements should cover safety, accuracy, fairness, security and other factors.
• Implement training, monitoring, documentation and other controls directly responding to top risks based on severity. Use organizational risk appetite to guide mitigations.
• Develop contingency plans detailing responses if unsafe conditions emerge, including disabling components or entire systems when warranted.
• Provide transparency to users on system limitations, redress processes and improvements via accessible communications.
• Monitor effectiveness of mitigations through continuous risk assessments and user feedback. Enhance and adjust based on data.
• Maintain full documentation of risks assessed, measurement outcomes, controls implemented and incident response plans.

The goal is pragmatic customization of practices based on specific risks versus one-size-fits-all requirements. Start with higher scrutiny for public-facing or safety-critical applications. Mature approaches can later be scaled across the organization.

Adapting Implementation to Organizational Needs

Every organization faces unique constraints in adopting new practices. Teams can creatively adapt AI RMF guidance to their situation:

• Small companies can implement lightweight self-assessments, ethics reviews and monitoring pre-deployment. External advisory boards provide independent input.
• Startups can focus initial governance on priority risks like security, safety and fairness. As products and data grow more complex, controls can scale in maturity.
• Large enterprises can roll out tailored toolkits for decentralized teams to assess risks consistently across regions and business units. Central coordinators enable synergies and shared learning.
• AI software vendors can provide questionnaires, documentation and transparency tools to help customers manage downstream risks. Internally, automated monitoring as code aids rapid detection of production issues.
• Regulated sectors can integrate mandatory requirements like HIPAA with supplemental practices from the AI RMF tailored to their specialized risks.

In each case, the framework allows pragmatic application based on organizational constraints without compromising on core principles.

Sustaining Responsible AI Practices Over Time

Of course, risk management remains dynamic over technology lifecycles. Changes in applications, data sources, attacks and regulations necessitate constant vigilance. To embed practices over the long-term:

• Review policies and training programs periodically to incorporate evolving laws, standards and organizational lessons.
• Empower and resource internal audit teams to perform regular independent risk assessments using the AI RMF.
• Maintain secure repositories of risk documentation that product teams continually update from design through retirement.
• Monitor AI research and incidents for new attack vectors, vulnerabilities and controls to stay ahead of emergent threats.
• Foster an ethical culture encouraging stakeholders at all levels to ask tough questions and raise concerns.

By ingraining AI risk awareness into organizational culture, the necessary diligence to uphold safety and ethics sustains even as applications and portfolios scale exponentially.

The Road Ahead

Implementing the NIST AI RMF enables organizations to realize AI’s benefits while proactively managing risks. As a living document grounded in diverse expertise and input, it provides credible guidance poised to evolve with technology and societal expectations.

Organizations like yours play a crucial role in advancing AI responsibly. While no framework can address every situation, the AI RMF provides a values-based blueprint to custom-fit practices to your unique needs and context. The journey of a thousand miles begins with a single step – but you don’t have to take the first step alone.

On their own, each AI RMF function provides value in addressing specific areas of risk management. However, utilized together they offer an integrated strategy grounded in an organization’s unique objectives, context and constraints.

While the framework is not mandatory, pressures for responsible AI continue growing. By providing an adaptable roadmap informed by diverse expertise, the AI RMF enables organizations to get ahead of risks proactively.

Those embarking on the AI risk journey should start by building foundations through governance policies and risk evaluation frameworks. With basics established, organizations can incrementally expand measurement practices and controls prioritized for their unique needs and applications.

Of course, risk management remains dynamic over technology life cycles. By ingraining it as a core capability, organizations can confidently innovate with AI and uphold their duty of care to society. The AI RMF provides the tools to undertake this vital responsibility.

With AI adoption accelerating, better risk management is essential to avoid potentially catastrophic failures or abuses while enabling transformative innovation. The NIST AI RMF equips organizations with practical guidance grounded in shared expertise from industry, research and civil society. As a living document, it will continue evolving with the technology landscape to support responsible advancement of AI.

Widespread use of frameworks like this can help chart a future where AI risks are well-understood and managed, leading to trust, broad adoption and maximum societal benefit. Both developers and users of AI have crucial roles to play in this journey. Ultimately, responsible risk management will allow our society to confidently embrace AI’s immense potential while minimizing its dangers.