Inside the NIST AI Risk Management Framework: Trustworthy AI Through Govern, Map, Measure, and Manage

The Framework was developed under the National Artificial Intelligence Initiative Act of 2020 and is meant to equip AI actors, defined as those playing an active role in the AI system lifecycle, with approaches that increase the trustworthiness of AI systems over time. It is designed to be practical, flexible across sectors and organization sizes, and to adapt as AI technologies evolve.P7¶34P7¶35

When considering negative impacts, risk is a function of the magnitude of harm that would arise if an event occurs and the likelihood of occurrence, and harm can be experienced by individuals, groups, communities, organizations, society, the environment, and the planet. Risk management is defined as coordinated activities to direct and control an organization with regard to risk.P9¶48P9¶49

Appendix B catalogs AI-specific risks that are new or increased, including training data that may misrepresent its context, drift and staleness of datasets, massive scale with billions or trillions of decision points, higher statistical uncertainty and bias-management issues from pre-trained models, privacy risk from enhanced data aggregation, increased opacity and reproducibility concerns, underdeveloped testing standards, and computational costs affecting the environment.

Existing frameworks, the document notes, do not yet comprehensively address harmful bias, generative AI risks, machine learning attacks, or third-party and off-label use.P43¶258P44¶274

These characteristics must be balanced against each other based on the system's context of use, since tradeoffs are usually involved, for example between predictive accuracy and interpretability or between privacy-enhancing techniques and accuracy under data sparsity. Validity and reliability is treated as a necessary condition of trustworthiness, and accountability and transparency relate to all other characteristics, but trustworthiness is only as strong as its weakest characteristic.P17¶98P17¶99

Systemic bias can be present in datasets, organizational norms and processes across the AI lifecycle, and broader society; computational and statistical bias often stems from systematic errors such as non-representative samples; and human-cognitive bias relates to how individuals or groups perceive AI system information when making decisions.

AI systems can increase the speed and scale of biases and amplify harms, and further detail is provided in NIST Special Publication 1270 on identifying and managing bias in AI.P23¶142P23¶143

GOVERN cultivates risk management culture, accountability structures, and policies; MAP establishes context and frames risks, ending in an initial go/no-go decision; MEASURE applies quantitative, qualitative, or mixed-method tools to analyze, assess, benchmark, and monitor risk and trustworthiness; and MANAGE allocates resources to prioritize, respond to, and recover from mapped and measured risks. Most users start with MAP after instituting GOVERN outcomes, but the process should be iterative with cross-referencing between functions as needed.P26¶164P27¶173

Risk tolerance is highly contextual and use-case specific, influenced by legal or regulatory requirements, policies, and norms, and likely to change over time. Where established sector guidelines do not exist, organizations should define reasonable risk tolerance themselves; only after tolerance is defined can the AI RMF be used to manage and document risk processes.P12¶68P12¶70

There is currently no consensus on robust and verifiable measurement methods for risk and trustworthiness applicable across use cases, and metrics can be oversimplified, gamed, lack nuance, or fail to account for differences among affected groups.

Risks measured earlier in the AI lifecycle may differ from those measured later, laboratory measurements may not match operational settings, and systems intended to augment or replace human decisions lack easy baseline metrics because AI performs tasks differently than humans.P11¶61P11¶63P11¶66

The first version number will change only with major revisions while minor revisions are tracked after the decimal, and all changes will be recorded in a Version Control Table. NIST plans frequent updates to the AI RMF Playbook, with comments accepted by email and integrated on a semi-annual basis, and the Framework will be aligned with applicable international standards as the AI landscape evolves.P3¶10P7¶37