[{"id": 520556, "type": "GenericWebPage", "mime_type": "text/html", "content_type": null, "url": "https://www.nist.gov/itl/ai-risk-management-framework", "slug": "nist", "title": "AI Risk Management Framework", "snippet": "Released on January 26, 2023, the Framework was developed through a consensus-driven, open, transparent, and collaborative process that included a Request for Information, several draft versions for public comments, multiple workshops, and other opportunities to provide input.", "image": null, "date_published": null, "media": [], "products": [], "source": {"id": 520556, "website_id": 1909, "url": "https://www.nist.gov/itl/ai-risk-management-framework", "title": "AI Risk Management Framework", "snippet": "Released on January 26, 2023, the Framework was developed through a consensus-driven, open, transparent, and collaborative process that included a Request for Information, several draft versions for public comments, multiple workshops, and other opportunities to provide input.", "cleaned_text": "NIST is seeking comments on a second draft of the NIST Artificial Intelligence Risk Management Framework (AI RMF). The AI RMF is intended for voluntary use in", "cleaned_text_custom_parser": "Official websites use .govA.govwebsite belongs to an official government organization in the United States.\nSecure .gov websites use HTTPSAlock(LockA locked padlock) orhttps://means you\u2019ve safely connected to the .gov website. Share sensitive information only on official, secure websites.\nAI Risk Management Framework\nQuick Links\nOverview of the AI RMF\nIn collaboration with the private and public sectors, NIST has developed a framework to better manage risks to individuals, organizations, and society associated with artificial intelligence (AI). TheNIST AI Risk Management Framework (AI RMF)is intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems.\nReleased on January 26, 2023, the Framework was developed through a consensus-driven, open, transparent, and collaborative process that included a Request for Information, several draft versions for public comments, multipleworkshops, and other opportunities to provide input. It is intended to build on, align with, and support AI risk management efforts by others (Fact Sheet).\nA companionNIST AI RMF Playbookalso has been published by NISTalong with\u00a0anAI RMF Roadmap,AI RMF Crosswalk,\u00a0and variousPerspectives.\nOn March 30, 2023, NIST launched theTrustworthy and Responsible AI Resource Center, which will facilitate implementation of, and international alignment with, the AI RMF.Examples of how other organizations are building on and using the AI RMF can be found via theAIRC\u2019s Use Case page.\nOn July 26, 2024, NIST releasedNIST-AI-600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. The profile \u00a0can help organizations identify unique risks posed by generative AI and\u00a0proposes\u00a0actions for generative AI risk management that best aligns with their goals and priorities.\nTo view public comments received on the previous drafts of the AI RMF and Requests for Information, see\u00a0theAI RMF Developmentpage.\nTranslations\nPrior Documents\nNews & Updates\nNIST Risk Management Framework Aims to Improve Trustworthiness of Artificial Intelligence\nNIST Seeks Comments on AI Risk Management Framework Guidance, Workshop Date Set\nSymposium Focuses on AI and the Economy and the Path for Responsible and Inclusive AI\nNIST Seeks Comments on Draft AI Risk Management Framework, Offers Guidance on AI Bias\nHEADQUARTERS", "cleaned_text_custom_parser_2": "\n\n \n \n Skip to main content\n \n

\n $\"\"$ \n

\n \n Official websites use .gov\n \n
\n A\n \n .gov\n \n website belongs to an official government organization in the United States.\n

\n $\"\"$ \n

\n \n Secure .gov websites use HTTPS\n \n
\n A\n \n lock\n \n (\n \n ) or\n \n https://\n \n means you\u2019ve safely connected to the .gov website. Share sensitive information only on official, secure websites.\n

\n https://www.nist.gov/itl/ai-risk-management-framework\n

\n $\"National$ \n \n

\n \n Information Technology Laboratory\n \n

\n AI Risk Management Framework\n

\n Quick Links\n

\n \n Download the AI RMF 1.0\n \n \n View the AI RMF Playbook\n \n \n Visit the AI Resource Center\n \n

\n Overview of the AI RMF\n

\n In collaboration with the private and public sectors, NIST has developed a framework to better manage risks to individuals, organizations, and society associated with artificial intelligence (AI). The\n \n NIST AI Risk Management Framework (AI RMF)\n \n is intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems.\n

\n Released on January 26, 2023, the Framework was developed through a consensus-driven, open, transparent, and collaborative process that included a Request for Information, several draft versions for public comments, multiple\n \n workshops\n \n , and other opportunities to provide input. It is intended to build on, align with, and support AI risk management efforts by others (\n \n Fact Sheet\n \n ).\n

\n A companion\n \n NIST AI RMF Playbook\n \n also has been published by NIST\n along with\u00a0an\n \n AI RMF Roadmap\n \n ,\n \n AI RMF Crosswalk\n \n ,\u00a0and various\n \n Perspectives\n \n .\n

\n On March 30, 2023, NIST launched the\n \n Trustworthy and Responsible AI Resource Center\n \n , which will facilitate implementation of, and international alignment with, the AI RMF.\n Examples of how other organizations are building on and using the AI RMF can be found via the\n \n AIRC\u2019s Use Case page\n \n .\n

\n On July 26, 2024, NIST released\n \n NIST-AI-\n \n 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile\n \n \n . The profile \u00a0can help organizations identify unique risks posed by generative AI and\u00a0proposes\u00a0actions for generative AI risk management that best aligns with their goals and priorities.\n

\n To view public comments received on the previous drafts of the AI RMF and Requests for Information, see\u00a0the\n \n AI RMF Development\n \n page.\n

\n \n Translations\n \n

\n \n NIST AI Risk Management Framework\n \n (Arabic)\n
\n \n NIST AI Risk Management Framework\n \n (Japanese)\n

\n \n Prior Documents\n \n

\n \n Second draft of the AI Risk Management Framework\n \n (August 18, 2022)\n
\n \n Initial draft of the AI Risk Management Framework\n \n (March 17, 2022)\n
\n \n Concept paper to help guide development of the AI Risk Management Framework\n \n (December 13, 2021)\n
\n \n Brief summary\n \n of\n \n responses\n \n to the July 29, 2021, RFI (October 15, 2021)\n
\n \n Draft -Taxonomy of AI Risk\n \n (October 15, 2021)\n
\n \n AI Risk Management Framework Request for Information\n \n (July 29, 2021)\n

\n News & Updates\n

\n New guidance seeks to cultivate trust in AI technologies and promote AI innovation while mitigating risk.\n

\n NIST is seeking comments on a second draft of the NIST Artificial Intelligence Risk Management Framework (AI RMF). The AI RMF is intended for voluntary use in\n

\n The U.S. Department of Commerce, NIST, FinRegLab, and the Stanford Institute for Human-Centered Artificial Intelligence (HAI) hosted a symposium on \"Artificial\n

\n Seeking to promote the development and use of artificial intelligence (AI) technologies and systems that are trustworthy and responsible, NIST today released\n

\n \n Was this page helpful?\n \n

\n \n\n", "best_field": "cleaned_text_custom_parser", "enhanced": true, "date_published": null, "image": null, "num_tokens_oai_tokenizer": null, "num_tokens_gemini_tokenizer": null}, "favicon": "https://www.nist.gov/themes/custom/nist_www/favicon.ico", "state": "candidate", "enhanced": true, "answer": null, "valid_answer": false, "source_citation_index": 1, "user_web_page_id": null, "status": "ready", "reader": false, "rank": 1, "reason": "The source is highly relevant to the query because it provides direct access to the official AI Risk Management Framework page from NIST."}, {"id": 648327, "type": "GenericWebPage", "mime_type": "text/html", "content_type": null, "url": "https://www.nist.gov/itl/ai-risk-management-framework/nist-ai-rmf-playbook", "slug": "nist", "title": "NIST AI RMF Playbook | NIST", "snippet": "In collaboration with the private ... which suggests ways to navigate and use the AI Risk Management Framework (AI RMF) to incorporate trustworthiness considerations in the design, development, deployment, and use of AI systems...", "image": null, "date_published": null, "media": [], "products": [], "source": {"id": 648327, "website_id": 1909, "url": "https://www.nist.gov/itl/ai-risk-management-framework/nist-ai-rmf-playbook", "title": "NIST AI RMF Playbook | NIST", "snippet": "In collaboration with the private ... which suggests ways to navigate and use the AI Risk Management Framework (AI RMF) to incorporate trustworthiness considerations in the design, development, deployment, and use of AI systems...", "cleaned_text": null, "cleaned_text_custom_parser": null, "cleaned_text_custom_parser_2": null, "best_field": null, "enhanced": false, "date_published": null, "image": null, "num_tokens_oai_tokenizer": null, "num_tokens_gemini_tokenizer": null}, "favicon": "https://www.nist.gov/themes/custom/nist_www/favicon.ico", "state": "candidate", "enhanced": false, "answer": null, "valid_answer": false, "source_citation_index": 2, "user_web_page_id": null, "status": "not_processed", "reader": false, "rank": 1, "reason": "The source is highly relevant as it is the official NIST AI RMF Playbook, which offers guidance on using the framework effectively."}, {"id": 648328, "type": "GenericWebPage", "mime_type": "text/html", "content_type": null, "url": "https://www.nist.gov/itl/ai-risk-management-framework/crosswalks-nist-artificial-intelligence-risk-management-framework", "slug": "nist", "title": "Crosswalks to the NIST Artificial Intelligence Risk Management Framework (AI RMF 1.0) | NIST", "snippet": "NIST has prepared several crosswalks to the Artificial Intelligence Risk Management Framework (AI RMF 1.0). As directed by the National Artificial Intelligence Initiative Act of 2020 (P.L.", "image": null, "date_published": null, "media": [], "products": [], "source": {"id": 648328, "website_id": 1909, "url": "https://www.nist.gov/itl/ai-risk-management-framework/crosswalks-nist-artificial-intelligence-risk-management-framework", "title": "Crosswalks to the NIST Artificial Intelligence Risk Management Framework (AI RMF 1.0) | NIST", "snippet": "NIST has prepared several crosswalks to the Artificial Intelligence Risk Management Framework (AI RMF 1.0). As directed by the National Artificial Intelligence Initiative Act of 2020 (P.L.", "cleaned_text": null, "cleaned_text_custom_parser": null, "cleaned_text_custom_parser_2": null, "best_field": null, "enhanced": false, "date_published": null, "image": null, "num_tokens_oai_tokenizer": null, "num_tokens_gemini_tokenizer": null}, "favicon": "https://www.nist.gov/themes/custom/nist_www/favicon.ico", "state": "candidate", "enhanced": false, "answer": null, "valid_answer": false, "source_citation_index": 3, "user_web_page_id": null, "status": "not_processed", "reader": false, "rank": 1, "reason": "The source is highly relevant as it provides official NIST crosswalks to the AI Risk Management Framework, aligning it with other guidelines."}, {"id": 561731, "type": "GenericWebPage", "mime_type": "text/html", "content_type": null, "url": "https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf", "slug": "nist", "title": null, "snippet": "Artificial Intelligence Risk Management \u00b7 Framework (AI RMF 1.0)", "image": null, "date_published": null, "media": [], "products": [], "source": {"id": 561731, "website_id": 95431, "url": "https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf", "title": null, "snippet": "Artificial Intelligence Risk Management \u00b7 Framework (AI RMF 1.0)", "cleaned_text": null, "cleaned_text_custom_parser": "NIST AI 100-1\nArtificial Intelligence Risk Management\nFramework (AI RMF 1.0)\n NIST AI 100-1\nArtificial Intelligence Risk Management\nFramework (AI RMF 1.0)\nThis publication is available free of charge from:\nhttps://doi.org/10.6028/NIST.AI.100-1\nJanuary 2023\nU.S. Department of Commerce\nGina M. Raimondo, Secretary\nNational Institute of Standards and Technology\nLaurie E. Locascio, NIST Director and Under Secretary of Commerce for Standards and Technology Certain commercial entities, equipment, or materials may be identified in this document in order to describe \nan experimental procedure or concept adequately. Such identification is not intended to imply recommenda-\ntion or endorsement by the National Institute of Standards and Technology, nor is it intended to imply that \nthe entities, materials, or equipment are necessarily the best available for the purpose. \nThis publication is available free of charge from: https://doi.org/10.6028/NIST.AI.100-1\nUpdate Schedule and Versions\nThe Artificial Intelligence Risk Management Framework (AI RMF) is intended to be a living document.\nNIST will review the content and usefulness of the Framework regularly to determine if an update is appro-\npriate; a review with formal input from the AI community is expected to take place no later than 2028. The\nFramework will employ a two-number versioning system to track and identify major and minor changes. The\nfirst number will represent the generation of the AI RMF and its companion documents (e.g., 1.0) and will\nchange only with major revisions. Minor revisions will be tracked using \u201c.n\u201d after the generation number\n(e.g., 1.1). All changes will be tracked using a Version Control Table which identifies the history, including\nversion number, date of change, and description of change. NIST plans to update the AI RMF Playbook\nfrequently. Comments on the AI RMF Playbook may be sent via email to AIframework@nist.gov at any time\nand will be reviewed and integrated on a semi-annual basis. Table of Contents\nExecutive Summary 1\nPart 1: Foundational Information 4\n1 Framing Risk 4\n1.1 Understanding and Addressing Risks, Impacts, and Harms 4\n1.2 Challenges for AI Risk Management 5\n1.2.1 Risk Measurement 5\n1.2.2 Risk Tolerance 7\n1.2.3 Risk Prioritization 7\n1.2.4 Organizational Integration and Management of Risk 8\n2 Audience 9\n3 AI Risks and Trustworthiness 12\n3.1 Valid and Reliable 13\n3.2 Safe 14\n3.3 Secure and Resilient 15\n3.4 Accountable and Transparent 15\n3.5 Explainable and Interpretable 16\n3.6 Privacy-Enhanced 17\n3.7 Fair \u2013 with Harmful Bias Managed 17\n4 Effectiveness of the AI RMF 19\nPart 2: Core and Profiles 20\n5 AI RMF Core 20\n5.1 Govern 21\n5.2 Map 24\n5.3 Measure 28\n5.4 Manage 31\n6 AI RMF Profiles 33\nAppendix A: Descriptions of AI Actor Tasks from Figures 2 and 3 35\nAppendix B: How AI Risks Differ from Traditional Software Risks 38\nAppendix C: AI Risk Management and Human-AI Interaction 40\nAppendix D: Attributes of the AI RMF 42\nList of Tables\nTable 1 Categories and subcategories for the GOVERN function. 22\nTable 2 Categories and subcategories for the MAP function. 26\nTable 3 Categories and subcategories for the MEASURE function. 29\nTable 4 Categories and subcategories for the MANAGE function. 32\ni NIST AI 100-1 AI RMF 1.0\nList of Figures\nFig. 1 Examples of potential harms related to AI systems. Trustworthy AI systems\nand their responsible use can mitigate negative risks and contribute to bene-\nfits for people, organizations, and ecosystems. 5\nFig. 2 Lifecycle and Key Dimensions of an AI System. Modified from OECD\n(2022) OECD Framework for the Classification of AI systems \u2014 OECD\nDigital Economy Papers. The two inner circles show AI systems\u2019 key di-\nmensions and the outer circle shows AI lifecycle stages. Ideally, risk man-\nagement efforts start with the Plan and Design function in the application\ncontext and are performed throughout the AI system lifecycle. See Figure 3\nfor representative AI actors. 10\nFig. 3 AI actors across AI lifecycle stages. See Appendix A for detailed descrip-\ntions of AI actor tasks, including details about testing, evaluation, verifica-\ntion, and validation tasks. Note that AI actors in the AI Model dimension\n(Figure 2) are separated as a best practice, with those building and using the\nmodels separated from those verifying and validating the models. 11\nFig. 4 Characteristics of trustworthy AI systems. Valid & Reliable is a necessary\ncondition of trustworthiness and is shown as the base for other trustworthi-\nness characteristics. Accountable & Transparent is shown as a vertical box\nbecause it relates to all other characteristics. 12\nFig. 5 Functions organize AI risk management activities at their highest level to\ngovern, map, measure, and manage AI risks. Governance is designed to be\na cross-cutting function to inform and be infused throughout the other three\nfunctions. 20\nPage ii NIST AI 100-1 AI RMF 1.0\nExecutive Summary\nArtificial intelligence (AI) technologies have significant potential to transform society and\npeople\u2019s lives \u2013 from commerce and health to transportation and cybersecurity to the envi-\nronment and our planet. AI technologies can drive inclusive economic growth and support\nscientific advancements that improve the conditions of our world. AI technologies, how-\never, also pose risks that can negatively impact individuals, groups, organizations, commu-\nnities, society, the environment, and the planet. Like risks for other types of technology, AI\nrisks can emerge in a variety of ways and can be characterized as long- or short-term, high-\nor low-probability, systemic or localized, and high- or low-impact.\nThe AI RMF refers to an AI system as an engineered or machine-based system that\ncan, for a given set of objectives, generate outputs such as predictions, recommenda-\ntions, or decisions influencing real or virtual environments. AI systems are designed\nto operate with varying levels of autonomy (Adapted from: OECD Recommendation\non AI:2019; ISO/IEC22989:2022).\nWhile there are myriad standards and best practices to help organizations mitigate the risks\nof traditional software or information-based systems, the risks posed by AI systems are in\nmany ways unique (See Appendix B). AI systems, for example, may be trained on data that\ncan change over time, sometimes significantly and unexpectedly, affecting system function-\nality and trustworthiness in ways that are hard to understand. AI systems and the contexts\nin which they are deployed are frequently complex, making it difficult to detect and respond\nto failures when they occur. AI systems are inherently socio-technical in nature, meaning\nthey are influenced by societal dynamics and human behavior. AI risks \u2013 and benefits \u2013\ncan emerge from the interplay of technical aspects combined with societal factors related\nto how a system is used, its interactions with other AI systems, who operates it, and the\nsocial context in which it is deployed.\nThese risks make AI a uniquely challenging technology to deploy and utilize both for orga-\nnizations and within society. Without proper controls, AI systems can amplify, perpetuate,\nor exacerbate inequitable or undesirable outcomes for individuals and communities. With\nproper controls, AI systems can mitigate and manage inequitable outcomes.\nAI risk management is a key component of responsible development and use of AI sys-\ntems. Responsible AI practices can help align the decisions about AI system design, de-\nvelopment, and uses with intended aim and values. Core concepts in responsible AI em-\nphasize human centricity, social responsibility, and sustainability. AI risk management can\ndrive responsible uses and practices by prompting organizations and their internal teams\nwho design, develop, and deploy AI to think more critically about context and potential\nor unexpected negative and positive impacts. Understanding and managing the risks of AI\nsystems will help to enhance trustworthiness, and in turn, cultivate public trust.\nPage 1 NIST AI 100-1 AI RMF 1.0\nSocial responsibility can refer to the organization\u2019s responsibility \u201cfor the impacts\nof its decisions and activities on society and the environment through transparent\nand ethical behavior\u201d ( ISO26000:2010). Sustainability refers to the \u201cstate of the\nglobal system, including environmental, social, and economic aspects, in which the\nneeds of the present are met without compromising the ability of future generations\nto meet their own needs\u201d ( ISO/IEC TR 24368:2022). Responsible AI is meant to\nresult in technology that is also equitable and accountable. The expectation is that\norganizational practices are carried out in accord with \u201c professional responsibility ,\u201d\ndefined by ISOas an approach that \u201caims to ensure that professionals who design,\ndevelop, or deploy AI systems and applications or AI-based products or systems,\nrecognize their unique position to exert influence on people, society, and the future\nof AI\u201d ( ISO/IEC TR 24368:2022).\nAs directed by the National Artificial Intelligence Initiative Act of 2020 (P.L. 116-283),\nthe goal of the AI RMF is to offer a resource to the organizations designing, developing,\ndeploying, or using AI systems to help manage the many risks of AI and promote trustwor-\nthy and responsible development and use of AI systems. The Framework is intended to be\nvoluntary , rights-preserving, non-sector-specific, and use-case agnostic, providing flexibil-\nity to organizations of all sizes and in all sectors and throughout society to implement the\napproaches in the Framework.\nThe Framework is designed to equip organizations and individuals \u2013 referred to here as\nAI actors \u2013 with approaches that increase the trustworthiness of AI systems, and to help\nfoster the responsible design, development, deployment, and use of AI systems over time.\nAI actors are defined by the Organisation for Economic Co-operation and Development\n(OECD) as \u201cthose who play an active role in the AI system lifecycle, including organiza-\ntions and individuals that deploy or operate AI\u201d [OECD (2019) Artificial Intelligence in\nSociety\u2014OECD iLibrary] (See Appendix A).\nThe AI RMF is intended to be practical, to adapt to the AI landscape as AI technologies\ncontinue to develop, and to be operationalized by organizations in varying degrees and\ncapacities so society can benefit from AI while also being protected from its potential\nharms.\nThe Framework and supporting resources will be updated, expanded, and improved based\non evolving technology, the standards landscape around the world, and AI community ex-\nperience and feedback. NIST will continue to align the AI RMF and related guidance with\napplicable international standards, guidelines, and practices. As the AI RMF is put into\nuse, additional lessons will be learned to inform future updates and additional resources.\nThe Framework is divided into two parts. Part 1 discusses how organizations can frame\nthe risks related to AI and describes the intended audience. Next, AI risks and trustworthi-\nness are analyzed, outlining the characteristics of trustworthy AI systems, which include\nPage 2 NIST AI 100-1 AI RMF 1.0\nvalid and reliable, safe, secure and resilient, accountable and transparent, explainable and\ninterpretable, privacy enhanced, and fair with their harmful biases managed.\nPart 2 comprises the \u201cCore\u201d of the Framework. It describes four specific functions to help\norganizations address the risks of AI systems in practice. These functions \u2013 GOVERN ,\nMAP ,MEASURE , and MANAGE \u2013 are broken down further into categories and subcate-\ngories. While GOVERN applies to all stages of organizations\u2019 AI risk management pro-\ncesses and procedures, the MAP ,MEASURE , and MANAGE functions can be applied in AI\nsystem-specific contexts and at specific stages of the AI lifecycle.\nAdditional resources related to the Framework are included in the AI RMF Playbook,\nwhich is available via the NIST AI RMF website:\nhttps://www.nist.gov/itl/ai-risk-management-framework.\nDevelopment of the AI RMF by NIST in collaboration with the private and public sec-\ntors is directed and consistent with its broader AI efforts called for by the National AI\nInitiative Act of 2020, the National Security Commission on Artificial Intelligence recom-\nmendations, and the Plan for Federal Engagement in Developing Technical Standards and\nRelated Tools. Engagement with the AI community during this Framework\u2019s development\n\u2013 via responses to a formal Request for Information, three widely attended workshops,\npublic comments on a concept paper and two drafts of the Framework, discussions at mul-\ntiple public forums, and many small group meetings \u2013 has informed development of the AI\nRMF 1.0 as well as AI research and development and evaluation conducted by NIST and\nothers. Priority research and additional guidance that will enhance this Framework will be\ncaptured in an associated AI Risk Management Framework Roadmap to which NIST and\nthe broader community can contribute.\nPage 3 NIST AI 100-1 AI RMF 1.0\nPart 1: Foundational Information\n1. Framing Risk\nAI risk management offers a path to minimize potential negative impacts of AI systems,\nsuch as threats to civil liberties and rights, while also providing opportunities to maximize\npositive impacts. Addressing, documenting, and managing AI risks and potential negative\nimpacts effectively can lead to more trustworthy AI systems.\n1.1 Understanding and Addressing Risks, Impacts, and Harms\nIn the context of the AI RMF, riskrefers to the composite measure of an event\u2019s probability\nof occurring and the magnitude or degree of the consequences of the corresponding event.\nThe impacts, or consequences, of AI systems can be positive, negative, or both and can\nresult in opportunities or threats (Adapted from: ISO31000:2018). When considering the\nnegative impact of a potential event, risk is a function of 1) the negative impact, or magni-\ntude of harm, that would arise if the circumstance or event occurs and 2) the likelihood of\noccurrence (Adapted from: OMB Circular A-130:2016). Negative impact or harm can be\nexperienced by individuals, groups, communities, organizations, society, the environment,\nand the planet.\n\u201cRisk management refers to coordinated activities to direct and control an organiza-\ntion with regard to risk\u201d (Source: ISO31000:2018).\nWhile risk management processes generally address negative impacts, this Framework of-\nfers approaches to minimize anticipated negative impacts of AI systems andidentify op-\nportunities to maximize positive impacts. Effectively managing the risk of potential harms\ncould lead to more trustworthy AI systems and unleash potential benefits to people (individ-\nuals, communities, and society), organizations, and systems/ecosystems. Risk management\ncan enable AI developers and users to understand impacts and account for the inherent lim-\nitations and uncertainties in their models and systems, which in turn can improve overall\nsystem performance and trustworthiness and the likelihood that AI technologies will be\nused in ways that are beneficial.\nThe AI RMF is designed to address new risks as they emerge. This flexibility is particularly\nimportant where impacts are not easily foreseeable and applications are evolving. While\nsome AI risks and benefits are well-known, it can be challenging to assess negative impacts\nand the degree of harms. Figure 1 provides examples of potential harms that can be related\nto AI systems.\nAI risk management efforts should consider that humans may assume that AI systems work\n\u2013 and work well \u2013 in allsettings. For example, whether correct or not, AI systems are\noften perceived as being more objective than humans or as offering greater capabilities\nthan general software.\nPage 4 NIST AI 100-1 AI RMF 1.0\nFig. 1. Examples of potential harms related to AI systems. Trustworthy AI systems and their\nresponsible use can mitigate negative risks and contribute to benefits for people, organizations, and\necosystems.\n1.2 Challenges for AI Risk Management\nSeveral challenges are described below. They should be taken into account when managing\nrisks in pursuit of AI trustworthiness.\n1.2.1 Risk Measurement\nAI risks or failures that are not well-defined or adequately understood are difficult to mea-\nsure quantitatively or qualitatively. The inability to appropriately measure AI risks does not\nimply that an AI system necessarily poses either a high or low risk. Some risk measurement\nchallenges include:\nRisks related to third-party software, hardware, and data: Third-party data or systems\ncan accelerate research and development and facilitate technology transition. They also\nmay complicate risk measurement. Risk can emerge both from third-party data, software or\nhardware itself and how it is used. Risk metrics or methodologies used by the organization\ndeveloping the AI system may not align with the risk metrics or methodologies uses by\nthe organization deploying or operating the system. Also, the organization developing\nthe AI system may not be transparent about the risk metrics or methodologies it used. Risk\nmeasurement and management can be complicated by how customers use or integrate third-\nparty data or systems into AI products or services, particularly without sufficient internal\ngovernance structures and technical safeguards. Regardless, all parties and AI actors should\nmanage risk in the AI systems they develop, deploy, or use as standalone or integrated\ncomponents.\nTracking emergent risks: Organizations\u2019 risk management efforts will be enhanced by\nidentifying and tracking emergent risks and considering techniques for measuring them.\nPage 5 NIST AI 100-1 AI RMF 1.0\nAI system impact assessment approaches can help AI actors understand potential impacts\nor harms within specific contexts.\nAvailability of reliable metrics: The current lack of consensus on robust and verifiable\nmeasurement methods for risk and trustworthiness, and applicability to different AI use\ncases, is an AI risk measurement challenge. Potential pitfalls when seeking to measure\nnegative risk or harms include the reality that development of metrics is often an institu-\ntional endeavor and may inadvertently reflect factors unrelated to the underlying impact. In\naddition, measurement approaches can be oversimplified, gamed, lack critical nuance, be-\ncome relied upon in unexpected ways, or fail to account for differences in affected groups\nand contexts.\nApproaches for measuring impacts on a population work best if they recognize that contexts\nmatter, that harms may affect varied groups or sub-groups differently, and that communities\nor other sub-groups who may be harmed are not always direct users of a system.\nRisk at different stages of the AI lifecycle: Measuring risk at an earlier stage in the AI\nlifecycle may yield different results than measuring risk at a later stage; some risks may\nbe latent at a given point in time and may increase as AI systems adapt and evolve. Fur-\nthermore, different AI actors across the AI lifecycle can have different risk perspectives.\nFor example, an AI developer who makes AI software available, such as pre-trained mod-\nels, can have a different risk perspective than an AI actor who is responsible for deploying\nthat pre-trained model in a specific use case. Such deployers may not recognize that their\nparticular uses could entail risks which differ from those perceived by the initial developer.\nAll involved AI actors share responsibilities for designing, developing, and deploying a\ntrustworthy AI system that is fit for purpose.\nRisk in real-world settings: While measuring AI risks in a laboratory or a controlled\nenvironment may yield important insights pre-deployment, these measurements may differ\nfrom risks that emerge in operational, real-world settings.\nInscrutability: Inscrutable AI systems can complicate risk measurement. Inscrutability\ncan be a result of the opaque nature of AI systems (limited explainability or interpretabil-\nity), lack of transparency or documentation in AI system development or deployment, or\ninherent uncertainties in AI systems.\nHuman baseline: Risk management of AI systems that are intended to augment or replace\nhuman activity, for example decision making, requires some form of baseline metrics for\ncomparison. This is difficult to systematize since AI systems carry out different tasks \u2013 and\nperform tasks differently \u2013 than humans.\nPage 6 NIST AI 100-1 AI RMF 1.0\n1.2.2 Risk Tolerance\nWhile the AI RMF can be used to prioritize risk, it does not prescribe risk tolerance. Risk\ntolerance refers to the organization\u2019s or AI actor\u2019s (see Appendix A) readiness to bear the\nrisk in order to achieve its objectives. Risk tolerance can be influenced by legal or regula-\ntory requirements (Adapted from: ISO GUIDE 73). Risk tolerance and the level of risk that\nis acceptable to organizations or society are highly contextual and application and use-case\nspecific. Risk tolerances can be influenced by policies and norms established by AI sys-\ntem owners, organizations, industries, communities, or policy makers. Risk tolerances are\nlikely to change over time as AI systems, policies, and norms evolve. Different organiza-\ntions may have varied risk tolerances due to their particular organizational priorities and\nresource considerations.\nEmerging knowledge and methods to better inform harm/cost-benefit tradeoffs will con-\ntinue to be developed and debated by businesses, governments, academia, and civil society.\nTo the extent that challenges for specifying AI risk tolerances remain unresolved, there may\nbe contexts where a risk management framework is not yet readily applicable for mitigating\nnegative AI risks.\nThe Framework is intended to be flexible and to augment existing risk practices\nwhich should align with applicable laws, regulations, and norms. Organizations\nshould follow existing regulations and guidelines for risk criteria, tolerance, and\nresponse established by organizational, domain, discipline, sector, or professional\nrequirements. Some sectors or industries may have established definitions of harm or\nestablished documentation, reporting, and disclosure requirements. Within sectors,\nrisk management may depend on existing guidelines for specific applications and\nuse case settings. Where established guidelines do not exist, organizations should\ndefine reasonable risk tolerance. Once tolerance is defined, this AI RMF can be used\nto manage risks and to document risk management processes.\n1.2.3 Risk Prioritization\nAttempting to eliminate negative risk entirely can be counterproductive in practice because\nnot all incidents and failures can be eliminated. Unrealistic expectations about risk may\nlead organizations to allocate resources in a manner that makes risk triage inefficient or\nimpractical or wastes scarce resources. A risk management culture can help organizations\nrecognize that not all AI risks are the same, and resources can be allocated purposefully.\nActionable risk management efforts lay out clear guidelines for assessing trustworthiness\nof each AI system an organization develops or deploys. Policies and resources should be\nprioritized based on the assessed risk level and potential impact of an AI system. The extent\nto which an AI system may be customized or tailored to the specific context of use by the\nAI deployer can be a contributing factor.\nPage 7 NIST AI 100-1 AI RMF 1.0\nWhen applying the AI RMF, risks which the organization determines to be highest for the\nAI systems within a given context of use call for the most urgent prioritization and most\nthorough risk management process. In cases where an AI system presents unacceptable\nnegative risk levels \u2013 such as where significant negative impacts are imminent, severe harms\nare actually occurring, or catastrophic risks are present \u2013 development and deployment\nshould cease in a safe manner until risks can be sufficiently managed. If an AI system\u2019s\ndevelopment, deployment, and use cases are found to be low-risk in a specific context, that\nmay suggest potentially lower prioritization.\nRisk prioritization may differ between AI systems that are designed or deployed to directly\ninteract with humans as compared to AI systems that are not. Higher initial prioritization\nmay be called for in settings where the AI system is trained on large datasets comprised of\nsensitive or protected data such as personally identifiable information, or where the outputs\nof the AI systems have direct or indirect impact on humans. AI systems designed to interact\nonly with computational systems and trained on non-sensitive datasets (for example, data\ncollected from the physical environment) may call for lower initial prioritization. Nonethe-\nless, regularly assessing and prioritizing risk based on context remains important because\nnon-human-facing AI systems can have downstream safety or social implications.\nResidual risk \u2013 defined as risk remaining after risk treatment (Source: ISO GUIDE 73) \u2013\ndirectly impacts end users or affected individuals and communities. Documenting residual\nrisks will call for the system provider to fully consider the risks of deploying the AI product\nand will inform end users about potential negative impacts of interacting with the system.\n1.2.4 Organizational Integration and Management of Risk\nAI risks should not be considered in isolation. Different AI actors have different responsi-\nbilities and awareness depending on their roles in the lifecycle. For example, organizations\ndeveloping an AI system often will not have information about how the system may be\nused. AI risk management should be integrated and incorporated into broader enterprise\nrisk management strategies and processes. Treating AI risks along with other critical risks,\nsuch as cybersecurity and privacy, will yield a more integrated outcome and organizational\nefficiencies.\nThe AI RMF may be utilized along with related guidance and frameworks for managing\nAI system risks or broader enterprise risks. Some risks related to AI systems are common\nacross other types of software development and deployment. Examples of overlapping risks\ninclude: privacy concerns related to the use of underlying data to train AI systems; the en-\nergy and environmental implications associated with resource-heavy computing demands;\nsecurity concerns related to the confidentiality, integrity, and availability of the system and\nits training and output data; and general security of the underlying software and hardware\nfor AI systems.\nPage 8 NIST AI 100-1 AI RMF 1.0\nOrganizations need to establish and maintain the appropriate accountability mechanisms,\nroles and responsibilities, culture, and incentive structures for risk management to be ef-\nfective. Use of the AI RMF alone will not lead to these changes or provide the appropriate\nincentives. Effective risk management is realized through organizational commitment at\nsenior levels and may require cultural change within an organization or industry. In addi-\ntion, small to medium-sized organizations managing AI risks or implementing the AI RMF\nmay face different challenges than large organizations, depending on their capabilities and\nresources.\n2. Audience\nIdentifying and managing AI risks and potential impacts \u2013 both positive and negative \u2013 re-\nquires a broad set of perspectives and actors across the AI lifecycle. Ideally, AI actors will\nrepresent a diversity of experience, expertise, and backgrounds and comprise demograph-\nically and disciplinarily diverse teams. The AI RMF is intended to be used by AI actors\nacross the AI lifecycle and dimensions.\nThe OECD has developed a framework for classifying AI lifecycle activities according to\nfive key socio-technical dimensions, each with properties relevant for AI policy and gover-\nnance, including risk management [OECD (2022) OECD Framework for the Classification\nof AI systems \u2014 OECD Digital Economy Papers]. Figure 2 shows these dimensions,\nslightly modified by NIST for purposes of this framework. The NIST modification high-\nlights the importance of test, evaluation, verification, and validation (TEVV) processes\nthroughout an AI lifecycle and generalizes the operational context of an AI system.\nAI dimensions displayed in Figure 2 are the Application Context, Data and Input, AI\nModel, and Task and Output. AI actors involved in these dimensions who perform or\nmanage the design, development, deployment, evaluation, and use of AI systems and drive\nAI risk management efforts are the primary AI RMF audience.\nRepresentative AI actors across the lifecycle dimensions are listed in Figure 3 and described\nin detail in Appendix A. Within the AI RMF, all AI actors work together to manage risks\nand achieve the goals of trustworthy and responsible AI. AI actors with TEVV-specific\nexpertise are integrated throughout the AI lifecycle and are especially likely to benefit from\nthe Framework. Performed regularly, TEVV tasks can provide insights relative to technical,\nsocietal, legal, and ethical standards or norms, and can assist with anticipating impacts and\nassessing and tracking emergent risks. As a regular process within an AI lifecycle, TEVV\nallows for both mid-course remediation and post-hoc risk management.\nThe People & Planet dimension at the center of Figure 2 represents human rights and the\nbroader well-being of society and the planet. The AI actors in this dimension comprise\na separate AI RMF audience who informs the primary audience. These AI actors may in-\nclude trade associations, standards developing organizations, researchers, advocacy groups,\nPage 9 NIST AI 100-1 AI RMF 1.0\nFig. 2. Lifecycle and Key Dimensions of an AI System. Modified from OECD (2022) OECD\nFramework for the Classification of AI systems \u2014 OECD Digital Economy Papers. The two inner\ncircles show AI systems\u2019 key dimensions and the outer circle shows AI lifecycle stages. Ideally,\nrisk management efforts start with the Plan and Design function in the application context and are\nperformed throughout the AI system lifecycle. See Figure 3 for representative AI actors.\nenvironmental groups, civil society organizations, end users, and potentially impacted in-\ndividuals and communities. These actors can:\n\u2022 assist in providing context and understanding potential and actual impacts;\n\u2022 be a source of formal or quasi-formal norms and guidance for AI risk management;\n\u2022 designate boundaries for AI operation (technical, societal, legal, and ethical); and\n\u2022 promote discussion of the tradeoffs needed to balance societal values and priorities\nrelated to civil liberties and rights, equity, the environment and the planet, and the\neconomy.\nSuccessful risk management depends upon a sense of collective responsibility among AI\nactors shown in Figure 3. The AI RMF functions, described in Section 5, require diverse\nperspectives, disciplines, professions, and experiences. Diverse teams contribute to more\nopen sharing of ideas and assumptions about the purposes and functions of technology \u2013\nmaking these implicit aspects more explicit. This broader collective perspective creates\nopportunities for surfacing problems and identifying existing and emergent risks.\nPage 10 NIST AI 100-1 AI RMF 1.0\nFig. 3. AI actors across AI lifecycle stages. See Appendix A for detailed descriptions of AI actor tasks, including details about testing,\nevaluation, verification, and validation tasks. Note that AI actors in the AI Model dimension (Figure 2) are separated as a best practice, with\nthose building and using the models separated from those verifying and validating the models.\nPage 11 NIST AI 100-1 AI RMF 1.0\n3. AI Risks and Trustworthiness\nFor AI systems to be trustworthy, they often need to be responsive to a multiplicity of cri-\nteria that are of value to interested parties. Approaches which enhance AI trustworthiness\ncan reduce negative AI risks. This Framework articulates the following characteristics of\ntrustworthy AI and offers guidance for addressing them. Characteristics of trustworthy AI\nsystems include: valid and reliable, safe, secure and resilient, accountable and trans-\nparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias\nmanaged. Creating trustworthy AI requires balancing each of these characteristics based\non the AI system\u2019s context of use. While all characteristics are socio-technical system at-\ntributes, accountability and transparency also relate to the processes and activities internal\nto an AI system and its external setting. Neglecting these characteristics can increase the\nprobability and magnitude of negative consequences.\nFig. 4. Characteristics of trustworthy AI systems. Valid & Reliable is a necessary condition of\ntrustworthiness and is shown as the base for other trustworthiness characteristics. Accountable &\nTransparent is shown as a vertical box because it relates to all other characteristics.\nTrustworthiness characteristics (shown in Figure 4) are inextricably tied to social and orga-\nnizational behavior, the datasets used by AI systems, selection of AI models and algorithms\nand the decisions made by those who build them, and the interactions with the humans who\nprovide insight from and oversight of such systems. Human judgment should be employed\nwhen deciding on the specific metrics related to AI trustworthiness characteristics and the\nprecise threshold values for those metrics.\nAddressing AI trustworthiness characteristics individually will not ensure AI system trust-\nworthiness; tradeoffs are usually involved, rarely do all characteristics apply in every set-\nting, and some will be more or less important in any given situation. Ultimately, trustwor-\nthiness is a social concept that ranges across a spectrum and is only as strong as its weakest\ncharacteristics.\nWhen managing AI risks, organizations can face difficult decisions in balancing these char-\nacteristics. For example, in certain scenarios tradeoffs may emerge between optimizing for\ninterpretability and achieving privacy. In other cases, organizations might face a tradeoff\nbetween predictive accuracy and interpretability. Or, under certain conditions such as data\nsparsity, privacy-enhancing techniques can result in a loss in accuracy, affecting decisions\nPage 12 NIST AI 100-1 AI RMF 1.0\nabout fairness and other values in certain domains. Dealing with tradeoffs requires tak-\ning into account the decision-making context. These analyses can highlight the existence\nand extent of tradeoffs between different measures, but they do not answer questions about\nhow to navigate the tradeoff. Those depend on the values at play in the relevant context and\nshould be resolved in a manner that is both transparent and appropriately justifiable.\nThere are multiple approaches for enhancing contextual awareness in the AI lifecycle. For\nexample, subject matter experts can assist in the evaluation of TEVV findings and work\nwith product and deployment teams to align TEVV parameters to requirements and de-\nployment conditions. When properly resourced, increasing the breadth and diversity of\ninput from interested parties and relevant AI actors throughout the AI lifecycle can en-\nhance opportunities for informing contextually sensitive evaluations, and for identifying\nAI system benefits and positive impacts. These practices can increase the likelihood that\nrisks arising in social contexts are managed appropriately.\nUnderstanding and treatment of trustworthiness characteristics depends on an AI actor\u2019s\nparticular role within the AI lifecycle. For any given AI system, an AI designer or developer\nmay have a different perception of the characteristics than the deployer.\nTrustworthiness characteristics explained in this document influence each other.\nHighly secure but unfair systems, accurate but opaque and uninterpretable systems,\nand inaccurate but secure, privacy-enhanced, and transparent systems are all unde-\nsirable. A comprehensive approach to risk management calls for balancing tradeoffs\namong the trustworthiness characteristics. It is the joint responsibility of all AI ac-\ntors to determine whether AI technology is an appropriate or necessary tool for a\ngiven context or purpose, and how to use it responsibly. The decision to commission\nor deploy an AI system should be based on a contextual assessment of trustworthi-\nness characteristics and the relative risks, impacts, costs, and benefits, and informed\nby a broad set of interested parties.\n3.1 Valid and Reliable\nValidation is the \u201cconfirmation, through the provision of objective evidence, that the re-\nquirements for a specific intended use or application have been fulfilled\u201d (Source: ISO\n9000:2015). Deployment of AI systems which are inaccurate, unreliable, or poorly gener-\nalized to data and settings beyond their training creates and increases negative AI risks and\nreduces trustworthiness.\nReliability is defined in the same standard as the \u201cability of an item to perform as required,\nwithout failure, for a given time interval, under given conditions\u201d (Source: ISO/IEC TS\n5723:2022). Reliability is a goal for overall correctness of AI system operation under the\nconditions of expected use and over a given period of time, including the entire lifetime of\nthe system.\nPage 13 NIST AI 100-1 AI RMF 1.0\nAccuracy and robustness contribute to the validity and trustworthiness of AI systems, and\ncan be in tension with one another in AI systems.\nAccuracy is defined by ISO/IEC TS 5723:2022 as \u201ccloseness of results of observations,\ncomputations, or estimates to the true values or the values accepted as being true.\u201d Mea-\nsures of accuracy should consider computational-centric measures (e.g., false positive and\nfalse negative rates), human-AI teaming, and demonstrate external validity (generalizable\nbeyond the training conditions). Accuracy measurements should always be paired with\nclearly defined and realistic test sets \u2013 that are representative of conditions of expected use\n\u2013 and details about test methodology; these should be included in associated documen-\ntation. Accuracy measurements may include disaggregation of results for different data\nsegments.\nRobustness orgeneralizability is defined as the \u201cability of a system to maintain its level\nof performance under a variety of circumstances\u201d (Source: ISO/IEC TS 5723:2022). Ro-\nbustness is a goal for appropriate system functionality in a broad set of conditions and\ncircumstances, including uses of AI systems not initially anticipated. Robustness requires\nnot only that the system perform exactly as it does under expected uses, but also that it\nshould perform in ways that minimize potential harms to people if it is operating in an\nunexpected setting.\nValidity and reliability for deployed AI systems are often assessed by ongoing testing or\nmonitoring that confirms a system is performing as intended. Measurement of validity,\naccuracy, robustness, and reliability contribute to trustworthiness and should take into con-\nsideration that certain types of failures can cause greater harm. AI risk management efforts\nshould prioritize the minimization of potential negative impacts, and may need to include\nhuman intervention in cases where the AI system cannot detect or correct errors.\n3.2 Safe\nAI systems should \u201cnot under defined conditions, lead to a state in which human life,\nhealth, property, or the environment is endangered\u201d (Source: ISO/IEC TS 5723:2022). Safe\noperation of AI systems is improved through:\n\u2022 responsible design, development, and deployment practices;\n\u2022 clear information to deployers on responsible use of the system;\n\u2022 responsible decision-making by deployers and end users; and\n\u2022 explanations and documentation of risks based on empirical evidence of incidents.\nDifferent types of safety risks may require tailored AI risk management approaches based\non context and the severity of potential risks presented. Safety risks that pose a potential\nrisk of serious injury or death call for the most urgent prioritization and most thorough risk\nmanagement process.\nPage 14 NIST AI 100-1 AI RMF 1.0\nEmploying safety considerations during the lifecycle and starting as early as possible with\nplanning and design can prevent failures or conditions that can render a system dangerous.\nOther practical approaches for AI safety often relate to rigorous simulation and in-domain\ntesting, real-time monitoring, and the ability to shut down, modify, or have human inter-\nvention into systems that deviate from intended or expected functionality.\nAI safety risk management approaches should take cues from efforts and guidelines for\nsafety in fields such as transportation and healthcare, and align with existing sector- or\napplication-specific guidelines or standards.\n3.3 Secure and Resilient\nAI systems, as well as the ecosystems in which they are deployed, may be said to be re-\nsilient if they can withstand unexpected adverse events or unexpected changes in their envi-\nronment or use \u2013 or if they can maintain their functions and structure in the face of internal\nand external change and degrade safely and gracefully when this is necessary (Adapted\nfrom: ISO/IEC TS 5723:2022). Common security concerns relate to adversarial examples,\ndata poisoning, and the exfiltration of models, training data, or other intellectual property\nthrough AI system endpoints. AI systems that can maintain confidentiality, integrity, and\navailability through protection mechanisms that prevent unauthorized access and use may\nbe said to be secure . Guidelines in the NIST Cybersecurity Framework and Risk Manage-\nment Framework are among those which are applicable here.\nSecurity and resilience are related but distinct characteristics. While resilience is the abil-\nity to return to normal function after an unexpected adverse event, security includes re-\nsilience but also encompasses protocols to avoid, protect against, respond to, or recover\nfrom attacks. Resilience relates to robustness and goes beyond the provenance of the data\nto encompass unexpected or adversarial use (or abuse or misuse) of the model or data.\n3.4 Accountable and Transparent\nTrustworthy AI depends upon accountability. Accountability presupposes transparency.\nTransparency reflects the extent to which information about an AI system and its outputs is\navailable to individuals interacting with such a system \u2013 regardless of whether they are even\naware that they are doing so. Meaningful transparency provides access to appropriate levels\nof information based on the stage of the AI lifecycle and tailored to the role or knowledge\nof AI actors or individuals interacting with or using the AI system. By promoting higher\nlevels of understanding, transparency increases confidence in the AI system.\nThis characteristic\u2019s scope spans from design decisions and training data to model train-\ning, the structure of the model, its intended use cases, and how and when deployment,\npost-deployment, or end user decisions were made and by whom. Transparency is often\nnecessary for actionable redress related to AI system outputs that are incorrect or otherwise\nlead to negative impacts. Transparency should consider human-AI interaction: for exam-\nPage 15 NIST AI 100-1 AI RMF 1.0\nple, how a human operator or user is notified when a potential or actual adverse outcome\ncaused by an AI system is detected. A transparent system is not necessarily an accurate,\nprivacy-enhanced, secure, or fair system. However, it is difficult to determine whether an\nopaque system possesses such characteristics, and to do so over time as complex systems\nevolve.\nThe role of AI actors should be considered when seeking accountability for the outcomes of\nAI systems. The relationship between risk and accountability associated with AI and tech-\nnological systems more broadly differs across cultural, legal, sectoral, and societal contexts.\nWhen consequences are severe, such as when life and liberty are at stake, AI developers\nand deployers should consider proportionally and proactively adjusting their transparency\nand accountability practices. Maintaining organizational practices and governing structures\nfor harm reduction, like risk management, can help lead to more accountable systems.\nMeasures to enhance transparency and accountability should also consider the impact of\nthese efforts on the implementing entity, including the level of necessary resources and the\nneed to safeguard proprietary information.\nMaintaining the provenance of training data and supporting attribution of the AI system\u2019s\ndecisions to subsets of training data can assist with both transparency and accountability.\nTraining data may also be subject to copyright and should follow applicable intellectual\nproperty rights laws.\nAs transparency tools for AI systems and related documentation continue to evolve, devel-\nopers of AI systems are encouraged to test different types of transparency tools in cooper-\nation with AI deployers to ensure that AI systems are used as intended.\n3.5 Explainable and Interpretable\nExplainability refers to a representation of the mechanisms underlying AI systems\u2019 oper-\nation, whereas interpretability refers to the meaning of AI systems\u2019 output in the context\nof their designed functional purposes. Together, explainability and interpretability assist\nthose operating or overseeing an AI system, as well as users of an AI system, to gain\ndeeper insights into the functionality and trustworthiness of the system, including its out-\nputs. The underlying assumption is that perceptions of negative risk stem from a lack of\nability to make sense of, or contextualize, system output appropriately. Explainable and\ninterpretable AI systems offer information that will help end users understand the purposes\nand potential impact of an AI system.\nRisk from lack of explainability may be managed by describing how AI systems function,\nwith descriptions tailored to individual differences such as the user\u2019s role, knowledge, and\nskill level. Explainable systems can be debugged and monitored more easily, and they lend\nthemselves to more thorough documentation, audit, and governance.\nPage 16 NIST AI 100-1 AI RMF 1.0\nRisks to interpretability often can be addressed by communicating a description of why\nan AI system made a particular prediction or recommendation. (See \u201cFour Principles of\nExplainable Artificial Intelligence\u201d and \u201cPsychological Foundations of Explainability and\nInterpretability in Artificial Intelligence\u201d found here.)\nTransparency, explainability, and interpretability are distinct characteristics that support\neach other. Transparency can answer the question of \u201cwhat happened\u201d in the system. Ex-\nplainability can answer the question of \u201chow\u201d a decision was made in the system. Inter-\npretability can answer the question of \u201cwhy\u201d a decision was made by the system and its\nmeaning or context to the user.\n3.6 Privacy-Enhanced\nPrivacy refers generally to the norms and practices that help to safeguard human autonomy,\nidentity, and dignity. These norms and practices typically address freedom from intrusion,\nlimiting observation, or individuals\u2019 agency to consent to disclosure or control of facets of\ntheir identities (e.g., body, data, reputation). (See The NIST Privacy Framework: A Tool\nfor Improving Privacy through Enterprise Risk Management.)\nPrivacy values such as anonymity, confidentiality, and control generally should guide choices\nfor AI system design, development, and deployment. Privacy-related risks may influence\nsecurity, bias, and transparency and come with tradeoffs with these other characteristics.\nLike safety and security, specific technical features of an AI system may promote or reduce\nprivacy. AI systems can also present new risks to privacy by allowing inference to identify\nindividuals or previously private information about individuals.\nPrivacy-enhancing technologies (\u201cPETs\u201d) for AI, as well as data minimizing methods such\nas de-identification and aggregation for certain model outputs, can support design for\nprivacy-enhanced AI systems. Under certain conditions such as data sparsity, privacy-\nenhancing techniques can result in a loss in accuracy, affecting decisions about fairness\nand other values in certain domains.\n3.7 Fair \u2013 with Harmful Bias Managed\nFairness in AI includes concerns for equality and equity by addressing issues such as harm-\nful bias and discrimination. Standards of fairness can be complex and difficult to define be-\ncause perceptions of fairness differ among cultures and may shift depending on application.\nOrganizations\u2019 risk management efforts will be enhanced by recognizing and considering\nthese differences. Systems in which harmful biases are mitigated are not necessarily fair.\nFor example, systems in which predictions are somewhat balanced across demographic\ngroups may still be inaccessible to individuals with disabilities or affected by the digital\ndivide or may exacerbate existing disparities or systemic biases.\nPage 17 NIST AI 100-1 AI RMF 1.0\nBias is broader than demographic balance and data representativeness. NIST has identified\nthree major categories of AI bias to be considered and managed: systemic, computational\nand statistical, and human-cognitive. Each of these can occur in the absence of prejudice,\npartiality, or discriminatory intent. Systemic bias can be present in AI datasets, the orga-\nnizational norms, practices, and processes across the AI lifecycle, and the broader society\nthat uses AI systems. Computational and statistical biases can be present in AI datasets\nand algorithmic processes, and often stem from systematic errors due to non-representative\nsamples. Human-cognitive biases relate to how an individual or group perceives AI sys-\ntem information to make a decision or fill in missing information, or how humans think\nabout purposes and functions of an AI system. Human-cognitive biases are omnipresent\nin decision-making processes across the AI lifecycle and system use, including the design,\nimplementation, operation, and maintenance of AI.\nBias exists in many forms and can become ingrained in the automated systems that help\nmake decisions about our lives. While bias is not always a negative phenomenon, AI sys-\ntems can potentially increase the speed and scale of biases and perpetuate and amplify\nharms to individuals, groups, communities, organizations, and society. Bias is tightly asso-\nciated with the concepts of transparency as well as fairness in society. (For more informa-\ntion about bias, including the three categories, see NIST Special Publication 1270, Towards\na Standard for Identifying and Managing Bias in Artificial Intelligence.)\nPage 18 NIST AI 100-1 AI RMF 1.0\n4. Effectiveness of the AI RMF\nEvaluations of AI RMF effectiveness \u2013 including ways to measure bottom-line improve-\nments in the trustworthiness of AI systems \u2013 will be part of future NIST activities, in\nconjunction with the AI community.\nOrganizations and other users of the Framework are encouraged to periodically evaluate\nwhether the AI RMF has improved their ability to manage AI risks, including but not lim-\nited to their policies, processes, practices, implementation plans, indicators, measurements,\nand expected outcomes. NIST intends to work collaboratively with others to develop met-\nrics, methodologies, and goals for evaluating the AI RMF\u2019s effectiveness, and to broadly\nshare results and supporting information. Framework users are expected to benefit from:\n\u2022 enhanced processes for governing, mapping, measuring, and managing AI risk, and\nclearly documenting outcomes;\n\u2022 improved awareness of the relationships and tradeoffs among trustworthiness char-\nacteristics, socio-technical approaches, and AI risks;\n\u2022 explicit processes for making go/no-go system commissioning and deployment deci-\nsions;\n\u2022 established policies, processes, practices, and procedures for improving organiza-\ntional accountability efforts related to AI system risks;\n\u2022 enhanced organizational culture which prioritizes the identification and management\nof AI system risks and potential impacts to individuals, communities, organizations,\nand society;\n\u2022 better information sharing within and across organizations about risks, decision-\nmaking processes, responsibilities, common pitfalls, TEVV practices, and approaches\nfor continuous improvement;\n\u2022 greater contextual knowledge for increased awareness of downstream risks;\n\u2022 strengthened engagement with interested parties and relevant AI actors; and\n\u2022 augmented capacity for TEVV of AI systems and associated risks.\nPage 19 NIST AI 100-1 AI RMF 1.0\nPart 2: Core and Profiles\n5. AI RMF Core\nThe AI RMF Core provides outcomes and actions that enable dialogue, understanding, and\nactivities to manage AI risks and responsibly develop trustworthy AI systems. As illus-\ntrated in Figure 5, the Core is composed of four functions: GOVERN ,MAP ,MEASURE ,\nand MANAGE . Each of these high-level functions is broken down into categories and sub-\ncategories. Categories and subcategories are subdivided into specific actions and outcomes.\nActions do not constitute a checklist, nor are they necessarily an ordered set of steps.\nFig. 5. Functions organize AI risk management activities at their highest level to govern, map,\nmeasure, and manage AI risks. Governance is designed to be a cross-cutting function to inform\nand be infused throughout the other three functions.\nRisk management should be continuous, timely, and performed throughout the AI system\nlifecycle dimensions. AI RMF Core functions should be carried out in a way that reflects\ndiverse and multidisciplinary perspectives, potentially including the views of AI actors out-\nside the organization. Having a diverse team contributes to more open sharing of ideas and\nassumptions about purposes and functions of the technology being designed, developed,\nPage 20 NIST AI 100-1 AI RMF 1.0\ndeployed, or evaluated \u2013 which can create opportunities to surface problems and identify\nexisting and emergent risks.\nAn online companion resource to the AI RMF, the NIST AI RMF Playbook, is available\nto help organizations navigate the AI RMF and achieve its outcomes through suggested\ntactical actions they can apply within their own contexts. Like the AI RMF, the Playbook\nis voluntary and organizations can utilize the suggestions according to their needs and\ninterests. Playbook users can create tailored guidance selected from suggested material\nfor their own use and contribute their suggestions for sharing with the broader community.\nAlong with the AI RMF, the Playbook is part of the NIST Trustworthy and Responsible AI\nResource Center.\nFramework users may apply these functions as best suits their needs for managing\nAI risks based on their resources and capabilities. Some organizations may choose\nto select from among the categories and subcategories; others may choose and have\nthe capacity to apply all categories and subcategories. Assuming a governance struc-\nture is in place, functions may be performed in any order across the AI lifecycle as\ndeemed to add value by a user of the framework. After instituting the outcomes in\nGOVERN , most users of the AI RMF would start with the MAP function and con-\ntinue to MEASURE orMANAGE . However users integrate the functions, the process\nshould be iterative, with cross-referencing between functions as necessary. Simi-\nlarly, there are categories and subcategories with elements that apply to multiple\nfunctions, or that logically should take place before certain subcategory decisions.\n5.1 Govern\nThe GOVERN function:\n\u2022 cultivates and implements a culture of risk management within organizations design-\ning, developing, deploying, evaluating, or acquiring AI systems;\n\u2022 outlines processes, documents, and organizational schemes that anticipate, identify,\nand manage the risks a system can pose, including to users and others across society\n\u2013 and procedures to achieve those outcomes;\n\u2022 incorporates processes to assess potential impacts;\n\u2022 provides a structure by which AI risk management functions can align with organi-\nzational principles, policies, and strategic priorities;\n\u2022 connects technical aspects of AI system design and development to organizational\nvalues and principles, and enables organizational practices and competencies for the\nindividuals involved in acquiring, training, deploying, and monitoring such systems;\nand\n\u2022 addresses full product lifecycle and associated processes, including legal and other\nissues concerning use of third-party software or hardware systems and data.\nPage 21 NIST AI 100-1 AI RMF 1.0\nGOVERN is a cross-cutting function that is infused throughout AI risk management and\nenables the other functions of the process. Aspects of GOVERN , especially those related to\ncompliance or evaluation, should be integrated into each of the other functions. Attention\nto governance is a continual and intrinsic requirement for effective AI risk management\nover an AI system\u2019s lifespan and the organization\u2019s hierarchy.\nStrong governance can drive and enhance internal practices and norms to facilitate orga-\nnizational risk culture. Governing authorities can determine the overarching policies that\ndirect an organization\u2019s mission, goals, values, culture, and risk tolerance. Senior leader-\nship sets the tone for risk management within an organization, and with it, organizational\nculture. Management aligns the technical aspects of AI risk management to policies and\noperations. Documentation can enhance transparency, improve human review processes,\nand bolster accountability in AI system teams.\nAfter putting in place the structures, systems, processes, and teams described in the GOV-\nERN function, organizations should benefit from a purpose-driven culture focused on risk\nunderstanding and management. It is incumbent on Framework users to continue to ex-\necute the GOVERN function as knowledge, cultures, and needs or expectations from AI\nactors evolve over time.\nPractices related to governing AI risks are described in the NIST AI RMF Playbook. Table\n1 lists the GOVERN function\u2019s categories and subcategories.\nTable 1: Categories and subcategories for the GOVERN function.\nGOVERN 1:\nPolicies, processes,\nprocedures, and\npractices across the\norganization related\nto the mapping,\nmeasuring, and\nmanaging of AI\nrisks are in place,\ntransparent, and\nimplemented\neffectively.GOVERN 1.1: Legal and regulatory requirements involving AI\nare understood, managed, and documented.\nGOVERN 1.2: The characteristics of trustworthy AI are inte-\ngrated into organizational policies, processes, procedures, and\npractices.\nGOVERN 1.3: Processes, procedures, and practices are in place\nto determine the needed level of risk management activities based\non the organization\u2019s risk tolerance.\nGOVERN 1.4:The risk management process and its outcomes are\nestablished through transparent policies, procedures, and other\ncontrols based on organizational risk priorities.Categories Subcategories\nContinued on next page\nPage 22 NIST AI 100-1 AI RMF 1.0\nTable 1: Categories and subcategories for the GOVERN function. (Continued)\nGOVERN 1.5: Ongoing monitoring and periodic review of the\nrisk management process and its outcomes are planned and or-\nganizational roles and responsibilities clearly defined, including\ndetermining the frequency of periodic review.\nGOVERN 1.6: Mechanisms are in place to inventory AI systems\nand are resourced according to organizational risk priorities.\nGOVERN 1.7: Processes and procedures are in place for decom-\nmissioning and phasing out AI systems safely and in a man-\nner that does not increase risks or decrease the organization\u2019s\ntrustworthiness.\nGOVERN 2:\nAccountability\nstructures are in\nplace so that the\nappropriate teams\nand individuals are\nempowered,\nresponsible, and\ntrained for mapping,\nmeasuring, and\nmanaging AI risks.GOVERN 2.1: Roles and responsibilities and lines of communi-\ncation related to mapping, measuring, and managing AI risks are\ndocumented and are clear to individuals and teams throughout\nthe organization.\nGOVERN 2.2: The organization\u2019s personnel and partners receive\nAI risk management training to enable them to perform their du-\nties and responsibilities consistent with related policies, proce-\ndures, and agreements.\nGOVERN 2.3: Executive leadership of the organization takes re-\nsponsibility for decisions about risks associated with AI system\ndevelopment and deployment.\nGOVERN 3:\nWorkforce diversity,\nequity, inclusion,\nand accessibility\nprocesses are\nprioritized in the\nmapping,\nmeasuring, and\nmanaging of AI\nrisks throughout the\nlifecycle.GOVERN 3.1: Decision-making related to mapping, measuring,\nand managing AI risks throughout the lifecycle is informed by a\ndiverse team (e.g., diversity of demographics, disciplines, expe-\nrience, expertise, and backgrounds).\nGOVERN 3.2: Policies and procedures are in place to define and\ndifferentiate roles and responsibilities for human-AI configura-\ntions and oversight of AI systems.\nGOVERN 4:\nOrganizational\nteams are committed\nto a cultureGOVERN 4.1: Organizational policies and practices are in place\nto foster a critical thinking and safety-first mindset in the design,\ndevelopment, deployment, and uses of AI systems to minimize\npotential negative impacts.Categories Subcategories\nContinued on next page\nPage 23 NIST AI 100-1 AI RMF 1.0\nTable 1: Categories and subcategories for the GOVERN function. (Continued)\nthat considers and\ncommunicates AI\nrisk.GOVERN 4.2: Organizational teams document the risks and po-\ntential impacts of the AI technology they design, develop, deploy,\nevaluate, and use, and they communicate about the impacts more\nbroadly.\nGOVERN 4.3: Organizational practices are in place to enable AI\ntesting, identification of incidents, and information sharing.\nGOVERN 5:\nProcesses are in\nplace for robust\nengagement with\nrelevant AI actors.GOVERN 5.1: Organizational policies and practices are in place\nto collect, consider, prioritize, and integrate feedback from those\nexternal to the team that developed or deployed the AI system\nregarding the potential individual and societal impacts related to\nAI risks.\nGOVERN 5.2: Mechanisms are established to enable the team\nthat developed or deployed AI systems to regularly incorporate\nadjudicated feedback from relevant AI actors into system design\nand implementation.\nGOVERN 6:Policies\nand procedures are\nin place to address\nAI risks and benefits\narising from\nthird-party software\nand data and other\nsupply chain issues.GOVERN 6.1: Policies and procedures are in place that address\nAI risks associated with third-party entities, including risks of in-\nfringement of a third-party\u2019s intellectual property or other rights.\nGOVERN 6.2: Contingency processes are in place to handle\nfailures or incidents in third-party data or AI systems deemed to\nbe high-risk.Categories Subcategories\n5.2 Map\nThe MAP function establishes the context to frame risks related to an AI system. The AI\nlifecycle consists of many interdependent activities involving a diverse set of actors (See\nFigure 3). In practice, AI actors in charge of one part of the process often do not have full\nvisibility or control over other parts and their associated contexts. The interdependencies\nbetween these activities, and among the relevant AI actors, can make it difficult to reliably\nanticipate impacts of AI systems. For example, early decisions in identifying purposes and\nobjectives of an AI system can alter its behavior and capabilities, and the dynamics of de-\nployment setting (such as end users or impacted individuals) can shape the impacts of AI\nsystem decisions. As a result, the best intentions within one dimension of the AI lifecycle\ncan be undermined via interactions with decisions and conditions in other, later activities.\nPage 24 NIST AI 100-1 AI RMF 1.0\nThis complexity and varying levels of visibility can introduce uncertainty into risk man-\nagement practices. Anticipating, assessing, and otherwise addressing potential sources of\nnegative risk can mitigate this uncertainty and enhance the integrity of the decision process.\nThe information gathered while carrying out the MAP function enables negative risk pre-\nvention and informs decisions for processes such as model management, as well as an\ninitial decision about appropriateness or the need for an AI solution. Outcomes in the\nMAP function are the basis for the MEASURE and MANAGE functions. Without contex-\ntual knowledge, and awareness of risks within the identified contexts, risk management is\ndifficult to perform. The MAP function is intended to enhance an organization\u2019s ability to\nidentify risks and broader contributing factors.\nImplementation of this function is enhanced by incorporating perspectives from a diverse\ninternal team and engagement with those external to the team that developed or deployed\nthe AI system. Engagement with external collaborators, end users, potentially impacted\ncommunities, and others may vary based on the risk level of a particular AI system, the\nmakeup of the internal team, and organizational policies. Gathering such broad perspec-\ntives can help organizations proactively prevent negative risks and develop more trustwor-\nthy AI systems by:\n\u2022 improving their capacity for understanding contexts;\n\u2022 checking their assumptions about context of use;\n\u2022 enabling recognition of when systems are not functional within or out of their in-\ntended context;\n\u2022 identifying positive and beneficial uses of their existing AI systems;\n\u2022 improving understanding of limitations in AI and ML processes;\n\u2022 identifying constraints in real-world applications that may lead to negative impacts;\n\u2022 identifying known and foreseeable negative impacts related to intended use of AI\nsystems; and\n\u2022 anticipating risks of the use of AI systems beyond intended use.\nAfter completing the MAP function, Framework users should have sufficient contextual\nknowledge about AI system impacts to inform an initial go/no-go decision about whether\nto design, develop, or deploy an AI system. If a decision is made to proceed, organizations\nshould utilize the MEASURE and MANAGE functions along with policies and procedures\nput into place in the GOVERN function to assist in AI risk management efforts. It is incum-\nbent on Framework users to continue applying the MAP function to AI systems as context,\ncapabilities, risks, benefits, and potential impacts evolve over time.\nPractices related to mapping AI risks are described in the NIST AI RMF Playbook. Table\n2 lists the MAP function\u2019s categories and subcategories.\nPage 25 NIST AI 100-1 AI RMF 1.0\nTable 2: Categories and subcategories for the MAP function.\nMAP 1:Context is\nestablished and\nunderstood.MAP 1.1:Intended purposes, potentially beneficial uses, context-\nspecific laws, norms and expectations, and prospective settings in\nwhich the AI system will be deployed are understood and docu-\nmented. Considerations include: the specific set or types of users\nalong with their expectations; potential positive and negative im-\npacts of system uses to individuals, communities, organizations,\nsociety, and the planet; assumptions and related limitations about\nAI system purposes, uses, and risks across the development or\nproduct AI lifecycle; and related TEVV and system metrics.\nMAP 1.2: Interdisciplinary AI actors, competencies, skills, and\ncapacities for establishing context reflect demographic diversity\nand broad domain and user experience expertise, and their par-\nticipation is documented. Opportunities for interdisciplinary col-\nlaboration are prioritized.\nMAP 1.3: The organization\u2019s mission and relevant goals for AI\ntechnology are understood and documented.\nMAP 1.4: The business value or context of business use has been\nclearly defined or \u2013 in the case of assessing existing AI systems\n\u2013 re-evaluated.\nMAP 1.5: Organizational risk tolerances are determined and\ndocumented.\nMAP 1.6: System requirements (e.g., \u201cthe system shall respect\nthe privacy of its users\u201d) are elicited from and understood by rel-\nevant AI actors. Design decisions take socio-technical implica-\ntions into account to address AI risks.\nMAP 2:\nCategorization of\nthe AI system is\nperformed.MAP 2.1: The specific tasks and methods used to implement the\ntasks that the AI system will support are defined (e.g., classifiers,\ngenerative models, recommenders).\nMAP 2.2: Information about the AI system\u2019s knowledge limits\nand how system output may be utilized and overseen by humans\nis documented. Documentation provides sufficient information\nto assist relevant AI actors when making decisions and taking\nsubsequent actions.Categories Subcategories\nContinued on next page\nPage 26 NIST AI 100-1 AI RMF 1.0\nTable 2: Categories and subcategories for the MAP function. (Continued)\nMAP 2.3: Scientific integrity and TEVV considerations are iden-\ntified and documented, including those related to experimental\ndesign, data collection and selection (e.g., availability, repre-\nsentativeness, suitability), system trustworthiness, and construct\nvalidation.\nMAP 3:AI\ncapabilities, targeted\nusage, goals, and\nexpected benefits\nand costs compared\nwith appropriate\nbenchmarks are\nunderstood.MAP 3.1: Potential benefits of intended AI system functionality\nand performance are examined and documented.\nMAP 3.2: Potential costs, including non-monetary costs, which\nresult from expected or realized AI errors or system functionality\nand trustworthiness \u2013 as connected to organizational risk toler-\nance \u2013 are examined and documented.\nMAP 3.3: Targeted application scope is specified and docu-\nmented based on the system\u2019s capability, established context, and\nAI system categorization.\nMAP 3.4: Processes for operator and practitioner proficiency\nwith AI system performance and trustworthiness \u2013 and relevant\ntechnical standards and certifications \u2013 are defined, assessed, and\ndocumented.\nMAP 3.5: Processes for human oversight are defined, assessed,\nand documented in accordance with organizational policies from\ntheGOVERN function.\nMAP 4:Risks and\nbenefits are mapped\nfor all components\nof the AI system\nincluding third-party\nsoftware and data.MAP 4.1: Approaches for mapping AI technology and legal risks\nof its components \u2013 including the use of third-party data or soft-\nware \u2013 are in place, followed, and documented, as are risks of in-\nfringement of a third party\u2019s intellectual property or other rights.\nMAP 4.2: Internal risk controls for components of the AI sys-\ntem, including third-party AI technologies, are identified and\ndocumented.\nMAP 5:Impacts to\nindividuals, groups,\ncommunities,\norganizations, and\nsociety are\ncharacterized.MAP 5.1: Likelihood and magnitude of each identified impact\n(both potentially beneficial and harmful) based on expected use,\npast uses of AI systems in similar contexts, public incident re-\nports, feedback from those external to the team that developed\nor deployed the AI system, or other data are identified and\ndocumented.Categories Subcategories\nContinued on next page\nPage 27 NIST AI 100-1 AI RMF 1.0\nTable 2: Categories and subcategories for the MAP function. (Continued)\nMAP 5.2: Practices and personnel for supporting regular en-\ngagement with relevant AI actors and integrating feedback about\npositive, negative, and unanticipated impacts are in place and\ndocumented.Categories Subcategories\n5.3 Measure\nThe MEASURE function employs quantitative, qualitative, or mixed-method tools, tech-\nniques, and methodologies to analyze, assess, benchmark, and monitor AI risk and related\nimpacts. It uses knowledge relevant to AI risks identified in the MAP function and informs\nthe MANAGE function. AI systems should be tested before their deployment and regu-\nlarly while in operation. AI risk measurements include documenting aspects of systems\u2019\nfunctionality and trustworthiness.\nMeasuring AI risks includes tracking metrics for trustworthy characteristics, social impact,\nand human-AI configurations. Processes developed or adopted in the MEASURE function\nshould include rigorous software testing and performance assessment methodologies with\nassociated measures of uncertainty, comparisons to performance benchmarks, and formal-\nized reporting and documentation of results. Processes for independent review can improve\nthe effectiveness of testing and can mitigate internal biases and potential conflicts of inter-\nest.\nWhere tradeoffs among the trustworthy characteristics arise, measurement provides a trace-\nable basis to inform management decisions. Options may include recalibration, impact\nmitigation, or removal of the system from design, development, production, or use, as well\nas a range of compensating, detective, deterrent, directive, and recovery controls.\nAfter completing the MEASURE function, objective, repeatable, or scalable test, evaluation,\nverification, and validation (TEVV) processes including metrics, methods, and methodolo-\ngies are in place, followed, and documented. Metrics and measurement methodologies\nshould adhere to scientific, legal, and ethical norms and be carried out in an open and trans-\nparent process. New types of measurement, qualitative and quantitative, may need to be\ndeveloped. The degree to which each measurement type provides unique and meaningful\ninformation to the assessment of AI risks should be considered. Framework users will en-\nhance their capacity to comprehensively evaluate system trustworthiness, identify and track\nexisting and emergent risks, and verify efficacy of the metrics. Measurement outcomes will\nbe utilized in the MANAGE function to assist risk monitoring and response efforts. It is in-\ncumbent on Framework users to continue applying the MEASURE function to AI systems\nas knowledge, methodologies, risks, and impacts evolve over time.\nPage 28 NIST AI 100-1 AI RMF 1.0\nPractices related to measuring AI risks are described in the NIST AI RMF Playbook. Table\n3 lists the MEASURE function\u2019s categories and subcategories.\nTable 3: Categories and subcategories for the MEASURE function.\nMEASURE 1:\nAppropriate\nmethods and metrics\nare identified and\napplied.MEASURE 1.1: Approaches and metrics for measurement of AI\nrisks enumerated during the MAP function are selected for imple-\nmentation starting with the most significant AI risks. The risks\nor trustworthiness characteristics that will not \u2013 or cannot \u2013 be\nmeasured are properly documented.\nMEASURE 1.2: Appropriateness of AI metrics and effectiveness\nof existing controls are regularly assessed and updated, including\nreports of errors and potential impacts on affected communities.\nMEASURE 1.3: Internal experts who did not serve as front-line\ndevelopers for the system and/or independent assessors are in-\nvolved in regular assessments and updates. Domain experts,\nusers, AI actors external to the team that developed or deployed\nthe AI system, and affected communities are consulted in support\nof assessments as necessary per organizational risk tolerance.\nMEASURE 2:AI\nsystems are\nevaluated for\ntrustworthy\ncharacteristics.MEASURE 2.1:Test sets, metrics, and details about the tools used\nduring TEVV are documented.\nMEASURE 2.2: Evaluations involving human subjects meet ap-\nplicable requirements (including human subject protection) and\nare representative of the relevant population.\nMEASURE 2.3: AI system performance or assurance criteria\nare measured qualitatively or quantitatively and demonstrated\nfor conditions similar to deployment setting(s). Measures are\ndocumented.\nMEASURE 2.4: The functionality and behavior of the AI sys-\ntem and its components \u2013 as identified in the MAP function \u2013 are\nmonitored when in production.\nMEASURE 2.5: The AI system to be deployed is demonstrated\nto be valid and reliable. Limitations of the generalizability be-\nyond the conditions under which the technology was developed\nare documented.Categories Subcategories\nContinued on next page\nPage 29 NIST AI 100-1 AI RMF 1.0\nTable 3: Categories and subcategories for the MEASURE function. (Continued)\nMEASURE 2.6: The AI system is evaluated regularly for safety\nrisks \u2013 as identified in the MAP function. The AI system to be de-\nployed is demonstrated to be safe, its residual negative risk does\nnot exceed the risk tolerance, and it can fail safely, particularly if\nmade to operate beyond its knowledge limits. Safety metrics re-\nflect system reliability and robustness, real-time monitoring, and\nresponse times for AI system failures.\nMEASURE 2.7: AI system security and resilience \u2013 as identified\nin the MAP function \u2013 are evaluated and documented.\nMEASURE 2.8: Risks associated with transparency and account-\nability \u2013 as identified in the MAP function \u2013 are examined and\ndocumented.\nMEASURE 2.9: The AI model is explained, validated, and docu-\nmented, and AI system output is interpreted within its context \u2013\nas identified in the MAP function \u2013 to inform responsible use and\ngovernance.\nMEASURE 2.10: Privacy risk of the AI system \u2013 as identified in\ntheMAP function \u2013 is examined and documented.\nMEASURE 2.11: Fairness and bias \u2013 as identified in the MAP\nfunction \u2013 are evaluated and results are documented.\nMEASURE 2.12: Environmental impact and sustainability of AI\nmodel training and management activities \u2013 as identified in the\nMAP function \u2013 are assessed and documented.\nMEASURE 2.13: Effectiveness of the employed TEVV met-\nrics and processes in the MEASURE function are evaluated and\ndocumented.\nMEASURE 3:\nMechanisms for\ntracking identified\nAI risks over time\nare in place.MEASURE 3.1: Approaches, personnel, and documentation are\nin place to regularly identify and track existing, unanticipated,\nand emergent AI risks based on factors such as intended and ac-\ntual performance in deployed contexts.\nMEASURE 3.2: Risk tracking approaches are considered for\nsettings where AI risks are difficult to assess using currently\navailable measurement techniques or where metrics are not yet\navailable.Categories Subcategories\nContinued on next page\nPage 30 NIST AI 100-1 AI RMF 1.0\nTable 3: Categories and subcategories for the MEASURE function. (Continued)\nMEASURE 3.3: Feedback processes for end users and impacted\ncommunities to report problems and appeal system outcomes are\nestablished and integrated into AI system evaluation metrics.\nMEASURE 4:\nFeedback about\nefficacy of\nmeasurement is\ngathered and\nassessed.MEASURE 4.1:Measurement approaches for identifying AI risks\nare connected to deployment context(s) and informed through\nconsultation with domain experts and other end users. Ap-\nproaches are documented.\nMEASURE 4.2: Measurement results regarding AI system trust-\nworthiness in deployment context(s) and across the AI lifecycle\nare informed by input from domain experts and relevant AI ac-\ntors to validate whether the system is performing consistently as\nintended. Results are documented.\nMEASURE 4.3: Measurable performance improvements or de-\nclines based on consultations with relevant AI actors, in-\ncluding affected communities, and field data about context-\nrelevant risks and trustworthiness characteristics are identified\nand documented.Categories Subcategories\n5.4 Manage\nThe MANAGE function entails allocating risk resources to mapped and measured risks on\na regular basis and as defined by the GOVERN function. Risk treatment comprises plans to\nrespond to, recover from, and communicate about incidents or events.\nContextual information gleaned from expert consultation and input from relevant AI actors\n\u2013 established in GOVERN and carried out in MAP \u2013 is utilized in this function to decrease\nthe likelihood of system failures and negative impacts. Systematic documentation practices\nestablished in GOVERN and utilized in MAP and MEASURE bolster AI risk management\nefforts and increase transparency and accountability. Processes for assessing emergent risks\nare in place, along with mechanisms for continual improvement.\nAfter completing the MANAGE function, plans for prioritizing risk and regular monitoring\nand improvement will be in place. Framework users will have enhanced capacity to man-\nage the risks of deployed AI systems and to allocate risk management resources based on\nassessed and prioritized risks. It is incumbent on Framework users to continue to apply\nthe MANAGE function to deployed AI systems as methods, contexts, risks, and needs or\nexpectations from relevant AI actors evolve over time.\nPage 31 NIST AI 100-1 AI RMF 1.0\nPractices related to managing AI risks are described in the NIST AI RMF Playbook. Table\n4 lists the MANAGE function\u2019s categories and subcategories.\nTable 4: Categories and subcategories for the MANAGE function.\nMANAGE 1:AI\nrisks based on\nassessments and\nother analytical\noutput from the\nMAP and MEASURE\nfunctions are\nprioritized,\nresponded to, and\nmanaged.MANAGE 1.1: A determination is made as to whether the AI\nsystem achieves its intended purposes and stated objectives and\nwhether its development or deployment should proceed.\nMANAGE 1.2: Treatment of documented AI risks is prioritized\nbased on impact, likelihood, and available resources or methods.\nMANAGE 1.3: Responses to the AI risks deemed high priority, as\nidentified by the MAP function, are developed, planned, and doc-\numented. Risk response options can include mitigating, transfer-\nring, avoiding, or accepting.\nMANAGE 1.4: Negative residual risks (defined as the sum of all\nunmitigated risks) to both downstream acquirers of AI systems\nand end users are documented.\nMANAGE 2:\nStrategies to\nmaximize AI\nbenefits and\nminimize negative\nimpacts are planned,\nprepared,\nimplemented,\ndocumented, and\ninformed by input\nfrom relevant AI\nactors.MANAGE 2.1: Resources required to manage AI risks are taken\ninto account \u2013 along with viable non-AI alternative systems, ap-\nproaches, or methods \u2013 to reduce the magnitude or likelihood of\npotential impacts.\nMANAGE 2.2: Mechanisms are in place and applied to sustain\nthe value of deployed AI systems.\nMANAGE 2.3: Procedures are followed to respond to and recover\nfrom a previously unknown risk when it is identified.\nMANAGE 2.4: Mechanisms are in place and applied, and respon-\nsibilities are assigned and understood, to supersede, disengage, or\ndeactivate AI systems that demonstrate performance or outcomes\ninconsistent with intended use.\nMANAGE 3:AI\nrisks and benefits\nfrom third-party\nentities are\nmanaged.MANAGE 3.1: AI risks and benefits from third-party resources\nare regularly monitored, and risk controls are applied and\ndocumented.\nMANAGE 3.2: Pre-trained models which are used for develop-\nment are monitored as part of AI system regular monitoring and\nmaintenance.Categories Subcategories\nContinued on next page\nPage 32 NIST AI 100-1 AI RMF 1.0\nTable 4: Categories and subcategories for the MANAGE function. (Continued)\nMANAGE 4:Risk\ntreatments,\nincluding response\nand recovery, and\ncommunication\nplans for the\nidentified and\nmeasured AI risks\nare documented and\nmonitored regularly.MANAGE 4.1: Post-deployment AI system monitoring plans\nare implemented, including mechanisms for capturing and eval-\nuating input from users and other relevant AI actors, appeal\nand override, decommissioning, incident response, recovery, and\nchange management.\nMANAGE 4.2: Measurable activities for continual improvements\nare integrated into AI system updates and include regular engage-\nment with interested parties, including relevant AI actors.\nMANAGE 4.3: Incidents and errors are communicated to relevant\nAI actors, including affected communities. Processes for track-\ning, responding to, and recovering from incidents and errors are\nfollowed and documented.Categories Subcategories\n6. AI RMF Profiles\nAI RMF use-case profiles are implementations of the AI RMF functions, categories, and\nsubcategories for a specific setting or application based on the requirements, risk tolerance,\nand resources of the Framework user: for example, an AI RMF hiring profile or an AI\nRMF fair housing profile . Profiles may illustrate and offer insights into how risk can be\nmanaged at various stages of the AI lifecycle or in specific sector, technology, or end-use\napplications. AI RMF profiles assist organizations in deciding how they might best manage\nAI risk that is well-aligned with their goals, considers legal/regulatory requirements and\nbest practices, and reflects risk management priorities.\nAI RMF temporal profiles are descriptions of either the current state or the desired, target\nstate of specific AI risk management activities within a given sector, industry, organization,\nor application context. An AI RMF Current Profile indicates how AI is currently being\nmanaged and the related risks in terms of current outcomes. A Target Profile indicates the\noutcomes needed to achieve the desired or target AI risk management goals.\nComparing Current and Target Profiles likely reveals gaps to be addressed to meet AI risk\nmanagement objectives. Action plans can be developed to address these gaps to fulfill\noutcomes in a given category or subcategory. Prioritization of gap mitigation is driven by\nthe user\u2019s needs and risk management processes. This risk-based approach also enables\nFramework users to compare their approaches with other approaches and to gauge the\nresources needed (e.g., staffing, funding) to achieve AI risk management goals in a cost-\neffective, prioritized manner.\nPage 33 NIST AI 100-1 AI RMF 1.0\nAI RMF cross-sectoral profiles cover risks of models or applications that can be used across\nuse cases or sectors. Cross-sectoral profiles can also cover how to govern, map, measure,\nand manage risks for activities or business processes common across sectors such as the\nuse of large language models, cloud-based services or acquisition.\nThis Framework does not prescribe profile templates, allowing for flexibility in implemen-\ntation.\nPage 34 NIST AI 100-1 AI RMF 1.0\nAppendix A:\nDescriptions of AI Actor Tasks from Figures 2 and 3\nAI Design tasks are performed during the Application Context and Data and Input phases\nof the AI lifecycle in Figure 2. AI Design actors create the concept and objectives of AI\nsystems and are responsible for the planning, design, and data collection and processing\ntasks of the AI system so that the AI system is lawful and fit-for-purpose. Tasks include ar-\nticulating and documenting the system\u2019s concept and objectives, underlying assumptions,\ncontext, and requirements; gathering and cleaning data; and documenting the metadata\nand characteristics of the dataset. AI actors in this category include data scientists, do-\nmain experts, socio-cultural analysts, experts in the field of diversity, equity, inclusion,\nand accessibility, members of impacted communities, human factors experts (e.g., UX/UI\ndesign), governance experts, data engineers, data providers, system funders, product man-\nagers, third-party entities, evaluators, and legal and privacy governance.\nAI Development tasks are performed during the AI Model phase of the lifecycle in Figure\n2. AI Development actors provide the initial infrastructure of AI systems and are responsi-\nble for model building and interpretation tasks, which involve the creation, selection, cali-\nbration, training, and/or testing of models or algorithms. AI actors in this category include\nmachine learning experts, data scientists, developers, third-party entities, legal and privacy\ngovernance experts, and experts in the socio-cultural and contextual factors associated with\nthe deployment setting.\nAI Deployment tasks are performed during the Task and Output phase of the lifecycle in\nFigure 2. AI Deployment actors are responsible for contextual decisions relating to how\nthe AI system is used to assure deployment of the system into production. Related tasks\ninclude piloting the system, checking compatibility with legacy systems, ensuring regu-\nlatory compliance, managing organizational change, and evaluating user experience. AI\nactors in this category include system integrators, software developers, end users, oper-\nators and practitioners, evaluators, and domain experts with expertise in human factors,\nsocio-cultural analysis, and governance.\nOperation and Monitoring tasks are performed in the Application Context/Operate and\nMonitor phase of the lifecycle in Figure 2. These tasks are carried out by AI actors who are\nresponsible for operating the AI system and working with others to regularly assess system\noutput and impacts. AI actors in this category include system operators, domain experts, AI\ndesigners, users who interpret or incorporate the output of AI systems, product developers,\nevaluators and auditors, compliance experts, organizational management, and members of\nthe research community.\nTest, Evaluation, Verification, and Validation (TEVV) tasks are performed throughout\nthe AI lifecycle. They are carried out by AI actors who examine the AI system or its\ncomponents, or detect and remediate problems. Ideally, AI actors carrying out verification\nPage 35 NIST AI 100-1 AI RMF 1.0\nand validation tasks are distinct from those who perform test and evaluation actions. Tasks\ncan be incorporated into a phase as early as design, where tests are planned in accordance\nwith the design requirement.\n\u2022 TEVV tasks for design, planning, and data may center on internal and external vali-\ndation of assumptions for system design, data collection, and measurements relative\nto the intended context of deployment or application.\n\u2022 TEVV tasks for development (i.e., model building) include model validation and\nassessment.\n\u2022 TEVV tasks for deployment include system validation and integration in production,\nwith testing, and recalibration for systems and process integration, user experience,\nand compliance with existing legal, regulatory, and ethical specifications.\n\u2022 TEVV tasks for operations involve ongoing monitoring for periodic updates, testing,\nand subject matter expert (SME) recalibration of models, the tracking of incidents\nor errors reported and their management, the detection of emergent properties and\nrelated impacts, and processes for redress and response.\nHuman Factors tasks and activities are found throughout the dimensions of the AI life-\ncycle. They include human-centered design practices and methodologies, promoting the\nactive involvement of end users and other interested parties and relevant AI actors, incor-\nporating context-specific norms and values in system design, evaluating and adapting end\nuser experiences, and broad integration of humans and human dynamics in all phases of the\nAI lifecycle. Human factors professionals provide multidisciplinary skills and perspectives\nto understand context of use, inform interdisciplinary and demographic diversity, engage\nin consultative processes, design and evaluate user experience, perform human-centered\nevaluation and testing, and inform impact assessments.\nDomain Expert tasks involve input from multidisciplinary practitioners or scholars who\nprovide knowledge or expertise in \u2013 and about \u2013 an industry sector, economic sector, con-\ntext, or application area where an AI system is being used. AI actors who are domain\nexperts can provide essential guidance for AI system design and development, and inter-\npret outputs in support of work performed by TEVV and AI impact assessment teams.\nAI Impact Assessment tasks include assessing and evaluating requirements for AI system\naccountability, combating harmful bias, examining impacts of AI systems, product safety,\nliability, and security, among others. AI actors such as impact assessors and evaluators\nprovide technical, human factor, socio-cultural, and legal expertise.\nProcurement tasks are conducted by AI actors with financial, legal, or policy management\nauthority for acquisition of AI models, products, or services from a third-party developer,\nvendor, or contractor.\nGovernance and Oversight tasks are assumed by AI actors with management, fiduciary,\nand legal authority and responsibility for the organization in which an AI system is de-\nPage 36 NIST AI 100-1 AI RMF 1.0\nsigned, developed, and/or deployed. Key AI actors responsible for AI governance include\norganizational management, senior leadership, and the Board of Directors. These actors\nare parties that are concerned with the impact and sustainability of the organization as a\nwhole.\nAdditional AI Actors\nThird-party entities include providers, developers, vendors, and evaluators of data, al-\ngorithms, models, and/or systems and related services for another organization or the or-\nganization\u2019s customers or clients. Third-party entities are responsible for AI design and\ndevelopment tasks, in whole or in part. By definition, they are external to the design, devel-\nopment, or deployment team of the organization that acquires its technologies or services.\nThe technologies acquired from third-party entities may be complex or opaque, and risk\ntolerances may not align with the deploying or operating organization.\nEnd users of an AI system are the individuals or groups that use the system for specific\npurposes. These individuals or groups interact with an AI system in a specific context. End\nusers can range in competency from AI experts to first-time technology end users.\nAffected individuals/communities encompass all individuals, groups, communities, or\norganizations directly or indirectly affected by AI systems or decisions based on the output\nof AI systems. These individuals do not necessarily interact with the deployed system or\napplication.\nOther AI actors may provide formal or quasi-formal norms or guidance for specifying\nand managing AI risks. They can include trade associations, standards developing or-\nganizations, advocacy groups, researchers, environmental groups, and civil society\norganizations .\nThe general public is most likely to directly experience positive and negative impacts of\nAI technologies. They may provide the motivation for actions taken by the AI actors. This\ngroup can include individuals, communities, and consumers associated with the context in\nwhich an AI system is developed or deployed.\nPage 37 NIST AI 100-1 AI RMF 1.0\nAppendix B:\nHow AI Risks Differ from Traditional Software Risks\nAs with traditional software, risks from AI-based technology can be bigger than an en-\nterprise, span organizations, and lead to societal impacts. AI systems also bring a set of\nrisks that are not comprehensively addressed by current risk frameworks and approaches.\nSome AI system features that present risks also can be beneficial. For example, pre-trained\nmodels and transfer learning can advance research and increase accuracy and resilience\nwhen compared to other models and approaches. Identifying contextual factors in the MAP\nfunction will assist AI actors in determining the level of risk and potential management\nefforts.\nCompared to traditional software, AI-specific risks that are new or increased include the\nfollowing:\n\u2022 The data used for building an AI system may not be a true or appropriate representa-\ntion of the context or intended use of the AI system, and the ground truth may either\nnot exist or not be available. Additionally, harmful bias and other data quality issues\ncan affect AI system trustworthiness, which could lead to negative impacts.\n\u2022 AI system dependency and reliance on data for training tasks, combined with in-\ncreased volume and complexity typically associated with such data.\n\u2022 Intentional or unintentional changes during training may fundamentally alter AI sys-\ntem performance.\n\u2022 Datasets used to train AI systems may become detached from their original and in-\ntended context or may become stale or outdated relative to deployment context.\n\u2022 AI system scale and complexity (many systems contain billions or even trillions of\ndecision points) housed within more traditional software applications.\n\u2022 Use of pre-trained models that can advance research and improve performance can\nalso increase levels of statistical uncertainty and cause issues with bias management,\nscientific validity, and reproducibility.\n\u2022 Higher degree of difficulty in predicting failure modes for emergent properties of\nlarge-scale pre-trained models.\n\u2022 Privacy risk due to enhanced data aggregation capability for AI systems.\n\u2022 AI systems may require more frequent maintenance and triggers for conducting cor-\nrective maintenance due to data, model, or concept drift.\n\u2022 Increased opacity and concerns about reproducibility.\n\u2022 Underdeveloped software testing standards and inability to document AI-based prac-\ntices to the standard expected of traditionally engineered software for all but the\nsimplest of cases.\n\u2022 Difficulty in performing regular AI-based software testing, or determining what to\ntest, since AI systems are not subject to the same controls as traditional code devel-\nopment.\nPage 38 NIST AI 100-1 AI RMF 1.0\n\u2022 Computational costs for developing AI systems and their impact on the environment\nand planet.\n\u2022 Inability to predict or detect the side effects of AI-based systems beyond statistical\nmeasures.\nPrivacy and cybersecurity risk management considerations and approaches are applicable\nin the design, development, deployment, evaluation, and use of AI systems. Privacy and\ncybersecurity risks are also considered as part of broader enterprise risk management con-\nsiderations, which may incorporate AI risks. As part of the effort to address AI trustworthi-\nness characteristics such as \u201cSecure and Resilient\u201d and \u201cPrivacy-Enhanced,\u201d organizations\nmay consider leveraging available standards and guidance that provide broad guidance to\norganizations to reduce security and privacy risks, such as, but not limited to, the NIST Cy-\nbersecurity Framework, the NIST Privacy Framework, the NIST Risk Management Frame-\nwork, and the Secure Software Development Framework. These frameworks have some\nfeatures in common with the AI RMF. Like most risk management approaches, they are\noutcome-based rather than prescriptive and are often structured around a Core set of func-\ntions, categories, and subcategories. While there are significant differences between these\nframeworks based on the domain addressed \u2013 and because AI risk management calls for\naddressing many other types of risks \u2013 frameworks like those mentioned above may inform\nsecurity and privacy considerations in the MAP ,MEASURE , and MANAGE functions of the\nAI RMF.\nAt the same time, guidance available before publication of this AI RMF does not compre-\nhensively address many AI system risks. For example, existing frameworks and guidance\nare unable to:\n\u2022 adequately manage the problem of harmful bias in AI systems;\n\u2022 confront the challenging risks related to generative AI;\n\u2022 comprehensively address security concerns related to evasion, model extraction, mem-\nbership inference, availability, or other machine learning attacks;\n\u2022 account for the complex attack surface of AI systems or other security abuses enabled\nby AI systems; and\n\u2022 consider risks associated with third-party AI technologies, transfer learning, and off-\nlabel use where AI systems may be trained for decision-making outside an organiza-\ntion\u2019s security controls or trained in one domain and then \u201cfine-tuned\u201d for another.\nBoth AI and traditional software technologies and systems are subject to rapid innovation.\nTechnology advances should be monitored and deployed to take advantage of those devel-\nopments and work towards a future of AI that is both trustworthy and responsible.\nPage 39 NIST AI 100-1 AI RMF 1.0\nAppendix C:\nAI Risk Management and Human-AI Interaction\nOrganizations that design, develop, or deploy AI systems for use in operational settings\nmay enhance their AI risk management by understanding current limitations of human-\nAI interaction. The AI RMF provides opportunities to clearly define and differentiate the\nvarious human roles and responsibilities when using, interacting with, or managing AI\nsystems.\nMany of the data-driven approaches that AI systems rely on attempt to convert or represent\nindividual and social observational and decision-making practices into measurable quanti-\nties. Representing complex human phenomena with mathematical models can come at the\ncost of removing necessary context. This loss of context may in turn make it difficult to\nunderstand individual and societal impacts that are key to AI risk management efforts.\nIssues that merit further consideration and research include:\n1.Human roles and responsibilities in decision making and overseeing AI systems\nneed to be clearly defined and differentiated. Human-AI configurations can span\nfrom fully autonomous to fully manual. AI systems can autonomously make deci-\nsions, defer decision making to a human expert, or be used by a human decision\nmaker as an additional opinion. Some AI systems may not require human oversight,\nsuch as models used to improve video compression. Other systems may specifically\nrequire human oversight.\n2.Decisions that go into the design, development, deployment, evaluation, and use\nof AI systems reflect systemic and human cognitive biases. AI actors bring their\ncognitive biases, both individual and group, into the process. Biases can stem from\nend-user decision-making tasks and be introduced across the AI lifecycle via human\nassumptions, expectations, and decisions during design and modeling tasks. These\nbiases, which are not necessarily always harmful, may be exacerbated by AI system\nopacity and the resulting lack of transparency. Systemic biases at the organizational\nlevel can influence how teams are structured and who controls the decision-making\nprocesses throughout the AI lifecycle. These biases can also influence downstream\ndecisions by end users, decision makers, and policy makers and may lead to negative\nimpacts.\n3.Human-AI interaction results vary. Under certain conditions \u2013 for example, in\nperceptual-based judgment tasks \u2013 the AI part of the human-AI interaction can am-\nplify human biases, leading to more biased decisions than the AI or human alone.\nWhen these variations are judiciously taken into account in organizing human-AI\nteams, however, they can result in complementarity and improved overall perfor-\nmance.\nPage 40 NIST AI 100-1 AI RMF 1.0\n4.Presenting AI system information to humans is complex. Humans perceive and\nderive meaning from AI system output and explanations in different ways, reflecting\ndifferent individual preferences, traits, and skills.\nThe GOVERN function provides organizations with the opportunity to clarify and define\nthe roles and responsibilities for the humans in the Human-AI team configurations and\nthose who are overseeing the AI system performance. The GOVERN function also creates\nmechanisms for organizations to make their decision-making processes more explicit, to\nhelp counter systemic biases.\nThe MAP function suggests opportunities to define and document processes for operator\nand practitioner proficiency with AI system performance and trustworthiness concepts, and\nto define relevant technical standards and certifications. Implementing MAP function cat-\negories and subcategories may help organizations improve their internal competency for\nanalyzing context, identifying procedural and system limitations, exploring and examining\nimpacts of AI-based systems in the real world, and evaluating decision-making processes\nthroughout the AI lifecycle.\nThe GOVERN and MAP functions describe the importance of interdisciplinarity and demo-\ngraphically diverse teams and utilizing feedback from potentially impacted individuals and\ncommunities. AI actors called out in the AI RMF who perform human factors tasks and\nactivities can assist technical teams by anchoring in design and development practices to\nuser intentions and representatives of the broader AI community, and societal values. These\nactors further help to incorporate context-specific norms and values in system design and\nevaluate end user experiences \u2013 in conjunction with AI systems.\nAI risk management approaches for human-AI configurations will be augmented by on-\ngoing research and evaluation. For example, the degree to which humans are empowered\nand incentivized to challenge AI system output requires further studies. Data about the fre-\nquency and rationale with which humans overrule AI system output in deployed systems\nmay be useful to collect and analyze.\nPage 41 NIST AI 100-1 AI RMF 1.0\nAppendix D:\nAttributes of the AI RMF\nNIST described several key attributes of the AI RMF when work on the Framework first\nbegan. These attributes have remained intact and were used to guide the AI RMF\u2019s devel-\nopment. They are provided here as a reference.\nThe AI RMF strives to:\n1. Be risk-based, resource-efficient, pro-innovation, and voluntary.\n2. Be consensus-driven and developed and regularly updated through an open, trans-\nparent process. All stakeholders should have the opportunity to contribute to the AI\nRMF\u2019s development.\n3. Use clear and plain language that is understandable by a broad audience, including\nsenior executives, government officials, non-governmental organization leadership,\nand those who are not AI professionals \u2013 while still of sufficient technical depth to\nbe useful to practitioners. The AI RMF should allow for communication of AI risks\nacross an organization, between organizations, with customers, and to the public at\nlarge.\n4. Provide common language and understanding to manage AI risks. The AI RMF\nshould offer taxonomy, terminology, definitions, metrics, and characterizations for\nAI risk.\n5. Be easily usable and fit well with other aspects of risk management. Use of the\nFramework should be intuitive and readily adaptable as part of an organization\u2019s\nbroader risk management strategy and processes. It should be consistent or aligned\nwith other approaches to managing AI risks.\n6. Be useful to a wide range of perspectives, sectors, and technology domains. The AI\nRMF should be universally applicable to any AI technology and to context-specific\nuse cases.\n7. Be outcome-focused and non-prescriptive. The Framework should provide a catalog\nof outcomes and approaches rather than prescribe one-size-fits-all requirements.\n8. Take advantage of and foster greater awareness of existing standards, guidelines, best\npractices, methodologies, and tools for managing AI risks \u2013 as well as illustrate the\nneed for additional, improved resources.\n9. Be law- and regulation-agnostic. The Framework should support organizations\u2019\nabilities to operate under applicable domestic and international legal or regulatory\nregimes.\n10. Be a living document. The AI RMF should be readily updated as technology, under-\nstanding, and approaches to AI trustworthiness and uses of AI change and as stake-\nholders learn from implementing AI risk management generally and this framework\nin particular.\nPage 42 This publication is available free of charge from:\nhttps://doi.org/10.6028/NIST.AI.100-1", "cleaned_text_custom_parser_2": null, "best_field": "cleaned_text_custom_parser", "enhanced": true, "date_published": null, "image": null, "num_tokens_oai_tokenizer": null, "num_tokens_gemini_tokenizer": null}, "favicon": null, "state": "candidate", "enhanced": true, "answer": null, "valid_answer": false, "source_citation_index": 4, "user_web_page_id": null, "status": "ready", "reader": false, "rank": 2, "reason": "The source is relevant as it is the downloadable PDF version of the AI Risk Management Framework, essential for detailed reference."}, {"id": 648330, "type": "GenericWebPage", "mime_type": "text/html", "content_type": null, "url": "https://digitalgovernmenthub.org/library/nist-ai-risk-management-framework-playbook/", "slug": "digitalgovernmenthub", "title": "NIST AI Risk Management Framework Playbook - Digital Government Hub", "snippet": "This resource provides practical ... the AI lifecycle. It includes detailed examples, risk mitigation strategies, and documentation templates to support trustworthy, transparent, and accountable AI use....", "image": null, "date_published": null, "media": [], "products": [], "source": {"id": 648330, "website_id": 124569, "url": "https://digitalgovernmenthub.org/library/nist-ai-risk-management-framework-playbook/", "title": "NIST AI Risk Management Framework Playbook - Digital Government Hub", "snippet": "This resource provides practical ... the AI lifecycle. It includes detailed examples, risk mitigation strategies, and documentation templates to support trustworthy, transparent, and accountable AI use....", "cleaned_text": null, "cleaned_text_custom_parser": null, "cleaned_text_custom_parser_2": null, "best_field": null, "enhanced": false, "date_published": null, "image": null, "num_tokens_oai_tokenizer": null, "num_tokens_gemini_tokenizer": null}, "favicon": null, "state": "candidate", "enhanced": false, "answer": null, "valid_answer": false, "source_citation_index": 5, "user_web_page_id": null, "status": "not_processed", "reader": false, "rank": 2, "reason": "The source is relevant as it highlights the companion playbook, which provides actionable strategies based on the AI RMF."}, {"id": 648329, "type": "GenericWebPage", "mime_type": "text/html", "content_type": null, "url": "https://airc.nist.gov/airmf-resources/crosswalks/", "slug": "nist", "title": "Crosswalk Documents - AIRC - NIST AI Resource Center", "snippet": "AI Risk Management Framework crosswalk documents that map concepts between the AI RMF and other guidelines, frameworks, and standards. Submit your own crosswalks for potential inclusion at mailto:AIframework@nist.gov", "image": null, "date_published": null, "media": [], "products": [], "source": {"id": 648329, "website_id": 123921, "url": "https://airc.nist.gov/airmf-resources/crosswalks/", "title": "Crosswalk Documents - AIRC - NIST AI Resource Center", "snippet": "AI Risk Management Framework crosswalk documents that map concepts between the AI RMF and other guidelines, frameworks, and standards. Submit your own crosswalks for potential inclusion at mailto:AIframework@nist.gov", "cleaned_text": null, "cleaned_text_custom_parser": null, "cleaned_text_custom_parser_2": null, "best_field": null, "enhanced": false, "date_published": null, "image": null, "num_tokens_oai_tokenizer": null, "num_tokens_gemini_tokenizer": null}, "favicon": null, "state": "candidate", "enhanced": false, "answer": null, "valid_answer": false, "source_citation_index": 6, "user_web_page_id": null, "status": "not_processed", "reader": false, "rank": 2, "reason": "The source is relevant as it offers additional crosswalk documents that help link the AI RMF to other standards."}, {"id": 648343, "type": "GenericWebPage", "mime_type": "text/html", "content_type": null, "url": "https://www.nist.gov/video/introduction-nist-ai-risk-management-framework-ai-rmf-10-explainer-video", "slug": "nist", "title": "Introduction to the NIST AI Risk Management Framework (AI RMF 1.0): An Explainer Video | NIST", "snippet": "Introduction to the NIST AI Risk Management Framework (AI RMF 1.0): An Explainer Video", "image": null, "date_published": null, "media": [], "products": [], "source": {"id": 648343, "website_id": 1909, "url": "https://www.nist.gov/video/introduction-nist-ai-risk-management-framework-ai-rmf-10-explainer-video", "title": "Introduction to the NIST AI Risk Management Framework (AI RMF 1.0): An Explainer Video | NIST", "snippet": "Introduction to the NIST AI Risk Management Framework (AI RMF 1.0): An Explainer Video", "cleaned_text": null, "cleaned_text_custom_parser": null, "cleaned_text_custom_parser_2": null, "best_field": null, "enhanced": false, "date_published": null, "image": null, "num_tokens_oai_tokenizer": null, "num_tokens_gemini_tokenizer": null}, "favicon": "https://www.nist.gov/themes/custom/nist_www/favicon.ico", "state": "candidate", "enhanced": false, "answer": null, "valid_answer": false, "source_citation_index": 7, "user_web_page_id": null, "status": "not_processed", "reader": false, "rank": 2, "reason": "The source is relevant as it provides an explainer video on the NIST AI Risk Management Framework, which can help users understand its application."}, {"id": 648346, "type": "GenericWebPage", "mime_type": "text/html", "content_type": null, "url": "https://www.nist.gov/news-events/events/2023/01/nist-ai-risk-management-framework-ai-rmf-10-launch", "slug": "nist", "title": "NIST AI Risk Management Framework (AI RMF 1.0) Launch | NIST", "snippet": "The NIST AI RMF is a framework to better manage risks to individuals, organizations, and society associated with artificial intelligence (AI).", "image": null, "date_published": null, "media": [], "products": [], "source": {"id": 648346, "website_id": 1909, "url": "https://www.nist.gov/news-events/events/2023/01/nist-ai-risk-management-framework-ai-rmf-10-launch", "title": "NIST AI Risk Management Framework (AI RMF 1.0) Launch | NIST", "snippet": "The NIST AI RMF is a framework to better manage risks to individuals, organizations, and society associated with artificial intelligence (AI).", "cleaned_text": null, "cleaned_text_custom_parser": null, "cleaned_text_custom_parser_2": null, "best_field": null, "enhanced": false, "date_published": null, "image": null, "num_tokens_oai_tokenizer": null, "num_tokens_gemini_tokenizer": null}, "favicon": "https://www.nist.gov/themes/custom/nist_www/favicon.ico", "state": "candidate", "enhanced": false, "answer": null, "valid_answer": false, "source_citation_index": 8, "user_web_page_id": null, "status": "not_processed", "reader": false, "rank": 2, "reason": "The source provides a comprehensive overview of the AI RMF's purpose and its voluntary use for managing AI risks."}]