NIST
Last reviewed
Sources
23 citations
Review status
Source-backed
Revision
v2 ยท 2,257 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
23 citations
Review status
Source-backed
Revision
v2 ยท 2,257 words
Add missing citations, update stale details, or suggest a clearer explanation.
The National Institute of Standards and Technology (NIST) is a non-regulatory agency of the US Department of Commerce responsible for measurement science, standards, and technology, and it is the US government's central technical institution for artificial intelligence. Founded on March 3, 1901 as the National Bureau of Standards and renamed in 1988, it is one of the oldest physical science laboratories in the United States, with campuses in Gaithersburg, Maryland and Boulder, Colorado.[1] During the 2020s NIST became the federal hub for AI governance and evaluation: it published the AI Risk Management Framework (AI RMF 1.0) on January 26, 2023,[2] housed the US AI Safety Institute from November 2023,[3] reorganized that body into the Center for AI Standards and Innovation (CAISI) in June 2025,[4] and runs evaluation programs that test everything from facial recognition algorithms to frontier language models.
NIST's involvement with AI long predates the current policy debate. The agency has run measurement-driven technology evaluations for decades: its Text REtrieval Conference (TREC) has benchmarked information retrieval systems since 1992, the MNIST handwritten digit dataset that became a foundational machine learning benchmark was derived from NIST's handwriting sample databases, and its Face Recognition Vendor Test began in 2000 and grew into the de facto global benchmark for the biometrics industry.[5]
Congress formalized NIST's AI mandate in the National Artificial Intelligence Initiative Act of 2020, which directed the agency to develop a voluntary risk management framework for trustworthy AI.[2] Further direction came from both administrations: President Biden's Executive Order 14110 (signed October 30, 2023) assigned NIST work on red-team testing guidance, synthetic content authentication, and dual-use foundation models,[3] while the Trump administration's America's AI Action Plan (July 23, 2025) re-tasked the agency toward pro-innovation standards, evaluations of Chinese frontier models, and revision of earlier guidance.[6] As of December 2025, NIST was led by acting director Craig Burkhardt.[7]
NIST released version 1.0 of the Artificial Intelligence Risk Management Framework (AI RMF, document NIST AI 100-1) on January 26, 2023, following an open consensus process that included a formal request for information, multiple public drafts, and a series of workshops.[2] The framework is voluntary, sector-agnostic, and organized around four functions, Govern, Map, Measure, and Manage, applied across the AI lifecycle. It defines seven characteristics of trustworthy AI: valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; and fair with harmful bias managed.[2] NIST treats the RMF as a living document, with a formal review planned no later than 2028. The agency frames the document as a resource "intended for voluntary use" to improve the ability to incorporate trustworthiness into the design, development, use, and evaluation of AI systems.[2]
On July 26, 2024, NIST published a Generative AI Profile (NIST AI 600-1) as a companion to the framework, fulfilling a task in Executive Order 14110. The profile maps the RMF's four functions onto twelve risks it considers unique to or amplified by generative AI, including confabulation, information integrity, information security, intellectual property, and easier access to information about chemical, biological, radiological, and nuclear weapons.[8]
The RMF has been widely referenced in corporate AI governance programs and US policy discussions. In July 2025, America's AI Action Plan directed NIST to revise the framework to "eliminate references to misinformation, Diversity, Equity, and Inclusion, and climate change," a change the plan framed as removing ideological bias from federal guidance; critics argued it politicized a technical document.[6]
The Department of Commerce announced the US AI Safety Institute (AISI) on November 1, 2023, two days after Executive Order 14110 was signed, with Vice President Kamala Harris unveiling the plan during the UK's AI Safety Summit at Bletchley Park. Housed within NIST, the institute was charged with developing standards for the safety, security, and testing of advanced models and with evaluating emerging risks.[3] In February 2024, Commerce Secretary Gina Raimondo named Elizabeth Kelly the institute's first director and launched the AI Safety Institute Consortium (AISIC), which gathered more than 200 member organizations including Apple, Meta, and Microsoft.[9] In April 2024, Paul Christiano, who pioneered reinforcement learning from human feedback at OpenAI and founded the Alignment Research Center, was appointed head of AI safety.[10]
In August 2024 the institute signed first-of-their-kind memoranda of understanding with OpenAI and Anthropic, giving it access to major new models both before and after public release and enabling joint safety research in collaboration with the UK AI Safety Institute.[11] Announcing the agreements on August 29, 2024, director Kelly said, "Safety is essential to fueling breakthrough technological innovation," adding that the institute looked forward to "beginning our technical collaborations with Anthropic and OpenAI to advance the science of AI safety."[11] In November 2024 it established the Testing Risks of AI for National Security (TRAINS) taskforce with partners across the defense and intelligence community, and it convened the inaugural meeting of the International Network of AI Safety Institutes in San Francisco.[12]
The institute's trajectory changed under the second Trump administration, which rescinded Executive Order 14110 in January 2025. Kelly stepped down as director on February 5, 2025.[13] On June 3, 2025, Commerce Secretary Howard Lutnick announced that the institute would be transformed into the Center for AI Standards and Innovation, stating that "for far too long, censorship and regulations have been used under the guise of national security" and casting the new center as pro-innovation and pro-science.[4] CAISI serves as industry's primary point of contact in the federal government for AI testing and collaborative research; it focuses on "demonstrable risks" such as cybersecurity, biosecurity, and chemical weapons, assesses security vulnerabilities and foreign influence in adversary AI systems, coordinates with national security agencies, and represents US interests in international standards development.[14] Observers noted that the reorganization dropped the safety-centric framing of the Biden era in favor of national security and competitiveness goals.[15]
CAISI's most visible output has been comparative evaluations of Chinese models, a task assigned by America's AI Action Plan. Its September 30, 2025 report on DeepSeek evaluated DeepSeek-R1, R1-0528, and V3.1 against GPT-5, GPT-5-mini, gpt-oss, and Claude Opus 4 across 19 benchmarks. It concluded that the DeepSeek models lagged US models on performance, cost, and security, were far more susceptible to agent hijacking and jailbreaking, echoed Chinese state narratives, and had nonetheless seen downloads on model-sharing platforms grow nearly 1,000 percent since January 2025.[16] A follow-up evaluation of DeepSeek V4 Pro, published May 1, 2026, estimated that the model trailed the US frontier by roughly eight months, with an Item Response Theory-estimated Elo score of 800 versus 1260 for GPT-5.5 on CAISI's benchmark suite, though it remained cheaper than GPT-5.4 mini on five of seven benchmarks tested.[17]
The center has also expanded its industry agreements. On May 5, 2026, CAISI announced pre-deployment testing agreements with Google DeepMind, Microsoft, and xAI, joining existing partners OpenAI and Anthropic, and reported having completed more than 40 evaluations, including of unreleased models.[18] Earlier in 2026 it signed a memorandum of understanding with the General Services Administration to bring evaluation science into federal AI procurement (March 18, 2026) and a cooperative research agreement with OpenMined on privacy-preserving evaluation methods (March 27, 2026).[14] The international network in which the US participates through CAISI was renamed the International Network for Advanced AI Measurement, Evaluation and Science in early 2026, with the United Kingdom serving as coordinator.[19]
| Program | Launched | Focus |
|---|---|---|
| FRTE / FATE (formerly FRVT) | 2000; split August 2023 | Face recognition accuracy (1:1 verification, 1:N identification); face analysis including morph detection, spoof detection, and age estimation |
| NIST GenAI | April 2024 | Challenge evaluations of generators and detectors of AI-generated text and images |
| ARIA | May 2024 | Sociotechnical testing of AI systems in realistic contexts of use |
| AISI/CAISI frontier evaluations | 2024 | Pre- and post-deployment capability and security testing of frontier models |
The Face Recognition Vendor Test (FRVT), run since 2000, was split in August 2023 into the Face Recognition Technology Evaluation (FRTE), which covers one-to-one verification and one-to-many identification accuracy, and the Face Analysis Technology Evaluation (FATE), which covers image quality, morph and presentation attack detection, and age estimation.[5] The program's December 2019 demographic effects study (NISTIR 8280) found that many algorithms produced false positive rates that differed across demographic groups; for one-to-one matching the differentials "often ranged from a factor of 10 to 100 times, depending on the individual algorithm," with higher false positive rates for Asian, African American, and Native American faces.[20] NIST scientist and lead author Patrick Grother cautioned, "While it is usually incorrect to make statements across algorithms, we found empirical evidence for the existence of demographic differentials in the majority of the face recognition algorithms we studied."[20] The result was widely cited in debates over law enforcement use of face recognition.
NIST GenAI, launched in April 2024, runs challenge problems that pit generative models against discriminator systems designed to detect AI-generated content. The 2024 pilot focused on text-to-text generation and detection and was documented in report NIST AI 700-1; the 2025 cycle extended to AI-generated images.[21]
The Assessing Risks and Impacts of AI (ARIA) program, launched May 28, 2024, evaluates AI systems in realistic settings rather than on static benchmarks, combining three levels of testing (model testing, red teaming, and field testing) and aggregating results into a Contextual Robustness Index (CoRIx). Its 0.1 pilot ran 508 testing sessions across three scenarios involving meal planning, travel pathfinding, and avoiding television spoilers.[22]
NIST anchors the US government's role in international AI standardization. Its Plan for Global Engagement on AI Standards (NIST AI 100-5, July 2024) set out a strategy for working through bodies such as ISO/IEC, and America's AI Action Plan assigns CAISI responsibility for advancing US interests, and countering Chinese influence, in those forums.[6][14]
The agency also publishes technical guidance and software for AI security. Dioptra, an open-source testbed re-released on July 26, 2024, lets developers measure how adversarial attacks degrade machine learning model performance; it shipped alongside the Generative AI Profile and SP 800-218A, which adapts NIST's secure software development framework to generative AI.[23] NIST's adversarial machine learning taxonomy (NIST AI 100-2), first issued in January 2024 and updated in March 2025, catalogs attacks such as data poisoning, evasion, and prompt injection together with mitigations.[23]
Under the Strategy for American Technology Leadership that followed the AI Action Plan, NIST announced two AI Economic Security Centers on December 22, 2025, operated with the MITRE Corporation under a $20 million investment: one to develop AI tools for US manufacturing productivity and one to defend critical infrastructure against cyberthreats. It also announced plans for an AI for Resilient Manufacturing Institute funded at up to $70 million over five years.[7]