NIST Generative AI Profile: AI 600-1 Actions
NIST AI 600-1 is a voluntary Generative AI Profile with 12 risk categories. It provides actions for confabulation and human-oversight evidence.

On July 26, 2024, NIST published the Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, designated NIST AI 600-1. The 64-page document gives organizations a GenAI-specific companion to AI RMF 1.0. It identifies 12 risk categories, then places suggested actions under the four AI RMF functions.
That last sentence carries two details that matter. NIST describes the AI RMF and this cross-sectoral Profile as voluntary. Section 3 calls its entries "suggested actions," and organizations choose actions based on their role and context. The document is useful as a risk and control map. It is a poor candidate for checkbox theatre.
I want to correct a second source of confusion at the outset. OMB M-24-10 has been obsolete since April 3, 2025. Its replacement, M-25-21, sets current federal agency policy for high-impact AI. The GenAI Profile remains useful guidance, while OMB's requirements arise under a separate memorandum.
TL;DR
NIST AI 600-1 is voluntary guidance for managing generative-AI risk. It identifies 12 risk categories and maps suggested actions to the AI RMF functions. MP-2.3-001 puts human oversight inside a ground-truth evaluation process; MG-4.3-002 calls for an error and near-miss record.
The official NIST AI 600-1 structure
The official NIST AI 600-1 publication has three working layers.
Section 1 begins by defining scope. NIST designed the Profile for organizations that design or develop generative AI. It also covers organizations that put it into service or evaluate it. The Profile extends the NIST AI Risk Management Framework, so its actions sit under the RMF's four functions instead of forming a separate certification system.
Section 2 describes 12 risks. CBRN Information or Capabilities concerns access to dangerous-material information and weapon-related design capabilities. Confabulation covers confidently presented erroneous content. The category for harmful content covers material that can cause harm or support illegal activity. Data Privacy concerns leakage and unauthorized use, including disclosure or de-anonymization of sensitive data.
Environmental Impacts covers compute alongside energy and water effects. Harmful Bias or Homogenization includes amplified bias with performance disparities and output homogeneity. Human-AI Configuration includes automation bias and over-reliance, plus anthropomorphism, algorithmic aversion, and emotional entanglement. Information Integrity addresses misinformation and disinformation, including obscured uncertainty or origin.
Information Security covers offensive cyber capability and attacks against AI components, including compromise of code or model weights. Intellectual Property includes trade-secret exposure and unauthorized reproduction, plus related rights issues. Obscene, Degrading, and/or Abusive Content includes abusive imagery plus synthetic child sexual abuse material and nonconsensual intimate imagery. Value Chain and Component Integration covers opaque third-party components and weak supplier vetting, including untraceable upstream software or data.
Section 3 contains the suggested actions. Each action has a precise identifier such as GV-3.2-003 or MG-4.3-002. The prefix identifies the RMF function: GV for GOVERN, MP for MAP, MS for MEASURE, and MG for MANAGE. The following numbers identify the RMF subcategory and the specific Profile action.
A search phrase such as "GenAI classification B, GenAI is permitted as an assistive tool" describes an organization's possible internal usage tier. It is absent from the official NIST taxonomy. Teams can create an assistive-use tier, but the policy should label it as an internal classification and map it to applicable NIST actions.
Confabulation and human oversight in NIST AI 600-1
NIST uses confabulation for confidently stated erroneous or false content, also commonly called hallucination or fabrication. The Profile notes that a generated answer can contradict its prompt. It can also conflict with known facts or earlier statements in the same context. It can invent logic and citations that make the answer look more credible.
For confabulation, NIST action MP-2.3-001 calls for comparing output against known ground truth with human oversight and automated evaluation. Human oversight is a defined evaluation activity in this action, not a vague instruction to look over model output after a decision has already reached a user.
Treat that activity as a bounded review record. For each sampled output, a domain reviewer compares the answer with an identified source or ground-truth set. The record identifies the model configuration and classifies the result against a threshold set before the evaluation begins. It also records the reviewer's disposition and the reason for it. Those fields turn a review into evidence that another person can retrace.
Human oversight also needs a trigger. A financial eligibility recommendation or clinical summary can route to a reviewer before the result informs the decision. Benefits determinations call for the same treatment. A low-consequence writing aid can use periodic samples, with the sample rate and escalation rule documented for the use case. The choice belongs in the risk analysis because NIST ties oversight to context and system performance, as well as the impact of an error.
When a reviewer finds a confabulation, retain the prompt or a privacy-safe reference with the output and source used for comparison. Keep the model route with the date and disposition. Include the corrective action in the same record. MP-2.3-003 calls for documented fact-checking techniques, while MG-4.3-002 calls for tracking reported errors and near-misses with negative impacts. An error register that omits the evaluation context leaves engineering without the detail needed to reproduce the failure.
I would reject a review process that records only "approved" or "rejected" in a ticket. That pair of labels hides the source and threshold, along with the correction that makes human oversight useful during a later incident review.
Picture a case reviewer with a model-generated summary on the left monitor and a 300-page source file on the right. The summary gives a wrong medication or an incorrect contract date in polished prose. Confidence in the wording becomes part of the failure mechanism.
The Profile ties confabulation to several concrete actions. MP-2.3-001 suggests comparing output with known ground truth through human and automated evaluation. MP-2.3-003 suggests documented fact-checking techniques, especially when output draws on multiple or unknown sources. MG-4.3-002 covers policies and procedures to record and track reported errors with near-misses or negative impacts. A practical implementation needs an evaluation set and reviewer instructions. It also needs an error intake path with a retained record connecting the faulty output to its model version and use context. Our guide to LLM hallucination controls covers that operational layer in more detail.
My opinion is blunt: a generic instruction telling staff to "verify AI output" is paperwork dressed as a control. A reviewer needs an authoritative source and a defined decision threshold. The review record needs a place to capture disagreement. High-consequence output should remain outside automated decision paths until that test exists.
Correct action IDs change the control map
A subcategory label summarizes an objective. A full action ID points to a suggested activity. Mixing them produces bad audit requests, and the original version of this article did exactly that.
These examples come directly from Section 3 of NIST AI 600-1:
- GV-3.2-003 suggests acceptable-use policies for GenAI interfaces and modalities, including human-AI configurations and refusal criteria. GV-3.2 concerns roles and responsibilities alongside oversight.
- GV-6.2-002 covers incidents involving third-party GenAI data or systems. GV-6.2-003 addresses a third-party incident-response plan, including rehearsals, retrospective learning, communications, and legal review.
- MP-2.3-001 addresses output accuracy and quality against known ground truth. MP-2.3-003 focuses on fact-checking. MP-2.3 covers scientific integrity plus testing, evaluation, verification, and validation.
- MP-4.1-001 suggests periodic monitoring of generated content for privacy exposure. MP-4.1-006 addresses policies for third-party intellectual property and training data. MP-4.1 spans technology and legal risks in components, including third-party software or data.
- MP-5.1-002 calls for ranking potential content-provenance harms by likelihood and impact. MP-5.1-003 addresses disclosure of GenAI use after considering purpose, audience, risk, and frequency.
- MS-2.7-001 covers security measures for compromised dependencies, data breaches, eavesdropping, model theft, extraction, and inference attacks. MS-2.7 evaluates security and resilience. It is broader than privacy testing.
- MS-2.12-002 suggests documenting anticipated environmental impacts in product design decisions. MS-2.12-003 addresses measurement or estimation of energy and water use during training, fine-tuning, and deployment. MS-2.12 concerns environmental impact and sustainability.
- MG-2.2-001 suggests comparing output against organizational risk tolerance and reviewing generated content against defined guidelines. MG-2.2-007 suggests real-time auditing to track and validate lineage or authenticity. MG-2.2 concerns sustaining value in deployed systems.
- MG-4.3-001 addresses after-action assessments and incident communication. MG-4.3-002 covers error and near-miss tracking. MG-4.3-003 addresses legally required incident reporting.
This precision affects ownership. A sustainability lead might own MS-2.12 evidence. Security engineering can own part of MS-2.7. Legal and data governance have central work under MP-4.1. Incident response owns much of MG-4.3, supported by records supplied by application and infrastructure teams.
AI RMF functions need different evidence
GOVERN evidence establishes authority and accountability. Useful artifacts include an approved acceptable-use policy and named owners. Add review records, third-party incident plans, and proof that staff received role-specific instructions. A policy PDF sitting untouched since 2024 offers weak evidence of active governance.
MAP evidence connects a system to its context. Record the intended purpose with its users and affected people. Document upstream data, external connections, foreseeable misuse, and the likelihood and magnitude of identified impacts. An AI inventory provides the index. The system card and data records provide detail, alongside supplier documentation and impact analysis.
MEASURE evidence demonstrates that somebody tested a claim. Preserve the evaluation dataset version with the model configuration. Record the test procedure, result, reviewer, date, threshold, and disposition. A production monitoring alert and a pre-deployment benchmark answer different questions. Our NIST AI RMF MEASURE controls guide gives teams a more detailed testing structure.
MANAGE evidence shows what happened after a risk became visible. Retain risk acceptance with mitigation changes and release decisions. Keep monitoring results, user feedback, incidents, corrective actions, and after-action review. The record should connect the decision to the policy and test result that informed it. A useful evidence package lets an assessor trace one sampled output without opening six dashboards and a spreadsheet called final-v7.
M-25-21 replaced M-24-10
OMB Memorandum M-25-21, issued April 3, 2025, expressly rescinded and replaced M-24-10. Current federal content should use M-25-21's term high-impact AI. The earlier "safety-impacting" and "rights-impacting" categories belong to superseded policy.
M-25-21 defines high-impact AI around output serving as a principal basis for decisions or actions with legal or material effects on rights and safety. For covered agency uses, Section 4 requires minimum practices. They include pre-deployment testing and a documented AI impact assessment, plus ongoing monitoring with periodic human review where feasible. Agencies also need operator training, suitable human oversight, a timely human-review and appeal route when appropriate, and public or end-user feedback where appropriate.
Those are OMB requirements for covered agency use. NIST AI 600-1 remains voluntary guidance, and M-25-21 does not convert every suggested Profile action into a federal mandate. An agency can use the Profile to organize risks and candidate mitigations while maintaining a separate M-25-21 compliance matrix tied to the memorandum's operative text.
Federal contractor scope also needs care. OMB Memorandum M-25-22 governs agency acquisition practices and excludes AI used incidentally at a contractor's option when the contract neither directs nor requires that use. Contract language and the acquired system determine the contractor's duties. Agency direction and actual performance also matter. A federal customer logo on a slide cannot answer that analysis.
A defensible implementation sequence
Start with one GenAI use case rather than a framework-wide spreadsheet. Name its owner and intended purpose. Document users, the model route, data classes, affected people, upstream components, and the decision consequence. Select the applicable risks among NIST's 12 categories and explain each exclusion.
Select full action IDs during the next working session, then assign an owner and evidence artifact to every chosen action. An acceptable-use policy plus a test of prohibited request handling can support action GV-3.2-003. A versioned ground-truth evaluation provides evidence for MP-2.3-001. The error register should connect sampled requests and corrective actions under MG-4.3-002.
Testing comes after the owners and artifacts are clear. Record the threshold before running the evaluation. Capture exceptions, reviewer disagreement, and failed cases. The AI compliance logging guide explains the fields needed to reconstruct individual runtime decisions.
Finally, review the map after a model or prompt-template change. Apply the same review when a retrieval source, integration, user population, or decision purpose changes. NIST's Profile repeatedly emphasizes context and lifecycle activity. A static annual attestation misses the point.
DeepInspect
DeepInspect sits on a narrow enforcement boundary: authenticated HTTP requests routed through the gateway between users or agents and LLM endpoints. At that point it can apply identity-aware policy and record the principal, route, model, and timestamp. It also retains the policy version, classification result, permit or deny decision, and related response metadata.
Those records can contribute evidence to GV-3.2-003 acceptable-use enforcement, MG-2.2 output monitoring, MS-2.7 security measurement for observable routed traffic, and MG-4.3 incident investigation. They also make a sampled event retrievable: an investigator can connect a person or agent, a model route, the policy active at that moment, and the gateway decision.
The wider Profile remains a multi-owner program. DeepInspect leaves training-data provenance and model weights to the teams that can observe them. Benchmark design, environmental measurement, IP review, supplier contracts, interface design, model red-teaming, local execution, and non-HTTP activity also sit outside the gateway's capture boundary. A gateway record contributes evidence. It never satisfies the whole NIST GenAI Profile.
Frequently asked questions
- Is the NIST GenAI Profile mandatory?
NIST describes the AI RMF and its Generative AI Profile as voluntary resources. A law or regulator can separately require practices that overlap with the Profile. A contract or agency policy can create the same outcome. Teams should identify that separate authority instead of relabeling NIST's suggested actions as mandates.
- Is "GenAI classification B" an official NIST category?
No official "classification B" appears in NIST AI 600-1. An organization may define an internal tier in which GenAI is permitted as an assistive tool, then map its restrictions and evidence to applicable Profile actions. The tier name remains internal taxonomy.
- What does NIST mean by confabulation?
NIST defines confabulation as confidently presented erroneous or false content. The term includes output that diverges from the prompt, contradicts earlier statements, or invents supporting logic and citations. The production control set can include ground-truth evaluation, documented fact-checking, human review, error intake, and near-miss tracking.
- How does M-25-21 relate to NIST AI 600-1?
M-25-21 is current OMB policy for federal agency AI use and contains minimum practices for covered high-impact AI. NIST AI 600-1 is voluntary cross-sectoral guidance. Agencies may use the Profile to structure risk analysis, but the source of a federal duty remains M-25-21 or another applicable authority.
- Can one audit log prove implementation of the Profile?
A runtime log can support selected actions by showing identity, model route, policy state, classifications, decisions, and timestamps for traffic within its capture boundary. Other actions require evaluations, supplier records, training-data documentation, environmental measurements, legal analysis, human-factors work, and incident procedures. A credible evidence package keeps those boundaries visible.