Skip to content

Generative AI Profile

NIST AI 600-1, published in July 2024, is a cross-sectoral companion Profile for generative artificial intelligence. It helps organisations identify generative-AI risks and select actions that fit their goals and priorities.

Use the Profile when the system generates or materially transforms:

  • Text.
  • Images.
  • Audio.
  • Video.
  • Code.
  • Synthetic data.
  • Multimodal content.

It is also relevant to applications built on foundation models, chatbots, copilots and other generative services.

The Generative AI Profile:

  • Supplements the AI RMF Core.
  • Does not replace GOVERN, MAP, MEASURE or MANAGE.
  • Does not create NIST certification.
  • Should be tailored to the selected system.
  • Should be revisited as models, data, suppliers and capabilities change.

Gamut highlights the risk families identified in NIST AI 600-1:

Risk familyAssessment questions
CBRN information or capabilitiesCould the system meaningfully facilitate harmful chemical, biological, radiological or nuclear knowledge or capability?
ConfabulationCan plausible but false output cause decisions, harm or loss of trust?
Dangerous, violent or hateful contentCan the system generate, amplify or enable harmful content?
Data privacyCan training, prompts, retrieval, memory or outputs expose or infer personal information?
Environmental impactsAre training, inference and infrastructure impacts understood and managed?
Harmful bias and homogenisationCan outputs disadvantage groups, erase variation or reinforce harmful patterns?
Human-AI configurationAre reliance, automation bias, anthropomorphism and human roles appropriately managed?
Information integrityCan the system create or amplify misleading, manipulated or unverifiable information?
Information securityCan prompts, models, tools, data, outputs or integrations be attacked or misused?
Intellectual propertyAre rights, licences, provenance and output-use risks understood?
Obscene, degrading or abusive contentCan the system generate or facilitate abusive or exploitative material?
Value chain and component integrationAre model, data, platform, tool and downstream responsibilities and dependencies controlled?

For each relevant family:

  1. Define the credible scenario.
  2. Identify affected people, assets and decisions.
  3. Identify model, data, prompt, retrieval, tool and supplier contributors.
  4. Review preventive and detective controls.
  5. Select evaluation methods.
  6. Define pass criteria and limitations.
  7. Record incidents, near misses and external evidence.
  8. Determine treatment and monitoring.
  9. Map the work back to relevant Core outcomes.
  • Model and system cards.
  • Supplier documentation and change notices.
  • Data provenance and rights records.
  • Prompt and retrieval architecture.
  • Red-team and misuse testing.
  • Hallucination or factuality evaluation.
  • Safety-filter and policy testing.
  • Privacy testing.
  • Bias and representation evaluation.
  • Security testing.
  • Human-oversight procedures.
  • Content labelling and provenance mechanisms.
  • Incident, complaint and takedown records.
  • Monitoring by use case, user group and model version.

Evaluate the actual use case. The same model may have very different risk when used for creative drafting, medical support, recruitment or autonomous tool use.

Consider what the system can do with:

  • Long context.
  • Retrieval.
  • Tools.
  • Code execution.
  • Memory.
  • Multimodal input.
  • Fine-tuning.
  • External data.

Consider who can use it, which information it can reach and whether outputs can trigger action.

Consider volume, speed, user reach, automation and the ability to replicate harmful output.

Model updates, system prompts, retrieval sources, safety settings and suppliers can change risk without changing the product name.

For a customer-support assistant:

  • MAP identifies high-impact topics, escalation rules and knowledge limits.
  • MEASURE evaluates factuality using representative enquiries and current source material.
  • MANAGE requires abstention, source presentation, human escalation and incident response.
  • GOVERN assigns ownership and review cadence.

An acceptable average accuracy does not excuse a dangerous error class. Evaluate the severity and distribution of errors, not only the mean.

Using a third-party model does not transfer all responsibility. Record:

  • Supplier responsibilities.
  • Application-provider responsibilities.
  • Deploying-organisation responsibilities.
  • Customer or user responsibilities.
  • Evidence dependencies.
  • Escalation and contingency arrangements.

Reassess after:

  • New foundation model or model version.
  • Fine-tuning or prompt change.
  • New retrieval source.
  • New tool or action permission.
  • New user population.
  • New high-impact use.
  • Safety-control change.
  • Material incident or misuse.
  • New external research on capability or risk.

NIST announced a critical-infrastructure Profile concept note in April 2026. A concept note should not be described as a final Profile. Check the official AI RMF page for current status.