Source: NIST AI 600-1, Generative AI Profile

NIST's cross-sector profile of the AI RMF for generative AI: 12 GAI risks and about 200 suggested actions keyed to Govern/Map/Measure/Manage subcategories.

Contents
  • What: NIST Trustworthy and Responsible AI publication (PDF)
  • Publisher: NIST
  • Date: 2024-07-26
  • Original: NIST AI 600-1

Summary

A companion to the AI RMF for generative AI. It defines risks that are new to GAI or made worse by it, and gives suggested actions tagged with AI RMF subcategory IDs (e.g. GV-1.6-001) and the risks they address. The public working group behind it focused on four themes: governance, content provenance, pre-deployment testing and incident disclosure.

Key claims

  • The 12 GAI risks: CBRN information or capabilities; confabulation; dangerous, violent or hateful content; data privacy; environmental impacts; harmful bias and homogenization; human-AI configuration; information integrity; information security; intellectual property; obscene, degrading or abusive content; value chain and component integration.
  • Confabulation is a natural result of generating from a statistical distribution. It is most dangerous when users act on confident falsehoods, and models can also produce confabulated justifications or citations.
  • Data privacy risk covers memorization and leakage of training data and the inference of sensitive attributes that were never disclosed.
  • Example actions: add GAI systems to the organization’s AI inventory, with provenance and third-party details (GV-1.6); keep a plan to halt a system posing unacceptable risk (GV-1.3-007); keep approved GAI vendor lists and contract clauses that allow evaluating third-party processes (GV-6.1).
  • Current pre-deployment testing is often inadequate or mismatched to the deployment context. Benchmarks and anecdotal tests (exams, games) do not establish validity in the field.
  • AI incidents should be documented and disclosed; at publication no formal reporting channel existed.

Feeds

AI governance and compliance

Suggest changes →