NIST wrote a voluntary risk map for generative AI systems

The conversation

In July 2024, the US National Institute of Standards and Technology published NIST AI 600-1, a generative-AI profile built on the agency's earlier AI Risk Management Framework, which NIST released in January 2023. The profile's own text describes itself as a 'cross-sectoral profile' applying that framework's functions, govern, map, measure and manage, specifically to generative systems such as chatbots. NIST's current framework page describes this document as retrieved on 16 September 2026, alongside the wider Risk Management Framework it extends.

What the documents show

NIST's own document is explicit that the underlying framework 'is intended for voluntary use', not a binding rule. It names twelve risks it treats as unique to or worsened by generative systems, including confabulation, which the document defines as a system generating 'confidently stated but erroneous or false content'; dangerous, violent, or hateful content; data privacy; harmful bias and homogenisation; human-AI configuration, covering how a person interacts with and comes to rely on an automated system; information integrity and information security; and intellectual property, among others. For each risk, a separate table maps suggested actions to the underlying framework's existing subcategories, tagged so a reader can see which risk a given action addresses, and the document notes plainly that not every suggested action applies to every organisation building or deploying a system.

The system boundary

NIST's document specifically discusses chatbots when describing risk from dangerous or unhelpful content, stating that 'studies have observed that users may disclose mental health issues in conversations with chatbots' and that users can react badly to an unhelpful response during a moment of distress, an observation the profile ties to output controls that can still miss novel or indirect prompts. The document is guidance for an organisation's own risk-management process, not a testing standard or a certification; it tells an organisation what to consider and suggests actions, but the responsibility for applying, verifying or ignoring any suggested action sits with the organisation itself.

Where it fails

Because AI 600-1 is voluntary guidance rather than an enforceable rule, its own text does not describe consequences for an organisation that identifies a risk and declines to act on it; that gap sits with whatever separate law or regulator might otherwise apply. The document also acknowledges limits within its own risk descriptions, noting for example that there is 'no agreed upon method to estimate' environmental impacts from generative systems, meaning some of its own risk categories lack a settled way to measure them. A builder or operator should read a claim of following NIST's framework with that voluntary status in mind.

• Which of the twelve named risks apply to this specific chatbot's use case, and which suggested actions has the organisation actually adopted?
• Does an internal claim of following the framework rest on a documented process, or only on having read the document?
• Where NIST's own guidance says a risk lacks an agreed measurement method, what does the organisation use instead?

AI 600-1 gives a shared vocabulary for generative-AI risk; it does not, by NIST's own description, decide what any particular deployment must do about it.