On 29 April 2024 the US National Institute of Standards and Technology published the initial public draft of AI 600-1, a cross-sector guide for managing generative AI risks. It identifies twelve generative-AI-specific risks — CBRN information, hallucination, harmful content, data privacy, environmental impacts, bias, human-AI configuration, information integrity, information security, intellectual property, sexual content and value-chain risks — and is designed to accompany AI RMF 1.0.
Grouping the twelve
The twelve are not a flat list. They fall into three groups by source. Risks generated by the model itself include hallucination, harmful content, bias and sexual content. Data and rights risks include privacy and intellectual property. System and ecosystem risks include information integrity, information security, CBRN information, environmental impact, human-AI configuration and value-chain risk.
The grouping matters because it locates responsibility. Model-level problems are largely within the developer’s control; information integrity and value-chain risks span developers, deployers and platforms, and no single party can address them alone.
Relationship with AI RMF 1.0
AI RMF 1.0 provides a general framework of govern, map, measure and manage. AI 600-1 adds generative-specific actions on top. The framework is voluntary and creates no legal obligation, but it supplies a vocabulary that regulators, customers and insurers already understand.
The practical benefit is lower communication cost. When a customer asks in diligence how hallucination risk is managed, a risk register aligned with AI 600-1 is more persuasive than a verbal explanation.
Building a risk register
A workable register maps each of the twelve risks onto the product: whether it applies, what controls exist, and who accepts the residual risk. Risks judged inapplicable should carry a stated reason rather than an empty cell.
Model evaluation should correspond to the register. Hallucination maps to fact-checking and citation mechanisms, bias to test datasets, intellectual property to provenance review of training data, and information security to access control and adversarial testing. Each item needs a record, not only a conclusion.
Transparency and documentation
AI 600-1 emphasises transparency and documentation in two places: disclosure to users and internal decision records. Disclosure answers whether users know they are interacting with a generative system; documentation answers whether the judgment can be explained afterwards.
The file does not need to be heavy, but it must be traceable. Version, data source, evaluation conclusion and reason for change are enough for most situations. An evaluation report without a version number, by contrast, carries almost no evidentiary weight in a dispute.
How Chinese companies can use it
The guide is a draft, not a compliance standard, and it does not replace mandatory rules in any jurisdiction. Its value is a risk taxonomy that lets companies compare EU, US and Chinese requirements on one page.
For AI companies serving several markets, organising internal material in one risk language and adding jurisdictional differences on top is more sustainable than maintaining three separate document sets.
For security controls and adversarial testing of generative systems, see our cybersecurity compliance service; for training data and privacy risk, see data compliance; for provenance and rights review, see IP protection.
This article is based on the initial public draft of NIST AI 600-1, the cross-sector guide for managing generative AI risks, published on 29 April 2024, and on public information. It is for reference only and does not constitute legal advice.