NIST AI 600-1: risk profile for generative AI systems
In one sentence NIST publishes AI 600-1, specific guidance for generative AI risks: 12 unique risk categories including data poisoning, hallucination, prompt injection, homogenization, and value chain risks. Complements the AI RMF and is referenced in Biden EO compliance.
NIST's general AI risk management framework (AI RMF) covers all AI systems. But generative AI — the kind that generates text, images, code — has specific problems that a general framework does not capture well. You can train an image classifier for years without it ever producing a hallucination. With generative models it is different.
NIST published AI 600-1 as a companion document to the AI RMF specifically dedicated to the unique risks of generative systems. It identifies 12 risk categories that emerge specifically from these systems: hallucinations (the model invents convincing facts), data poisoning (training data is manipulated), prompt injection, cultural homogenization (all systems tend toward the same outputs), confabulation, amplified bias, and others.
For each category, the document provides descriptions, examples, suggested metrics, and mitigation practices. It is written for three types of audience: technical teams building systems, governance teams defining policies, and risk officers who must report to company boards.
Anyone trying to demonstrate compliance with US or European AI regulations will find in AI 600-1 a shared vocabulary and a structure recognized by authorities.
Companies
NIST
Tools
—
Tags
Sources