The National Institute of Standards and Technology (NIST) published its Generative AI Profile, document NIST AI 600-1, on July 26, 2024, and it names 12 risks that generative AI creates or makes worse. They run from confabulation (confident false output) to untraceable third-party components, and the profile is voluntary guidance, not a regulation. It is the most specific official U.S. list of what can go wrong with systems that generate text, images, audio, and video.
Key takeaways
- NIST AI 600-1 is a companion to the AI Risk Management Framework (AI RMF 1.0, January 2023) and is intended for voluntary use.
- The 12 named risks include confabulation, data privacy, harmful bias, information integrity, intellectual property, and environmental impacts.
- Several risks come from how people use the system, not from the model alone. Over-reliance and emotional entanglement are on the list.
- The profile covers only risks with an existing evidence base, and it says measurement of AI safety is still immature.
- The executive order that commissioned the profile was revoked on January 20, 2025, yet NIST still publishes the document.
What the profile is and what status it has
A NIST "profile" applies the AI RMF's four functions (Govern, Map, Measure, Manage) to a specific technology. This one targets generative AI, defined in the document through Executive Order 14110 as models that emulate the structure of input data to produce synthetic content such as images, video, audio, and text. NIST's publication page lists it as report 600-1 and shows an update on April 8, 2026.
The profile was developed under Section 4.1 of Executive Order 14110, which a White House order of January 20, 2025 listed among the revoked orders. The profile itself did not disappear with it. Because the AI RMF is voluntary, the profile has no enforcement mechanism: it informs governance choices, procurement questions, and internal policies, and it can be cited as a reference point for what reasonable risk management looks like.
The 12 risks in plain language
The document labels each risk by its source or target. The wording below paraphrases the definitions in section 2 of the profile.
| Risk | What it means |
|---|---|
| CBRN information or capabilities | Easier access to design knowledge for chemical, biological, radiological, or nuclear weapons |
| Confabulation | Confidently stated content that is false, often called hallucination |
| Dangerous, violent, or hateful content | Easier production of threatening or radicalizing material and of advice on self-harm or illegal acts |
| Data privacy | Leakage, unauthorized use, or de-anonymization of personal or sensitive data |
| Environmental impacts | Energy and resource use in training and running models |
| Harmful bias or homogenization | Amplified social bias, uneven performance across groups or languages, and outputs that all look alike |
| Human-AI configuration | Over-reliance, automation bias, anthropomorphizing, or emotional entanglement |
| Information integrity | Cheaper mass production of misinformation, disinformation, and deepfakes |
| Information security | Lower barriers to hacking, phishing, and malware, plus new attack surfaces on models and data |
| Intellectual property | Replication of copyrighted or licensed content without authorization, and exposure of trade secrets |
| Obscene, degrading, or abusive content | Synthetic child sexual abuse material and nonconsensual intimate images |
| Value chain and component integration | Untraceable or poorly vetted third-party data, models, and libraries |
Risks that depend on the person using the tool
Three entries describe problems that appear in the interaction, not in the software. Confabulation arises, the profile explains, because generative models approximate the statistical distribution of their training data: a language model predicts the next word, which can yield accurate text or fluent error. The document notes the risk is greatest in open-ended long answers and in fields that need domain expertise, and it gives the example of a made-up summary of patient records leading to a wrong diagnosis.
The human-AI configuration entry covers the opposite failure as well. Experts may be needlessly averse and skip useful applications, while others over-rely on the output, a pattern the profile calls automation bias. NIST adds that automation bias can intensify confabulation and bias problems, because a confident wrong answer is accepted without a check. Even the terms "hallucination" and "fabrication" are flagged in a footnote as anthropomorphizing.
Risks that sit above a single product
Some entries only make sense at the level of the whole ecosystem. The profile describes "algorithmic monocultures": when many organizations rely on the same model for consequential decisions such as hiring or lending, one flaw can cause correlated failures. The value chain entry makes a related point, that foundation models are hard to build and so are reused widely, which makes a problem in one upstream component hard to trace.
The environmental entry is notable for what it admits. NIST repeats an estimate that training one transformer language model can emit as much carbon as 300 round-trip flights between San Francisco and New York, and then states that no agreed method exists to estimate the environmental impact of generative AI. That is a statement about the state of measurement, not a verdict on any particular model.
What the profile does not cover
The scope is deliberately narrow. NIST says it limited the list to risks with an existing empirical evidence base and left out speculative risks that might arise in more advanced future systems. Its working group concentrated on four considerations: governance, content provenance, pre-deployment testing, and incident disclosure, so the suggested actions concentrate there. The document also says that the science of AI measurement and safety is generally immature and that visibility into training data is limited, which makes some risks hard to estimate. A future revision is expected to add subcategories and risks, and NIST's AI RMF page says the base framework is being revised under the White House AI Action Plan with no completion date given.
This article describes a government document and does not give legal, security, or compliance advice. Organizations deciding how to use the profile should consult qualified professionals and the primary text.
FAQ
Is the NIST Generative AI Profile a law?
No. NIST describes the AI RMF and its profiles as intended for voluntary use, so nothing in the document is legally binding on its own. Other laws or contracts may refer to it.
Did the revocation of Executive Order 14110 cancel the profile?
The order was revoked on January 20, 2025, but NIST's publication page still lists AI 600-1, with an update dated April 8, 2026. Its future revision status is not stated.
Does the profile say how serious each risk is?
Not as a ranking. It says risk is a combination of likelihood and magnitude, that levels vary by use case, and that organizations should allocate effort according to their own context.
Is the profile useful outside the United States?
It is a U.S. document, but a cross-sectoral list of risk categories can be read by anyone, including Canadian organizations, as one reference among several.
Related reading
- Computer vision applications and challenges
- Deep learning applications and advances
- A research paper on where AI is heading
- Limits of toy problems in AI research
More in AI Basics.
Last verified: October 2026
Sources
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1), National Institute of Standards and Technology, July 2024
- Publication page for NIST AI 600-1, National Institute of Standards and Technology, updated April 8, 2026
- AI Risk Management Framework overview, National Institute of Standards and Technology, 2026
- Initial Rescissions of Harmful Executive Orders and Actions, The White House, January 20, 2025