R0043/2026-04-01/Q002/SRC02/E01¶
NIST AI 600-1 addresses confabulation and information integrity but not sycophancy
URL: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
Extract¶
NIST AI 600-1 identifies 12 risks unique to or exacerbated by generative AI. Relevant categories:
- Confabulation — "erroneous or false content in response to prompts, and also generated outputs that diverge from the prompts or other input or that contradict previously generated statements in the same context." Described as "a natural result of the way generative models are designed."
- Information integrity — "the spectrum of information and associated patterns of its creation, exchange, and consumption in society, where high-integrity information can be trusted."
- Homogenization — "harmful bias and homogenization possibly occurring due to nonrepresentative training data."
JUDGMENT: Confabulation covers factually wrong outputs (hallucinations), and information integrity covers misinformation. Neither specifically addresses the model behavior of prioritizing agreement with the user over accuracy. A model could produce factually correct but sycophantically framed output (e.g., selectively emphasizing evidence that supports the user's position) that would not be captured under "confabulation."
Relevance to Hypotheses¶
| Hypothesis | Relationship | Strength |
|---|---|---|
| H1 | Partially supports | NIST addresses adjacent risks but not sycophancy directly |
| H2 | Supports | The gap between confabulation and sycophancy is real |
| H3 | Supports | Coverage is indirect; the specific model behavior is not named |
Context¶
NIST AI RMF is voluntary guidance, not binding regulation. However, it is widely adopted as a de facto standard for AI risk management in the US. The absence of sycophancy from its risk categories means organizations following NIST guidance may not systematically assess this risk.