Skip to content

R0043/2026-04-01/Q002/H2

Research R0043 — Sycophancy Vocabulary
Run 2026-04-01
Query Q002
Hypothesis H2

Statement

Regulated industries have not addressed the sycophancy phenomenon in their formal requirements, leaving a regulatory gap where agreeable-but-wrong AI output is not covered.

Status

Current: Partially Supported

Supporting Evidence

Evidence Summary
SRC06-E01 No explicit regulatory framework addresses sycophancy; OpenAI's response was voluntary self-regulation
SRC02-E01 NIST AI RMF does not name sycophancy; addresses "confabulation" but this covers factual errors, not agreement-seeking behavior

Contradicting Evidence

Evidence Summary
SRC01-E01 EU AI Act does address automation bias, a related concept
SRC03-E01 SR 11-7 effective challenge requirement indirectly addresses the problem

Reasoning

The gap is real but not total. No regulation directly addresses sycophancy-as-model-behavior, but several address adjacent concerns (automation bias, confabulation, model risk). The gap is specifically in the nexus between model behavior and regulatory language.

Relationship to Other Hypotheses

H2 overstates the gap — coverage exists but under different terminology. H3 is the more accurate characterization.