Skip to content

R0043/2026-03-28/Q001/H1

Research R0043 — Sycophancy Vocabulary
Run 2026-03-28
Query Q001
Hypothesis H1

Statement

Each domain has developed its own specific terminology for the sycophancy phenomenon, creating a rich but fragmented vocabulary across all eight specified fields.

Status

Current: Partially supported

The evidence shows that several domains — particularly aviation/human factors, healthcare, and defense — have well-developed terminology. However, this terminology largely addresses the human side of the problem (overreliance, complacency, automation bias) rather than the system behavior side (sycophancy, people-pleasing). The vocabulary is rich but addresses a different causal framing than AI safety's "sycophancy."

Supporting Evidence

Evidence Summary
SRC01-E01 Parasuraman & Manzey define automation bias and complacency as distinct but overlapping constructs used across aviation, healthcare, and military
SRC04-E01 CSET report identifies automation bias as a cross-domain term used in Tesla, aviation, and military contexts
SRC07-E01 DoD uses "calibrated trust," "overtrust," and "appropriate trust" as distinct domain terminology
SRC08-E01 FAA/aviation uses "automation complacency" as established terminology
SRC09-E01 Healthcare uses "acquiescence problem" and "automation bias" for clinical decision support

Contradicting Evidence

Evidence Summary
SRC02-E01 Within AI safety research, the term "sycophancy" is used consistently — not fragmented — suggesting convergence rather than fragmentation within a domain
SRC10-E01 The vocabulary gap paper argues existing terminology is fundamentally insufficient, not merely fragmented

Reasoning

H1 is partially supported because substantial domain-specific vocabulary does exist: aviation has "automation complacency," defense has "calibrated trust" and "overtrust," healthcare has "acquiescence problem" and "automation bias," and enterprise evaluation references "agreeableness bias." However, this vocabulary is not as richly developed for the system behavior as H1 predicts. Most domains frame the problem from the human operator's perspective (why humans over-trust) rather than from the system's perspective (why AI agrees when it should not). The AI safety term "sycophancy" has no direct equivalent in most regulated industries.

Relationship to Other Hypotheses

H1 overlaps significantly with H3 because the evidence shows vocabulary exists but with systematic asymmetries. The key distinction: H1 predicts each domain has its own terms, while H3 predicts some domains have terms and others borrow or lack them. The evidence more strongly supports H3.