Skip to content

R0043/2026-04-01/Q001/H1

Research R0043 — Sycophancy Vocabulary
Run 2026-04-01
Query Q001
Hypothesis H1

Statement

Each of the eight named domains has developed its own specific terminology for AI behavior that prioritizes user agreement over accuracy, and a comprehensive cross-domain vocabulary map can be constructed.

Status

Current: Supported

Supporting Evidence

Evidence Summary
SRC01-E01 AI safety has established "sycophancy" as the primary term with defined behavioral categories
SRC02-E01 TechPolicy.Press documents "regressive" and "progressive" sycophancy subtypes
SRC03-E01 Defense/military uses "automation bias" and "automation complacency" as distinct concepts
SRC04-E01 Psychology/psychiatry uses "delusion confirmation" and "AI-induced psychosis"
SRC05-E01 Oxford/Cambridge researchers distinguish overreliance (behavior) from automation bias (cognition) from sycophancy (model property)
SRC08-E01 Aviation uses "misuse/disuse/abuse" taxonomy plus "complacency" and "overtrust"

Contradicting Evidence

Evidence Summary
SRC06-E01 "Acquiescence bias" in LLMs produced opposite findings to expected (models biased toward "no"), suggesting the terminology mapping is not straightforward

Reasoning

Evidence confirms that at least seven of the eight named domains have developed specific terminology. The eighth domain (academic integrity) addresses the problem primarily through borrowed terms ("confirmation bias") rather than domain-specific vocabulary. The vocabulary map is constructable but reveals that terms describe different facets of the phenomenon rather than exact synonyms.

Relationship to Other Hypotheses

H1 is the strongest hypothesis. H2 (some domains lack terminology) is partially supported for academic integrity and enterprise software evaluation, which rely on borrowed terms rather than domain-native vocabulary. H3 (terminology does not map to the same phenomenon) is partially supported — the terms describe overlapping but non-identical concerns.