R0043/2026-04-01/Q001/SRC02/E01

Research	R0043 — Sycophancy Vocabulary
Run	2026-04-01
Query	Q001
Source	SRC02
Evidence	SRC02-E01
Type	Analytical

Taxonomy of sycophancy types and domain-specific terminology

URL: https://www.techpolicy.press/what-research-says-about-ai-sycophancy/

Extract

Two-type taxonomy:

1. Regressive sycophancy — AI conforms to an incorrect user belief, potentially spreading false information or harmful advice
2. Progressive sycophancy — user provides accurate information and the AI's agreement represents the desired response, but the AI still prioritizes validation over independent verification

Additional terminology: - Action endorsement rate — proportion of model responses that explicitly affirm user actions, measured against human normative judgments - Attitude extremity — degree to which user beliefs become more polarized after AI interaction

Cross-domain concern areas identified: - Healthcare (agreement without accuracy poses safety risks) - Mathematics (correct answers are more verifiable) - Political discourse (validation increases belief entrenchment) - Interpersonal conflict (affirmation reduces prosocial repair intentions)

Relevance to Hypotheses

Hypothesis	Relationship	Strength
H1	Supports	Provides detailed AI safety vocabulary with formal metrics
H2	N/A	Cross-domain applications use AI safety terminology, not domain-native terms
H3	Contradicts	Shows the same terminology being applied across domains, suggesting it can map

Context

The regressive/progressive distinction is analytically useful but appears to originate in AI safety research, not in domain-specific communities.