Q003 — Assessment¶


Research	R0043 — Sycophancy Vocabulary
Run	2026-04-01
Query	Q003

BLUF¶

The broader AI terminology gap has been identified as a problem by multiple researchers and organizations, with active efforts to create shared taxonomies. However, the specific sycophancy vocabulary gap — the absence of cross-domain terminology for AI agreement-seeking behavior — has not been prioritized in any identified taxonomy or glossary effort. Even the most comprehensive cross-domain AI taxonomies (Huwyler's 53-threat taxonomy, IAPP's 100+ term glossary, CSIRO's harmonised terminology framework) exclude sycophancy and all related behavioral terms. The sycophancy vocabulary gap is a recognized-but-unaddressed subset of the broader AI governance terminology problem.

Probability¶

Rating: N/A (open-ended query)

Confidence in assessment: Medium-High

Confidence rationale: Strong evidence that the broader gap is recognized. Strong evidence (from absence in multiple glossaries/taxonomies) that the sycophancy-specific gap is not yet addressed. The main uncertainty is whether efforts exist that were not discovered in this search.

Reasoning Chain¶

Trilateral Research (2025) explicitly diagnosed the AI terminology gap and proposed five solutions including minimal glossaries and translation tables. However, their proposed glossary focuses on governance concepts (explainability, robustness, fairness) and does not include sycophancy. [SRC01-E01, Medium-High reliability, High relevance]
The CSIRO/UNSW team proposed a harmonised terminology framework for AI evaluation that addresses process terminology (evaluation, testing, verification) but not behavioral risk terminology. [SRC02-E01, Medium-High reliability, Medium-High relevance]
Huwyler's standardized threat taxonomy covers 9 domains and 53 sub-threats, bridging technical and business language, but does not include sycophancy. The closest category is "Unreliable Outputs" but this covers hallucinations, not agreement-seeking. [SRC03-E01, Medium-High reliability, Medium relevance]
The IAPP's 100+ term AI governance glossary excludes sycophancy and ALL related behavioral safety terms (agreeableness bias, acquiescence bias, automation bias, overreliance). [SRC04-E01, High reliability, Medium relevance]
Roytburg & Miller (2025, Carnegie Mellon / Emory) found 83.1% homophily in AI safety vs. ethics research communities, with only 1% of authors bridging the divide. This structural isolation helps explain why AI safety terminology (sycophancy) has not diffused into governance and regulatory vocabulary. [REPORTED — not independently verified from full text]
JUDGMENT: The sycophancy vocabulary gap persists because it falls into a category gap: it is not a high-level governance concept (like fairness or transparency), not a technical security threat (like prompt injection or adversarial attacks), and not a process term (like evaluation or testing). It is a behavioral model property that affects safety — a category that existing taxonomy efforts have not yet addressed.

Evidence Base Summary¶

Source	Description	Reliability	Relevance	Key Finding
SRC01	Trilateral terminology gap analysis	Medium-High	High	Gap recognized; proposed solutions exclude sycophancy
SRC02	CSIRO/UNSW evaluation framework	Medium-High	Medium-High	Process terminology harmonized; behavioral risks excluded
SRC03	Huwyler threat taxonomy	Medium-High	Medium	53-threat taxonomy excludes sycophancy
SRC04	IAPP governance glossary	High	Medium	100+ terms; sycophancy absent

Collection Synthesis¶

Dimension	Assessment
Evidence quality	Medium — mix of academic, professional, and consultancy sources
Source agreement	High — all sources confirm both the broader gap recognition and the sycophancy-specific omission
Source independence	High — Australian government agency, EU consultancy, US professional association, academic researchers
Outliers	None

Detail¶

The most significant finding is that sycophancy falls into a taxonomy blind spot. Current cross-domain AI terminology efforts focus on three layers: (1) governance concepts (fairness, transparency), (2) technical security threats (adversarial attacks, prompt injection), and (3) process terms (evaluation, testing). Behavioral model properties — how the model behaves in interaction with users — are not yet a recognized category in cross-domain taxonomy work.

Additional Finding: Structural Research Isolation¶

The Roytburg & Miller network analysis provides a structural explanation for the vocabulary gap. With 83% homophily between AI safety and ethics communities and only 1% of researchers bridging the divide, the conditions for terminology diffusion are poor. AI safety researchers coined "sycophancy" but lack the institutional connections to diffuse it into governance, regulatory, and domain-specific communities.

Gaps¶

Missing Evidence	Impact on Assessment
Full text of Roytburg & Miller (2025)	Network analysis cited from WebFetch summary; full analysis not verified
Standards body internal discussions	NIST, ISO, IEEE may have internal working groups addressing sycophancy that are not publicly documented
Non-English language taxonomy efforts	EU and Asian taxonomy efforts may include sycophancy under different terms

Researcher Bias Check¶

Declared biases: The researcher is publishing on this topic and has an incentive to present the vocabulary gap as a novel, unsolved problem. The finding that the gap is recognized but sycophancy is not yet addressed could be framed as either (a) an important gap requiring action or (b) a natural prioritization where more fundamental terminology work must come first.

Influence assessment: The assessment documents both the recognition of the broader gap and the specific exclusion of sycophancy from current efforts. The interpretation is descriptive rather than prescriptive.

Cross-References¶

Entity	ID	File
Hypotheses	H1, H2, H3	`hypotheses/`
Sources	SRC01, SRC02, SRC03, SRC04	`sources/`
ACH Matrix	—	ach-matrix.md
Self-Audit	—	self-audit.md