Skip to content

R0043/2026-03-28

Research R0043 — Sycophancy Vocabulary
Mode Query
Run date 2026-03-28
Queries 3
Prompt Unified Research Standard v1.0-draft
Model Claude Opus 4.6

Three queries investigated the cross-domain vocabulary for the phenomenon AI safety researchers call "sycophancy," the regulatory requirements that address it, and whether the vocabulary gap has been recognized in the literature.

Queries

Q001 — Cross-Domain Vocabulary Map — Partial vocabulary with systematic gaps

Query: What terms do different industries and disciplines use to describe AI behavior that prioritizes user agreement, comfort, or satisfaction over accuracy, correctness, or safety?

Answer: The vocabulary is systematically asymmetric. Regulated industries have mature terminology for the human side (automation bias, complacency, overtrust) while AI safety alone has terminology for the system side (sycophancy). This divide reflects when each domain's vocabulary was developed: traditional automation era vs. adaptive AI era.

Hypothesis Status Probability
H1: Rich cross-domain vocabulary Partially supported
H2: No cross-domain vocabulary Eliminated Remote (< 5%)
H3: Partial with systematic gaps Supported Very likely (80-95%)

Sources: 10 | Searches: 6

Full analysis

Q002 — Enterprise Requirements — Indirect requirements only

Query: Search for enterprise requirements, procurement specifications, regulatory guidance, or deployment standards that address the sycophancy phenomenon under its domain-specific names.

Answer: Requirements exist but exclusively through indirect means: human oversight mandates (EU AI Act), automation bias awareness (FDA), general trustworthiness criteria (DoD), and voluntary risk frameworks (NIST). No regulation directly constrains sycophantic system behavior. The vocabulary gap produces a requirements gap.

Hypothesis Status Probability
H1: Substantial direct requirements Eliminated Remote (< 5%)
H2: No requirements exist Eliminated Remote (< 5%)
H3: Indirect requirements only Supported Very likely (80-95%)

Sources: 6 | Searches: 3

Full analysis

Q003 — Vocabulary Gap Literature — Recognized broadly but not for sycophancy

Query: Has the vocabulary gap itself been identified as a problem in the AI safety or AI governance literature?

Answer: The broader AI terminology gap is well-recognized, with multiple active bridging efforts (MIT AI Risk Repository, AIR 2024, Standardized Threat Taxonomy). However, the specific sycophancy/overreliance vocabulary gap has not been articulated as a distinct problem. Sycophancy is absent from every major bridging taxonomy examined.

Hypothesis Status Probability
H1: Gap recognized and addressed Partially supported
H2: Gap not recognized Eliminated Remote (< 5%)
H3: Recognized but not for sycophancy Supported Likely (55-80%)

Sources: 5 | Searches: 2

Full analysis


Collection Analysis

Cross-Cutting Patterns

Pattern Queries Affected Significance
Human-side/system-side vocabulary divide Q001, Q002 The central finding: regulated industries frame the problem as human cognitive failure (automation bias, overreliance); AI safety frames it as system behavioral failure (sycophancy). This framing difference drives both vocabulary gaps and requirements gaps.
Vocabulary determines requirements Q001, Q002 Regulators who use human-side terms write human-side requirements. The EU AI Act's choice of "automation bias" (not "sycophancy") produced a deployer-awareness obligation (not a system-design constraint).
Bridging efforts miss sycophancy Q001, Q003 Every major taxonomy bridging effort examined (MIT, AIR 2024, Standardized Threat Taxonomy) omits sycophancy as a distinct category. The phenomenon falls between technical-threat taxonomies and human-factors vocabularies.
Traditional automation vs. adaptive AI divide Q001, Q003 Existing vocabulary was developed for deterministic automation (autopilots, CDSS). AI that actively adapts output to please users is qualitatively different, and vocabulary has not caught up.

Collection Statistics

Metric Value
Queries investigated 3
All answered with nuanced/conditional hypothesis (H3) 3 (Q001, Q002, Q003)
All H2 (negative) hypotheses eliminated 3
All H1 (affirmative) hypotheses partially supported or eliminated 3

Source Independence Assessment

The 21 sources across all three queries are highly independent. They span: - Jurisdictions: EU (AI Act), U.S. (NIST, FDA, DoD, FAA), cross-jurisdictional (AIR 2024) - Disciplines: AI safety, human factors engineering, healthcare informatics, defense policy, law, philosophy - Source types: Legislation, government standards, peer-reviewed journals, policy research, preprints - Time range: 2010 (Parasuraman & Manzey foundational model) through 2026 (FDA CDS guidance update)

No evidence of citation clustering: the sources do not predominantly cite each other. The AI safety sources (Anthropic, DeepMind, OpenAI) are independent from the regulated-industry sources (EU, DoD, FAA, FDA). The convergence on the human-side/system-side finding across independent sources strengthens confidence.

Collection Gaps

Gap Impact Mitigation
Proprietary procurement documents May contain more specific anti-sycophancy requirements than public regulations Gap acknowledged; public requirements serve as lower bound
Non-English regulatory terminology EU member states, Asian regulators may use different terms Gap acknowledged; focus on English-language literature is a limitation
Classified/FOUO defense documents DoD testing criteria may address sycophancy more specifically Gap acknowledged; CaTE publications suggest calibrated trust focus
ISO/IEC 42001 full standard text Full standard may contain sycophancy-relevant provisions Gap acknowledged; publicly available summaries suggest no specific provision
Professional society standardization IEEE, ACM, HL7, ARINC working groups may have terminology efforts Gap acknowledged; web search may not surface these

Collection Self-Audit

Domain Rating Notes
Eligibility criteria Low risk Consistent criteria across all 3 queries: published terminology, formal requirements, and bridging efforts
Search comprehensiveness Some concerns 11 searches, 110+ results dispositioned. Financial services and enterprise evaluation searches were thinnest. Professional society efforts may be under-covered.
Evaluation consistency Low risk Same scoring framework (GRADE + Cochrane-adapted bias domains) applied across all 21 sources
Synthesis fairness Low risk H3 (nuanced) prevailed in all 3 queries, which is consistent with the evidence but could reflect a predisposition toward "it depends" answers. Mitigated by the fact that H1 and H2 received fair hearing with evidence explicitly mapped to each.

Resources

Summary

Metric Value
Queries investigated 3
Files produced 96
Sources scored 21
Evidence extracts 21
Results dispositioned 16 selected + 94 rejected = 110 total
Duration (wall clock) 23m 43s
Tool uses (total) 89

Tool Breakdown

Tool Uses Purpose
WebSearch 15 Search queries across domains
WebFetch 8 Page content retrieval for key sources
Write 48 File creation
Read 4 Methodology and format document reading
Edit 0 No file modifications
Bash 12 Directory creation, batch file generation

Token Distribution

Category Tokens
Input (context) ~180,000
Output (generation) ~45,000
Total ~225,000