Skip to content

R0043/2026-03-28/Q002 — Assessment

BLUF

Regulated industries have requirements that address the sycophancy phenomenon, but exclusively through indirect means: human oversight mandates, automation bias awareness obligations, and general trustworthiness criteria. No regulation, standard, or procurement specification found directly requires AI systems to avoid producing agreeable-but-wrong output. The regulatory gap mirrors the vocabulary gap identified in Q001 — because regulators frame the problem as human overreliance (not system sycophancy), they write human-side solutions (train the operator) rather than system-side solutions (constrain the model).

Probability

Rating: H3 (indirect requirements) is Very likely (80-95%)

Confidence in assessment: High

Confidence rationale: Evidence from 6 authoritative sources (EU legislation, FDA guidance, DoD principles, NIST framework, cross-jurisdictional taxonomy, financial services framework) consistently shows indirect-only approach. No counter-evidence found.

Reasoning Chain

  1. The EU AI Act Article 14 is the most explicit regulatory provision: it names "automation bias" and requires deployer awareness and override capability. But the obligation is on deployers, not providers, and mandates awareness, not system design constraints [SRC01-E01, High reliability, High relevance].

  2. FDA's 2026 CDS guidance explicitly uses "automation bias" and requires independent review capability. However, it treats this as a transparency obligation and cites a 2004 paper — 22 years behind current sycophancy research [SRC02-E01, Medium-High reliability, High relevance].

  3. DoD Responsible AI tenets (Responsible, Equitable, Traceable, Reliable, Governable) are used as procurement evaluation criteria. "Governable" requires human intervention and control mechanisms. None specifically address sycophancy-adjacent system behavior [SRC03-E01, High reliability, Medium-High relevance].

  4. NIST AI RMF identifies "overreliance" as a risk category with mitigation guidance but is a voluntary framework, not a binding standard [SRC04-E01, High reliability, Medium-High relevance].

  5. The AIR 2024 cross-jurisdictional taxonomy found that "risks associated with AI overreliance or excessive autonomy are less frequently specified in detail" across corporate policies — confirming the gap extends beyond regulation to industry practice [SRC05-E01, Medium-High reliability, High relevance].

  6. Financial services has 230 control objectives but none specifically targeting AI agreeableness; controls focus on governance, validation, and monitoring [SRC06-E01, Medium-High reliability, Medium-High relevance].

Evidence Base Summary

Source Description Reliability Relevance Key Finding
SRC01 EU AI Act Article 14 High High Automation bias awareness mandate — deployer obligation
SRC02 FDA CDS Guidance Medium-High High Transparency requirement citing 2004 research
SRC03 DoD RAI Principles High Medium-High General trustworthiness tenets in procurement
SRC04 NIST GAI Profile High Medium-High Risk identification without binding requirements
SRC05 AIR 2024 Taxonomy Medium-High High Overreliance risks underspecified in corporate policies
SRC06 FSSCC Framework Medium-High Medium-High 230 controls, none sycophancy-specific

Collection Synthesis

Dimension Assessment
Evidence quality Robust — legislation, government frameworks, cross-jurisdictional analysis
Source agreement High — all sources confirm indirect-only approach
Source independence High — EU, U.S., and cross-jurisdictional sources
Outliers None — remarkable consistency across all jurisdictions and sectors

Detail

The most significant finding is the consistency: across 4 sectors, 3 jurisdictions, and multiple regulatory approaches, every requirement addresses the sycophancy phenomenon from the human side. This is not coincidence — it reflects the vocabulary finding from Q001. Regulators who frame the problem as "automation bias" (human cognitive failure) naturally write requirements for human awareness and override capability. A regulator who framed it as "sycophancy" (system behavioral failure) would naturally write requirements for system design constraints.

Gaps

Missing Evidence Impact on Assessment
Proprietary enterprise procurement RFPs Private-sector requirements may be more specific than public regulatory guidance
OCC/Fed bank examination guidance for AI May contain AI-specific requirements not found in public searches
Aviation-specific AI deployment standards (EASA, FAA) Aviation may have more specific automation requirements than found
Defense-specific AI system acceptance testing criteria Classified or FOUO documents may contain sycophancy-relevant requirements

Researcher Bias Check

Declared biases: No researcher profile provided.

Influence assessment: The query assumes requirements exist "under domain-specific names" — presupposing that Q001's vocabulary would map to requirements. This framing was tested by including H2 (no requirements exist) as a hypothesis. H2 was eliminated, confirming that the framing was warranted but not self-fulfilling.

Cross-References

Entity ID File
Hypotheses H1, H2, H3 hypotheses/
Sources SRC01-SRC06 sources/
ACH Matrix ach-matrix.md
Self-Audit self-audit.md