R0043/2026-03-28/Q002/H3¶
Statement¶
Requirements address the sycophancy phenomenon indirectly through human oversight mandates, accuracy requirements, or general AI trustworthiness criteria rather than directly naming or targeting sycophancy-adjacent system behavior.
Status¶
Current: Supported
This is the hypothesis best supported by the evidence. Across all four regulated industries searched (defense, healthcare, finance, aviation), requirements address the phenomenon exclusively through human-side mandates: ensure operators can override AI, ensure operators are aware of automation bias risk, ensure systems are "valid and reliable." No requirement addresses why or how AI systems produce agreeable output in the first place.
Supporting Evidence¶
| Evidence | Summary |
|---|---|
| SRC01-E01 | EU AI Act mandates deployer awareness of automation bias and ability to override — indirect human-oversight approach |
| SRC02-E01 | FDA requires independent review capability as safeguard against automation bias — transparency approach |
| SRC03-E01 | DoD Responsible AI tenets require "Governable" systems with human intervention — general trustworthiness approach |
| SRC04-E01 | NIST identifies overreliance as risk category with mitigation guidance — risk management approach |
| SRC05-E01 | AIR 2024 taxonomy notes "overreliance" risks are "less frequently specified in detail" across corporate policies |
| SRC06-E01 | Financial services 230 control objectives focus on governance, validation, and monitoring — no sycophancy-specific controls |
Contradicting Evidence¶
No evidence contradicts H3.
Reasoning¶
H3 is supported by a consistent pattern across all four sectors: every regulatory framework found addresses the problem through human oversight requirements (the user must be able to override) rather than system design requirements (the system must not agree when wrong). This mirrors Q001's finding about the human-side/system-side vocabulary divide — the requirements reflect the vocabulary. If regulators have only human-side terms, they write human-side requirements.
Relationship to Other Hypotheses¶
H3 incorporates the valid elements of both H1 (requirements exist — just not direct ones) and H2 (direct sycophancy requirements don't exist). The "indirect" qualifier is the key finding.