---
type: "synthesis"
tags: ["synthesis", "governance", "tension", "trust"]
sources: ["execution"]
id: "cross-governance-vs-psychological-safety"
sourceVaultSlug: "hbr-seg-execution"
originDay: 8
articleStem: "hbr-seg-execution"
sourceUrl: "(unified vault: 7 sources)"
sourceTitle: "HBR — Firm Ⅱ-C · Execution quality — correct execution of AI"
---
## Structural control vs. social contract

The corpus houses two rival theories of controlling AI in the enterprise:

- **The structural/governance camp (A054).** Since [[claim-policing-ai-impossible|you can't police AI]], engineer the environment: [[action-track-provenance|track provenance]], [[action-restrict-unstructured-inputs|restrict to structured inputs]], control model roles. Ethics-by-design also shows up as A060's [[concept-ethical-stewardship]] and Moody's [[action-integrate-risk-and-compliance]].
- **The trust/psychological-safety camp (A076).** [[claim-governance-targets-wrong-problem]] — neither an AI policy nor approved tools predict whether employees hide AI. Worse, [[contrarian-governance-increases-hiding]]: in low-trust settings, sanctioned logging *increases* hiding. The fix is the [[framework-leadership-commitments-for-disclosure|five leadership commitments]].

## The reconciliation

They solve different problems. Governance solves **security, compliance, and content integrity** (A054's real worry is process pollution, not surveillance). Trust solves **disclosure and diffusion** (A076's worry is that value stays hidden). The honest synthesis both articles' enrichment layers reach: **governance is necessary but insufficient.** A054 restricts *inputs* to defuse an arms race; A076 warns that restricting *people* backfires. The unresolved paradox is A076's [[question-sanctioned-tool-extraction]] — the same logging that credits a discoverer enables their replacement — which A054's provenance tracking would only intensify. See [[cross-shadow-ai-fuels-decay]] for where these two crash into each other.