Filed under ai-governance
Every brief tagged "ai-governance".
- § 07 · AI-GOVERNANCE
Prompt Stacks and Prompt Governance — Why System-Level Prompts Are Emerging as a Regulatory Lever (and Where They Fall Short)
A Chinese AI-law reading of Neumann, Sargeant and Singh's FAccT 2026 paper Prompt Governance? — and what it means for how China, the EU, and the US treat 'system prompts' as a regulatory object. Li Wenlong (科技利维坦) walks through the four-layer 'prompt stack' (system instructions → system guidelines → developer instructions → user prompts), five properties practitioners need to understand (layered, hidden, natural-language, malleable, loosely coupled to behaviour), and the comparative regulatory landscape: the EU GPAI Code of Practice requires signatories to disclose system prompts to regulators in model reports; the Trump EO 14319 / OMB M-26-04 stops at model / system / data cards and leaves system-prompt disclosure voluntary; the UK's AI Cybersecurity Code says effectively nothing. China's current GenAI safety regime (TC260-003 plus the GenAI Interim Measures) is output-evaluation-based — filing and pre-launch scoring, with no architectural hook into system prompts. Li predicts a Brussels Effect: system-prompt disclosure to regulators will become a global compliance baseline, analogous to the DPIA in data law. For overseas counsel: this is what is coming, what to start archiving now, and why 'what you write' in a system prompt is not 'what the model executes.'
- § 08 · AI-GOVERNANCE
Zhu Xiaofeng — Who Pays When GenAI Causation Is Unclear? Applying Civil Code Article 1254 by Analogy
Zhu Xiaofeng (Central University of Finance and Economics Law School) takes on the GenAI causation black hole — when a personal-information harm clearly arises from a GenAI service but specific causation among model designer, model provider, model user, and data provider cannot be established, who pays? Zhu's structural answer: when conventional construction-element-analysis and Article 998 interest-balancing both fail (and they do), apply Civil Code Article 1254's 'unclear-causation' rule by analogy — the same rule used for falling-object-from-building cases. The doctrinal scaffolding: communication-safety theory, gain-and-risk allocation theory, causation proof + harm prevention. Critically: each potential injurer compensates the full damage; among themselves, allocation is proportional, with judges determining specific amounts case-by-case. Highly relevant for multinationals deploying GenAI in China — the proposed framework restructures the operating liability surface.
- § 09 · AI-AGENTS
Mapping the AI Agent Risk Surface — A Ten-Category Taxonomy Under China's New 智能体新规
China's Cyberspace Administration jointly issued the Implementation Opinions on Standardized Application and Innovation Development of AI Agents (the '智能体新规' or 'Agent Rules') on May 8, 2026 — the first dedicated regulatory document on AI agents anywhere in the world. This DCC brief works through the ten-category risk taxonomy that practitioners are now using to map the agent attack surface: goal hijacking, tool misuse, identity/permission abuse, supply-chain compromise, unintended code execution, memory and context poisoning, inter-agent communication insecurity, cascade failures, human-machine trust exploitation, and rogue agents. With the agent risk mapped, the brief works the legal-liability vector: how each risk maps to administrative, civil, and criminal exposure under existing PIPL, CSL, Anti-Unfair Competition, and trade-secret regimes. Closes with the Guangzhou Internet Court's recent dual-authorization ruling against an open-source agent that bypassed a chat platform's risk controls — the first Chinese case to articulate the dual-authorization principle for AI agents accessing third-party platforms.
- § 10 · AI-AGENTS
Operationalizing AI Agent Governance — A Ten-Step Internal Control Framework
Part 2 of DCC's brief on the Chinese Agent Rules (《智能体规范应用与创新发展实施意见》, May 2026). After mapping the ten-category risk taxonomy in Part 1, this brief works through the ten-step internal governance framework practitioners are now building to operationalize agent compliance: cross-functional governance organization + agent asset inventory; use-case admission and classification (L1 read-only / L2 limited-write / L3 sensitive-data / L4 high-impact); security assessment and AI red-team testing; identity authorization and permission control (with the under-discussed 'permission inheritance' trap); data protection; tool and protocol security; human-in-the-loop design; supply-chain security; continuous monitoring; and AI-specific incident response. Closes with five operational priorities for teams that need to start now without waiting for the 'big-and-comprehensive' regime build.
- § 11 · AI-GOVERNANCE
Open-Source Does Not Mean Open Data — Zhang Ping on Training-Data Compliance for Open-Source AI
Peking University Law School professor Zhang Ping, writing in 人民论坛 (People's Tribune), takes apart two misconceptions that have dominated the Chinese open-source AI discussion: that 'open source' means training data has no copyright protection, and that 'algorithm open-source' compels 'training data publication.' Both false. Zhang lays out the structural distinction: 'open source is conditional authorization under license' — applied to model weights, not to the training corpus, which is a legally independent object. She then maps the full-chain compliance risk (acquisition / processing / output) and proposes a four-tier differentiated governance framework that finance, healthcare, and government AI deployments can actually use to map their training-data inventory against compliance gates.
- § 12 · AI-GOVERNANCE
Where China's Draft AI Anthropomorphic-Interaction Measures Need Work — A Scholar's Reform Map
Li Wenlong (科技利维坦) walks through the directions in which he would amend China's draft Interim Measures for the Administration of AI Anthropomorphic Interaction Services (人工智能拟人化互动服务管理办法) — the country's first dedicated rule on 'companion'-style AI. His critique is structural, not cosmetic: the core definition of '拟人化 (anthropomorphisation)' is too broad because it anchors on human-like expression rather than the real harm (relational dependency); the invented concept of '交互数据 (interaction data)' should be deleted and folded back into PIPL rather than blanket-prohibited; Chapter 2 mixes three incompatible duty types and should be split; the '1M registered / 100k MAU' security-assessment trigger is borrowed from other regimes and does not track real risk; and the training-data duties are horizontal obligations misplaced in a vertical rule. For overseas counsel building companion-AI or emotional-AI products for the China market: this is a map of where the draft is likely to move, and which duties fall on deployers versus base-model providers.