A Nine-Principle Framework, Voluntarily Mandatory
On July 1, 2026, China's National Cybersecurity Standardization Technical Committee (TC260) put into force its Ethics-Safety Guidelines for Artificial Intelligence Applications 1.0 — a voluntary but politically significant document adopted six weeks ago on May 19. The guidelines set out nine core ethical principles that generative AI developers, service providers, and end users are expected to embed into their systems: enhancing human welfare, respect for life and dignity, fairness, transparency, privacy protection, safety, controllability and trustworthiness, inclusivity, and agile co-governance.
The guidelines are framed as a principled, reference-oriented technical document — not a statute. TC260 technical documents do not carry the force of binding legislation on their own. Yet in China's regulatory ecosystem, that distinction matters less than it first appears. Voluntary guidance from TC260 has historically served as the soft-law prelude to mandatory standards and enforcement expectations. The 2024 Basic Requirements for GenAI Safety, itself a TC260 product, quickly shaped how the Cyberspace Administration of China (CAC) approached large-language-model oversight. Today's guidelines are likely to follow the same trajectory.
Three Obligations Across the Supply Chain
The framework maps its nine principles to obligations across three stakeholder groups:
- Developers must integrate ethical and safety considerations throughout the entire AI lifecycle — data selection, algorithm design, testing, deployment, and ongoing maintenance — with specific attention to managing hallucination risks and protecting vulnerable populations including minors.
- Service providers are expected to maintain human oversight mechanisms in high-risk deployment scenarios, establish pathways for users to reject or terminate AI services, minimise personal data collection, and build emergency response capabilities for safety incidents.
- Users are expected to improve AI literacy, avoid over-reliance on AI systems, use services lawfully and responsibly, and refrain from disclosing sensitive personal or commercial information to AI systems.
Each principle maps to concrete technical controls in accompanying guidance: transparency demands traceable logs and explainability reports; fairness requires bias testing during model review; controllability implies human-in-the-loop checkpoints for consequential decisions.
Context: China's Layered AI Governance Since 2021
Today's guidelines land on top of an already-substantial regulatory stack. China issued its Algorithm Recommendation Measures in December 2021, its Deep Synthesis Measures covering deepfakes and synthetic media in November 2022, and the landmark Interim Measures for the Management of Generative AI Services in July 2023 — widely regarded as the world's first binding generative AI regulation. That 2023 framework, jointly issued by the CAC and six co-regulators, requires providers to act as content gatekeepers, prohibit outputs that violate "core socialist values," conduct algorithm filing and security assessments, and cooperate with regulatory inspections by disclosing training data provenance and model mechanisms.
The TC260 guidelines do not replace that architecture. They sit atop it, adding an ethical-principles layer and operationalising the soft-law consensus about what responsible AI development looks like before harder rules arrive.
What Arrives July 15
Two binding instruments — with actual penalties — follow in two weeks. The Interim Measures for the Administration of Anthropomorphic AI Interaction Services, jointly issued by the CAC, NDRC, MIIT, the Ministry of Public Security, and SAMR, regulate AI chatbots, virtual companions, and emotional interaction services. They mandate algorithm filing, security assessments, clear disclosure of AI status, and addiction-prevention mechanisms. They ban virtual intimate relationship services for minors under 18 and require parental consent for children under 14 to access any anthropomorphic AI service. Violations carry fines reaching 200,000 RMB (approximately US$27,500) for cases involving harm to users.
Also on July 15, the CAC, NDRC, and MIIT jointly activate Implementation Opinions on the Standardised Application and Innovative Development of Intelligent Agents — China's first coherent framework for autonomous AI agents. Healthcare, transportation, media, and public safety deployments face mandatory filing and compliance testing. The opinions also set a target of 70% intelligent agent adoption in smart terminals by 2027, signalling that this regulation is designed to accelerate as much as to constrain.
Steelmanning the Regulatory Wave
The case for this month's regulatory activity is not trivial. AI emotional manipulation of minors, autonomous agent failures in high-stakes sectors, and generative outputs that spread disinformation are genuine risks — not regulatory fictions. Human oversight checkpoints in consequential applications, transparency mandates, and meaningful user control mechanisms align broadly with responsible AI principles articulated by the OECD and embedded in the EU AI Act. The TC260 framework's nine principles are recognisably universal in form.
Nor is China alone in reaching for supply-chain governance. The EU AI Act imposes layered obligations on providers, deployers, and users by risk tier. What makes China's approach distinctive is not its layered structure but its speed of iteration and its content.
Where Proportionality Breaks Down
China's 2023 GenAI Measures' requirement that AI outputs align with "core socialist values" — a political concept without a neutral technical definition — blurs the boundary between content safety and content control. Compliance requires providers not merely to prevent harm but to enforce a specific ideological orientation at inference time. For global companies serving Chinese users, that demands algorithmic accommodation of political speech constraints. For domestic providers, it entrenches self-censorship as a baseline design requirement, not an edge-case exception.
The TC260 guidelines are less politically explicit, staying at the level of principles. But they sit within a regulatory stack that operationalises "human welfare" partly as state-defined value alignment. Advisory frameworks from TC260 have a track record of hardening into binding standards faster than their initial framing suggests. Developers who treat today's guidelines as merely optional do so at their own risk.
What to Watch
China's national AI law remains in the legislative pipeline, expected to consolidate and supersede much of the current patchwork. The July 2026 wave — TC260 ethics guidelines, Anthropomorphic AI measures, and Intelligent Agents opinions — likely serves two purposes: filling near-term governance gaps before that legislation arrives, and signalling to domestic and foreign developers what the compliance baseline will look like once it does.
For developers and providers operating in China, the operational imperative is to map existing practices against the nine TC260 principles now, before the advisory posture hardens. For policymakers elsewhere watching China's approach, the July 2026 wave offers an instructive case study in layered AI governance — one that moves faster than democratic legislatures typically can and with less procedural transparency at each step.