A procurement team at a state-owned plant sits across from a model vendor who slides a compliance report over. The last page says the system's records are complete and its risks are under control. The buyer asks who wrote it — because a report produced by the system under review is a description, not a verdict.
That question now has three documents behind it.
On 1 November 2026, a batch of Chinese AI standards took effect. Three define the compliance surface a firm actually meets: one national standard on how generative AI providers should manage their social impact, and two telecom-industry standards — one on risk management capability for AI systems, one on the quality of training data for embodied AI (具身智能). None is a law; all three are recommended, not mandatory. Together they do something quieter than a ban: they hand buyers and inspectors a shared vocabulary for asking vendors harder questions.
What actually came into force
GB/T 47863-2026 — Social impact of generative artificial intelligence technology application: Service providers compliance management guideline (《生成式人工智能技术应用社会影响 服务提供者合规管理指南》). Issued 2 July 2026 by the State Administration for Market Regulation and the Standardization Administration, implemented 1 November 2026, status current. A 20-page recommended national standard administered by the working group on social application and assessment of intelligent technology (SAC/SWG 35) and led by the China National Institute of Standardization. Its job is to turn "social impact" from a press-release adjective into a management system a provider can be audited against: leadership, planning, support, operation, evaluation, improvement. It builds on two earlier standards.
YD/T 7173-2026 — Artificial intelligence: Safety governance — System risk management capability requirements (《人工智能 安全治理 系统风险管理能力要求》). Issued 15 July 2026, implemented 1 November 2026, filed under the Ministry of Industry and Information Technology (MIIT). Drafted by the China Academy of Information and Communications Technology (CAICT) with 22 universities, institutes and companies. It splits capability into three blocks — organisational capacity, process management, risk protection — grades risk into four levels (severe, high, medium, low), and names eight dimensions a system must be able to demonstrate: reliability, controllability, security, fairness, transparency, explainability, privacy protection and accountability.
YD/T 6771-2026 — Artificial intelligence: Key foundational technology — Quality requirements and evaluation methods for embodied intelligence datasets (《人工智能 关键基础技术 具身智能数据集质量要求及评价方法》). Issued 15 July 2026, implemented 1 November 2026. Also CAICT-led, with more than 40 organisations. Its drafters describe it as the first industry standard for the data robots learn from; it sets eight quality indicators: completeness, consistency, diversity, authenticity, usability, practicality, scalability and security.
Voluntary, and therefore load-bearing
Legal status deserves a blunt sentence: GB/T means recommended national standard, YD/T means recommended communications-industry standard. Neither carries a fine. A company that ignores all three breaks no rule.
The reason they matter is where they land. Recommended standards feed three practical processes: procurement scoring for state-owned buyers, inspection checklists, and the evidence packs large customers demand before signing. Optional on paper becomes close to compulsory in a tender. That is the mechanism to watch, not the penalty clause — there isn't one.
The clause that changes the meeting
Of the eight dimensions in YD/T 7173, seven map onto things most enterprises already own: disaster recovery, access control, data masking, content filters. Accountability is the awkward one.
As CAICT described the standard in an August 2026 briefing, accountability rests on three things: a model audit mechanism, independent verification of the operating process and its results, and an external oversight channel. Read together, they require logs that another party can recompute. A log kept for internal eyes does not satisfy the test.
Self-attestation keeps its use as a source of evidence and loses its standing as a conclusion. A vendor can still say "we are compliant." The buyer now has a standard that asks a different question — who checked, and were they outside the system?
A yardstick for robot data
YD/T 6771 addresses a bottleneck anyone building humanoids runs into. Real-world manipulation data is expensive, slow and uneven, and the industry's own complaint is that demonstrations look better than deployments. The standard's three pillars — production practice, organisational assurance, quality evaluation — make "is this dataset any good" answerable before training rather than after deployment.
Its eight indicators include two that generic data-quality frameworks do not carry: diversity (coverage of lighting, materials, poses, clutter) and authenticity (whether real contact force, friction and deformation survive, or simulation is dressed as measurement). For anyone selling robot training data into China, those two words are the new acceptance criteria.
Why all three at once
The coincidence is bureaucratic, not conspiratorial. Approvals for this tranche landed in the same July window, and implementation dates cluster at the start of the following quarter. Filing records show the batch was larger: the same release carried YD/T 7171 on inference platforms and YD/T 7172 on large-model dataset management. The same pattern shows up elsewhere: MIIT's notice on agent-capable devices, issued 6 November 2026, points vendors at a separate family of national standards for agent interconnection.
China is not banning capability. It is building the paperwork layer that lets a buyer, an auditor or a regulator ask a capability to prove itself — and, in the accountability clause, insisting the proof come from someone other than the accused.
According to Li An, chief scientist at BrainNet (脑机网), the batch's practical effect is that compliance stops being a document you write and becomes a record you keep — a shift visible in procurement long before enforcement.
What readers can do now
- If you sell AI into China: treat YD/T 7173's eight dimensions as a pre-bid questionnaire. Accountability is the dimension most vendors cannot answer, so it is also the one that differentiates.
- If you buy AI in China: ask for the audit trail, not the compliance report — who produced it, and whether that party sits outside the system.
- If you build robots or sell robot data: map your dataset documentation onto YD/T 6771's eight indicators before a customer does it for you.
Honest limitations
I read the public registry entries, titles, issue and implementation dates, drafting-unit lists and scope statements for all three standards, plus official and state-media summaries. I did not read the full licensed text of any of them, so my description of internal structure relies on published summaries and CAICT's briefing as reported in the Chinese financial press. The claim that accountability requires independent verification comes from that briefing, not from a clause I quoted directly. I did not verify how many procurers use the standards as scoring criteria; that argument is inference.
Key takeaways
- Three AI standards took effect on 1 November 2026: GB/T 47863-2026 on generative-AI social-impact management, YD/T 7173-2026 on AI system risk capability, YD/T 6771-2026 on embodied-AI dataset quality.
- YD/T 7173's accountability dimension asks for a model audit, independent verification and an external oversight channel, downgrading vendor self-attestation from conclusion to evidence.
- YD/T 6771 gives robot training data its first industry-wide quality yardstick, with diversity and authenticity as the two non-standard indicators.