A civil servant asks an AI to draft a policy document; the model reaches for a news article instead of the statute book, and a confident but fabricated citation lands in a memo headed to a mayor's desk. As government agencies across China rush to bolt large models onto approvals, reception desks, and confidential workflows, the question stopped being "can we" and became "who checks the model, the data, and the audit trail."
A new national standard answers that question with a baseline. In late September 2026, China published its first recommended national standard (推荐性国家标准) written specifically for government-affairs large models — and it is less about capability than about trust.
What the standard is — and isn't
The document is GB/T 48324—2026《人工智能 政务大模型系统技术要求》 (Technical Requirements for Government-Affairs Large Model Systems). It was issued by the State Administration for Market Regulation (国家市场监督管理总局) and the Standardization Administration (国家标准化管理委员会, SAC), published on 28 September 2026, and is scheduled to take effect on 1 April 2027 (listed as 即将实施, "about to be implemented").
Crucially, it is a recommended standard, not a mandatory one. That shapes everything below: agencies are expected to converge on it through procurement and review, not because a hard legal switch flips. Drafting participants include PeopleAI (人民中科) as a main drafting unit and Kingsoft Office / WPS (金山办公), which co-wrote the agency-office application module.
The standard organizes its requirements around four pillars:
- Model capability (模型能力)
- Scenario application (场景应用)
- Security compliance (安全合规)
- Deployment and operations (部署运维)
It applies to the design, development, and deployment of 政务大模型 (government large models) and can also guide system evaluation.
Why government models need a different rulebook
A government model is not a consumer chatbot. The sensitivity of the data, the finality of some decisions, and the public-accountability angle all raise the bar above a general-purpose assistant. The standard bakes in principles that vendors have already been marketing as differentiators:
- Data must not leave the domain (数据不出域) — sensitive content stays inside the government network.
- Operations must be authorized, with high-risk actions requiring re-confirmation.
- Human-in-the-loop stays on the critical path; the model drafts, a person decides.
Real deployments show the pull. Per Kingsoft Office and China Daily, Zhongshan city has equipped about 34,000 civil servants with an AI drafting assistant built on a knowledge base of 600,000 government documents, cutting drafting time from roughly three hours to 40 minutes and claiming a 95% drop in content errors. In Guangdong, a WPS private-deployment edition mirrors each agency's org structure and permission system, constraining document access, edits, and AI calls with a full audit trail. These are vendor-stated figures, not independent audits — useful as illustration, not proof.
What the technical requirements actually demand
Pulling the four pillars into practice, the standard asks for:
- Reliability and grounding in model capability — hallucination control and knowledge faithfulness, not just benchmark scores.
- Clear scenario scope in application — defined tasks, with humans owning the key links.
- Data lifecycle security in compliance — covering personal information and important/classified data handling.
- Isolated, auditable deployment in operations — private or sandboxed environments, permission by user role (人机同权), and full-chain logging.
The through-line is that a government model is judged on whether a supervisor can trace every output back to a source, a permission, and a person.
The wider standards cluster
GB/T 48324—2026 did not arrive alone. Two companion guidance documents (指导性技术文件, GB/Z) round out the picture:
- GB/Z 298—2026《数字政府 智能客服系统建设指南》 — released 24 September 2026, currently in force; tells agencies how to build smart customer-service systems (the 12345 hotline, service halls, online portals).
- GB/Z 244—2026《人工智能 办公大模型系统建设指南》 — released 27 August 2026, currently in force; guidance on building office large models.
Together they answer three questions: how to build a gov smart hotline, what level a gov large model must reach, and how to build an office model. The context is a market that has scaled fast — China's AI-related output passed ¥700 billion (≈ US$97 billion / HK$760 billion) in 2024, per the China Internet Network Information Center.
Honest limitations
- GB/T 48324—2026 is 推荐性 (recommended), not 强制性 (mandatory); uptake will come through procurement and review cycles, not a binding legal mandate.
- It is 正式通过 (adopted) on 28 September 2026 but reaches 生效实施 only on 1 April 2027 — it is not yet in force.
- The deployment statistics (34,000 users, 40-minute drafting, 95% fewer errors) come from the vendor and a China Daily report, not an independent audit; treat them as illustrative of direction, not measured outcome.
- The standard sets a technical floor, not a certification regime. How local governments write it into tenders and enforcement will vary, and the companion GB/Z guides are guidance, not law.
What readers can do now
- If you sell AI to government in China, align your government-large-model product to the four pillars (capability / scenario / security / operations) before April 2027 — tenders will start quoting them.
- If you run a public-sector AI project, require private or isolated deployment and a full-chain audit as a baseline regardless of model size.
- Read the companion guidance docs (GB/Z 298, GB/Z 244) — they increasingly shape how procurement language is written, even where the main standard is recommended.