---
title: "The Assistant, Not the Author: How China Wants Government Offices to Use Large Models"
date: 2026-11-18
category: Policy & Governance
site: NeuroAI
canonical: https://neuroai.site/a/na-policy-gov-llm-deployment-guidelines
language: en
---

# The Assistant, Not the Author: How China Wants Government Offices to Use Large Models

> The Guidelines for the Deployment and Application of AI Large Models in Government Affairs, jointly issued by the Cyberspace Administration of China and the National Development and Reform Commission, set a full-process discipline for government use of large models, anchored in an "auxiliary" positioning. It is normative guidance, not law.

A document circulating through China's government offices draws one bright line about artificial intelligence: the model may help write the report, but it may never be the report.

The Guidelines for the Deployment and Application of AI Large Models in Government Affairs (政务领域人工智能大模型部署应用指引), jointly issued by the Cyberspace Administration of China (CAC) and the National Development and Reform Commission, is one of the most explicit attempts yet by Beijing to answer a question every government is now facing: when a civil servant opens a chatbot, what exactly is the machine allowed to do?

## What the Document Is, and Is Not

The guidelines are a normative document (规范性文件), not a law or administrative regulation. They set requirements and boundaries for government bodies deploying large models, but they carry no penalty schedule of their own.

Timing matters for reading coverage of it. Multiple reports date the guidelines' initial issuance to 2025, and a November 11, 2026 roundup in China Economic Times grouped them with the recent wave of sector-specific "AI+" implementation measures across ministries, which is why the document resurfaced in the news cycle this month. The guidelines predate that roundup; November 2026 coverage recycled their requirements, it did not create new ones.

Their stated purpose is to advance the deployment, application and continuous optimization of large-model technology, products and services in government affairs, in an orderly rather than a rush-everywhere manner.

## The Auxiliary Rule

The guidelines' central discipline is what Chinese commentary calls the "auxiliary" positioning: the large model is a tool for staff, never a replacement for them.

That translates into specific prohibitions. AI must not replace administrative personnel in making administrative decisions, imposing administrative penalties or granting administrative permits. It must not directly issue official documents, public notices or policy answers to the public. And critically, anything a model produces has no official effect: every output must go through substantive human review, verification and revision before it can circulate, be filed or be released.

A provincial party-media commentary in November 2026 quoted the guidelines' own scope for writing work: a model may provide writing suggestions, assist in drafting documents, and check, proofread and optimize format and content, to raise efficiency and reduce grassroots workload. The commentary's summary of the boundary: the AI is a helper, not a ghostwriter.

## The Security Red Lines

Two further rules address risk. Secrets stay out: state secrets, work secrets and unredacted sensitive government data must not be entered into non-classified models. And models must be compliant: deployment is expected to use models that have completed the required registration and filing procedures under China's generative-AI rules, with security assessments part of the process.

The guidelines also name the failure modes they are designed to prevent: model hallucination, data leakage, algorithmic bias and, notably, digital formalism, the risk that ministries buy systems to look modern rather than to work better.

## The Full-Process Discipline

Beyond red lines, the document structures the entire lifecycle: how models are deployed, how they are applied, and how they are continuously optimized.

Three institutional mechanisms give it teeth over time. Government bodies are directed to build a full-process monitoring and evaluation system for government large-model deployment and application, with continuous iteration. CAC and the NDRC are directed to accelerate a national standards system for government large models and to develop key standards. And agencies are told to run training so staff can actually use the systems, alongside broader public digital-literacy work.

The drafting-to-delivery chain also embeds accountability: whoever drafts is responsible, whoever reviews is accountable. The machine's output inherits a human signature.

## Why the Formalism Worry Is the Interesting Part

The most revealing public response came not from technologists but from a commentary on official writing itself. A Xinhua-affiliated provincial commentary in November 2026 warned against "AI new formula prose": bland, templated, empty documents that models produce when officials use them to satisfy material-writing quotas. Its point cuts both ways: if assessment rewards good-looking documents, officials will use AI to mass-produce them, and bureaucracy will have industrialized its own empty rituals.

The commentary's remedy mirrors the guidelines: secrets never enter public platforms, AI output undergoes substantive human revision, and accountability follows the human signature. It ends with a question worth keeping: if AI saves officials time on paperwork, does the saved time go into fieldwork, or into generating more paperwork?

## Honest limitations

- This article relies on state media summaries and published commentary on the guidelines; the full official text was not independently reviewed, and clause-level citations are therefore not possible here.

- The document is normative guidance without disclosed enforcement mechanisms; actual compliance varies by locality and no systematic audit results are public.

- The precise initial issuance date differs across reports, with most dating it to 2025; this article states the year rather than a contested month.

## What readers can do now

- Compare these guidelines with your own jurisdiction's rules on government AI use; the "auxiliary" positioning and human-signature accountability are the two design choices most worth borrowing or challenging.

- Watch for the national standards on government large models that CAC and the NDRC are directed to develop; they will define conformance testing for vendors selling into Chinese government IT.

- If you supply AI systems to any government, note that monitoring-and-evaluation obligations are becoming part of the product, not an afterthought.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-gov-llm-deployment-guidelines
Free to quote with attribution and a link to the original.
