China's new AI guidelines for government: what to do next
China has released guidelines to accelerate practical AI use in government work. The goal is clear: improve service delivery, decision support and internal efficiency without adding process for the sake of process.
The directives come from the Cyberspace Administration of China and the National Development and Reform Commission. They set expectations for planning, deployment, safety, and ongoing optimization of large-model applications across agencies.
What this means for your agency
- Adopt a scenario-first approach, starting with common, high-frequency tasks.
- Build once, reuse often: coordinate models, compute and data across levels of government.
- Run AI as an auxiliary tool with clear boundaries, not a replacement for responsibility.
- Institutionalize safety, confidentiality and lifecycle management from day one.
- Measure outcomes, iterate fast, and raise staff proficiency through training.
Prioritize high-value scenarios
Focus pilots where demand is constant: public services, social governance, office operations and decision support. Select typical scenarios that fit local conditions and your current tech stack.
Standardize how these scenarios are deployed to reduce duplication, speed approvals and simplify oversight.
Coordinate infrastructure and models
Follow provincial frameworks at the prefectural level. At the county level, reuse computing power and model resources from higher-level governments instead of building new silos.
This reduces cost, improves consistency and simplifies security review.
Strengthen data governance
Improve data quality, lineage and access control. Better data yields better model performance and safer outputs.
Prioritize shared data catalogs, de-duplication, and clear rules for model training datasets.
Operate with discipline
Balance workload reduction with meaningful outcomes. Avoid "digital formalism" - work that looks advanced but adds no value to the public or to policy execution.
Assign ownership for process change, not just tool rollout.
Lifecycle management and risk controls
Define the full lifecycle: design, approval, deployment, monitoring, updates and decommissioning. Set the application boundaries and keep AI in an assistive role where appropriate.
Mitigate model hallucinations and other failures with human-in-the-loop checks, escalation paths and clear fallback procedures.
Safety and confidentiality
Establish a safety responsibility system that assigns duties, tasks and escalation for AI-related risks. Test for security issues before and after deployment.
Enforce confidentiality rules: do not input State secrets, work secrets or sensitive data into nonclassified models. Reduce leak risks from data aggregation and cross-association.
Standards and interoperability
Accelerate a national standards system for government-facing large models. Align on key standards that enable sharing, auditing and portability.
Document and promote typical scenarios and proven practices to shorten time-to-value across agencies.
Monitoring and evaluation
Track the entire process end to end. Define metrics for accuracy, service quality, response time, cost, security incidents and user satisfaction.
Use results to drive iterative optimization, not one-off launches.
Training and public literacy
Upgrade staff capabilities through structured training and hands-on practice. Build playbooks, prompt libraries and decision guidelines for frontline use.
Support broader digital literacy efforts so residents can use AI-enabled services with confidence.
Immediate next steps
- Map 3-5 high-frequency scenarios and draft standard operating procedures for each.
- Inventory existing compute, models and datasets; plan reuse with higher-level resources.
- Stand up a cross-functional AI working group (business, IT, security, legal).
- Publish lifecycle and safety policies, including confidentiality and escalation paths.
- Launch a small pilot, define KPIs, and run a 60-90 day evaluation cycle.
- Roll out role-based training and track adoption.
For reference on the issuing bodies, see the Cyberspace Administration of China and the National Development and Reform Commission.
If you're planning training programs for staff, explore role-based options here: Complete AI Training - Courses by Job.
The message is straightforward: start with clear scenarios, share infrastructure, secure the data, manage risk, measure outcomes and train your people. Do that, and AI will improve government work where it counts.
Your membership also unlocks:
