In October 2025 the Cyberspace Administration of China and the National Development and Reform Commission jointly issued guidance on deploying AI large models in government affairs, implementing the “AI plus” initiative and providing direction for government bodies at all levels.
Assistive positioning as the first principle
The central formulation is that government large models are “assistive”. This means model output does not directly constitute an administrative decision, and decision responsibility remains with humans.
In practice this means two things: key judgments in administrative processes cannot be replaced by a model, which can only provide reference; and where model output is adopted, the adopter bears the corresponding judgment responsibility. For vendors supplying government AI, product design must retain human review rather than pursuing full automation.
Three explicit risk lines
Guarding against model hallucination. Government information demands high accuracy; fabricated content can directly mislead decisions or public statements. The guidance requires measures against hallucination, meaning vendors need to provide verifiable, traceable capability.
Strict confidentiality. The guidance expressly requires preventing classified information from being input into non-classified large models. This is the most rigid requirement for government AI — the security level of the data environment must match the classification of the information.
Preventing “digital formalism”. This targets the tendency to treat AI adoption as a showcase project while ignoring actual results, and stresses that applications should be oriented to solving real problems.
Connection to existing regimes
The guidance is not isolated; it works with the technical security specification for government large models: the guidance sets positioning and direction, the technical specification sets security requirements and test methods. See our note on the government large model security specification.
Confidentiality requirements also stack with duties under the Cybersecurity Law and the Data Security Law. See our notes on the AI clause in the revised Cybersecurity Law and network data security risk assessment.
The guidance also notes that counties and below should in principle reuse higher-level compute and model resources, controlling duplicated build cost and unifying security management. See our note on the national computing network platform.
What vendors must actually do
First, build human review and logging into the product, to match assistive positioning and traceable responsibility.
Second, design deployment by classification, using different models and environments for classified and non-classified contexts.
Third, provide evidence of traceability, including knowledge sources, retrieval basis and output basis, to address hallucination risk.
Scaling government AI requires positioning, confidentiality and responsibility to be designed clearly from the start. See our note on the State Council’s “AI plus” deployment.
This article is based on the Guidance on Deploying AI Large Models in Government Affairs and on public information. It is for general reference only and does not constitute legal advice.