The National Data Administration's Action Plan for High-Quality Industrial Datasets

On 8 June 2026 the National Data Administration released the Implementation Plan for the Initiative to Promote the Construction of High-Quality Industrial Datasets. This is China’s first systemic national-level deployment on “how data empowers AI development”, issued as document No. 25 of 2026 with a document date of 3 June.

Why datasets have become a focus

Data is the “fuel” of AI development. The ceiling of model capability is largely set by the quality, scale and compliance of training data. In recent years, computing power and algorithms received more attention, while data — especially high-quality, licensable and traceable data that can be used directly for training and development — has long been a weak link.

The plan’s purpose is to fill this gap: serving the needs of AI applications, it continues to advance multimodal high-quality datasets in text, images, audio and video; and focuses on key fields such as scientific research, industrial manufacturing, the low-altitude economy and embodied intelligence to accelerate industrial dataset building.

Six special actions

The plan deploys six special actions: strengthening the base and expanding capacity; attacking annotation; improving quality and efficiency; enabling applications; managing and servicing; and releasing value.

“Strengthening the base and expanding capacity” targets total supply and foundational capability. “Attacking annotation” targets data annotation, a labour- and technology-intensive step and a key process for quality improvement. “Improving quality and efficiency” concerns the transition from “having data” to “usable data”, emphasising standardisation and quality assessment.

“Enabling applications” connects datasets with “AI Plus” scenarios so data actually enters model training and industry use. “Managing and servicing” points to supporting systems such as registration, catalogues, circulation and security management of datasets. “Releasing value” concerns market-based allocation of data elements — letting datasets participate in circulation and distribution as factors of production.

The relationship with “AI Plus”

The plan stresses resonance with the “AI Plus” initiative. Dataset building is therefore not an isolated supply-side task but part of the broader layout serving industry intelligence.

This resonates with what companies feel: for firms building industry models, what is most lacking is often not algorithms but high-quality, lawfully usable industry data. The plan’s direction is to build up such supply at both public and industry levels.

Data-compliance prerequisites dataset building cannot avoid

From “collection” to “usable”, a dataset crosses a compliance chain. Whatever the plan deploys, companies participating in dataset building face several prerequisites.

First, lawful sourcing. Whether data is self-collected, purchased from third parties or publicly crawled determines the completeness of the authorisation chain. A dataset of unclear origin, however large, cannot be used safely.

Second, personal information processing. If a dataset contains personal information, there must be a lawful basis, notice duties must be met, and the minimum-necessary principle observed. As dataset scale rises and it circulates across entities, personal information protection requirements increase markedly.

Third, data security and classification. Identification and protection of important and core data, and cross-border arrangements, affect whether a dataset can circulate across entities or borders.

Fourth, compliance at the annotation stage. Data annotation involves personal information, intellectual property and employment relationships — an easily overlooked step.

Without resolving these, dataset building stalls at the circulation and use stage.

Three pointers for companies

First, put compliance at the collection stage. Compliance costs rise the later they are handled; building a source ledger and authorisation chain at collection and annotation is far cheaper than remediation.

Second, distinguish “can build” from “can circulate”. A dataset usable internally is not necessarily one that can be provided externally or across borders. Purpose determines the level of compliance required.

Third, watch the registration and catalogue system. As the management and service system develops, registration, labelling and quality assessment of datasets may become prerequisites for participating in circulation. For support on dataset compliance, personal information processing and data circulation arrangements, see our Data Compliance service.

This article is based on public information from the National Data Administration’s Implementation Plan of 8 June 2026. It is for reference only and does not constitute legal advice.