The opinion on high-quality development of the data labelling industry: policy on the training-data supply side

On 13 January 2025 the National Development and Reform Commission and other departments issued an opinion on promoting the high-quality development of the data labelling industry. The document looks upstream in the AI chain — the production of training data — and supports a shift in data labelling from a labour-intensive process to a technology-intensive one.

Three levels of attention

First, key technology work. The opinion calls for relying on national research programmes to advance cross-domain and cross-modal semantic alignment, 4D labelling and large-model labelling.

Second, tool development. It supports intelligent tools for multimodal labelling, label review, quality assessment and chain-of-thought expert labelling. This points to efficiency and quality control rather than simply scaling up labelling volume.

Third, the industrial ecosystem. By cultivating a data labelling industry, it aims to provide stable and predictable data supply for AI model training.

Why labelling deserves its own document

In the AI chain, data labelling has long been treated as a supporting step, yet it directly determines the quality and lawfulness of training data. As model capability converges and competition shifts to data quality, the strategic value of labelling rises. See our notes on the high-quality datasets action plan and on the National Data Administration datasets guide.

From a compliance angle, labelling is also where data provenance and the authorisation chain most often break. The lawfulness of the source and the rights status of labelled content travel into the model through training. See our notes on the model-data resonance action and on the opinion on data circulation service institutions.

What it means for companies

For companies training models, the opinion signals a predictable direction: data labelling supply will become more professional and tool-based, and companies buying labelling services should also look at the provider’s data sourcing and quality process.

The industrialisation of data labelling affects not only cost but also the traceability and compliance basis of training data. See our notes on the national integrated computing network service platform.

This article is based on the opinion on promoting the high-quality development of the data labelling industry issued by the National Development and Reform Commission and other departments and is for reference only; it does not constitute legal advice.