Interim Measures for Generative AI
Where must the corpus for training a large model come from to be lawful?
Basis: Interim Measures for the Administration of Generative AI Services (《生成式人工智能服务管理暂行办法》)
Conclusion: Article 7 of these Measures requires providers to carry out pre-training, optimisation training and other training data processing activities in accordance with the law, and sets three bottom lines: use data and base models with lawful sources; where intellectual property is involved, do not infringe the intellectual property that others enjoy in accordance with the law; where personal information is involved, obtain the consent of the individual or meet other circumstances provided by laws and administrative regulations. Training data must also be made more truthful, accurate, objective and diverse.
What you may need to do
- Build a corpus ledger recording the source and the proof of authorisation batch by batch
- De-identify corpora containing personal information, or obtain consent
- Screen corpora for copyrighted works and keep the authorisation documents
- Write labelling rules and verify labelling accuracy by sampling
- Train labelling staff on compliance and keep records
- Verify the lawful source of the base model as well
Provisions relied on
Interim Measures for the Administration of Generative AI Services, Article 7
(1) use data and base models with lawful sources;
Interim Measures for the Administration of Generative AI Services, Article 7
(2) where intellectual property is involved, do not infringe the intellectual property that others enjoy in accordance with the law;
Interim Measures for the Administration of Generative AI Services, Article 7
(3) where personal information is involved, obtain the consent of the individual or meet other circumstances provided by laws and administrative regulations;
One-line conclusion
Whether a corpus is compliant does not depend on “whether it can be found online”, but on three things: whether the source is lawful, whether it infringes anyone's intellectual property, and, where personal information is involved, whether consent has been obtained or a statutory circumstance applies. These three are the bottom lines for training data drawn by Article 7 of Interim Measures for the Administration of Generative AI Services.
Basis
Article 7 requires providers to carry out pre-training, optimisation training and other training data processing activities in accordance with the law, and lists five items. Three of them are the ones most likely to be questioned:
Item (1): use data and base models with lawful sources. Note that this governs the base model as well as the data; taking open-source weights of unknown provenance for further training is equally problematic.
Item (2): where intellectual property is involved, do not infringe the intellectual property that others enjoy in accordance with the law. The presence of others' works in a corpus does not mean they can be used freely; the basis for use must be explainable.
Item (3): where personal information is involved, obtain the consent of the individual or meet other circumstances provided by laws and administrative regulations. The key phrasing is “or other statutory circumstances”, which means consent is not necessarily the only route, but a specific legal basis must be identified.
Item (4) requires effective measures to improve the quality of training data and to enhance its truthfulness, accuracy, objectivity and diversity. Item (5) is a catch-all: the Cybersecurity Law, the Data Security Law, the Personal Information Protection Law and other relevant provisions and the regulatory requirements of the competent authorities also apply.
Article 8 separately addresses data labelling: where data labelling is carried out during research and development, the provider shall formulate clear, specific and operational labelling rules, carry out labelling quality assessment and sample verification of the accuracy of labelled content, and provide necessary training for labelling staff.
What you need to do
- Build a corpus ledger. For each batch of data record four things: the source channel, the method of acquisition, the proof of authorisation, and whether personal information is included. Without a ledger you cannot evidence your position when inspected.
- Classify the corpora containing personal information: obtain consent where it can be obtained, and where it cannot, identify which statutory circumstance applies and keep the reasoning on file.
- Keep authorisation documents for corpora containing the works of others; where there is no authorisation, assess whether the use falls within fair use, and remove from the training set anything whose source cannot be explained.
- Put the labelling rules into a document. The rules must be specific about what is labelled, what counts as acceptable, the sampling ratio and the rework process.
- Train labelling staff and keep records, with the training focused on law-abiding awareness and standardised operations.
- Archive the source and licence of the base model as well, in particular any commercial-use restrictions.
Common pitfalls
- Treating “publicly accessible” as “free to use for training”. Public accessibility and permission for commercial training are two different things.
- Checking the data but not the model. Article 7, item (1) puts the base model and the data in the same item.
- Outsourcing labelling and paying no further attention. Article 8 places the duties of labelling rules, quality assessment and staff training on the provider, and outsourcing does not exempt it.
- Doing a one-off review only. The corpus is continually supplemented, so the ledger must be updated along with it.
Common follow-up questions
Can publicly scraped data from the internet be used for training directly? You need to show that the source is lawful, and assess whether it infringes anyone's intellectual property or contains personal information. Public accessibility does not mean it may be used freely for model training.
If data labelling is outsourced to a third party, does the contractor bear liability when something goes wrong? Under Article 8 the responsible subject is the provider. Formulating the labelling rules, assessing quality and training staff are the provider's responsibility, and outsourcing does not transfer those duties.
Can an open-source model be used commercially? That depends on the specific licence. Article 7 requires the use of base models with lawful sources, and lawfulness of source includes whether the licence conditions are satisfied for a commercial scenario.
Common follow-up questions
- If the corpus contains someone else's article, must we always obtain authorisation?
- The requirement in Article 7 is not to infringe the intellectual property that others enjoy in accordance with the law. Either obtain authorisation or be able to explain that the use falls within a use permitted by law, and keep the reasoning on file for inspection.
- Can user conversation records be used for further training?
- Where personal information is involved, the consent of the individual must be obtained or another circumstance provided by laws and administrative regulations must apply. User input information is also protected by Article 11 and cannot be retained or repurposed at will.
- What counts as adequate data quality?
- Article 7 requires effective measures to improve the quality of training data and to enhance its truthfulness, accuracy, objectivity and diversity. In practice this means building a verifiable cleaning and sampling process, not giving a verbal assurance.