⁠​​‌‌​‌​‌​‌​‌‌​‌‌​‌​‌​‌​​‌​‌‌‌​‌‌⁠Interim Measures for Generative AI

What the user types to the AI -- how long can it be stored, and can it be used for training?

What you may need to do

1. Map the conversation log fields and mark the parts that can identify an individual.
2. Set a retention period and a deletion mechanism for each category of input information.
3. Turn off personal information collection items you do not need.
4. Set up an intake point and a time limit for personal information rights requests.
5. Define the approval route and the legal basis for providing logs to others.

Provisions relied on

One-line conclusion

The fact that a user typed something does not mean the provider may store and use it freely. Article 11 draws three prohibitions and one obligation for providers: do not collect unnecessary personal information, do not unlawfully retain identifiable input information and usage records, and do not unlawfully provide them to others; and do promptly accept and handle requests to access, copy, correct, supplement and delete personal information.

Basis

Article 11, paragraph 1 names the protected objects precisely: input information and usage records. These two categories of data are natural by-products of an AI service, and many people assume they are not sensitive; in fact they may contain names, mobile numbers, identity card numbers, health-related information or internal company material.

The three prohibitions are: do not collect unnecessary personal information; do not unlawfully retain input information and usage records that can identify a user; do not unlawfully provide users' input information and usage records to others. Note the word unlawfully -- it does not mean retention and provision are absolutely forbidden, but that they must have a legal basis.

Article 11, paragraph 2 is the duty to respond to individual rights: a provider shall promptly accept and handle, in accordance with the law, individuals' requests to access, copy, correct, supplement and delete their personal information. This requires an entry point, a time limit and feedback on the outcome.

In addition, Article 9, paragraph 1 provides that where personal information is involved, the provider bears the responsibility of a personal information processor in accordance with the law and performs personal information protection obligations. Article 11 is where that responsibility lands concretely in the AI context.

What you need to do

1. Do a data map. List every conversation input, generated result and call log stored in the system, marking whether it contains personal information, what purpose it serves and how long it is kept.
2. Every category of data needs a retention period. Data with no period is retained permanently by default, which is where problems most often arise.
3. Trim what you collect. Gather only the fields needed to deliver the function, and give a clear statement of purpose for items such as mobile numbers and identity card numbers.
4. Confirm the legal basis before providing data to others (including affiliates, suppliers, or a model vendor for optimisation), and keep the approval records.
5. Set up a personal information rights request channel, publish how to submit a request, and record the times of acceptance and handling.
6. Filter or de-identify inputs containing sensitive information so that they do not land in logs directly.

Common pitfalls

• Logs kept permanently by default. Article 11 explicitly prohibits unlawful retention, and long-term indiscriminate retention lacks a basis.
• Training models on user conversations by default. That involves both retention and provision under Article 11 and the basis for personal information processing under Article 7, item (3).
• Deleting only from the primary database and not from backups and logs. A deletion request must cover every copy.
• Handling rights requests verbally through customer service, with no entry point, no record and no time limit.

Common follow-up questions

How soon must user input be deleted? The Measures do not prescribe a uniform number of days; the requirement is not to retain unlawfully. In practice, set a retention period based on business necessity, tell users about it, delete automatically when the period expires, and keep a trace of the deletion.

If we call a third-party model API and the input goes to the vendor, is that unlawful provision? Look at whether there is a legal basis, whether notice obligations are met, and whether the vendor processes the data as agreed. Article 11 prohibits unlawful provision; a compliant path must be explainable and supported by agreements and records.

Can we refuse a user's request to delete their chat history? Article 11 requires prompt acceptance and handling in accordance with the law. Without legitimate grounds you should not refuse, and the outcome should be fed back and recorded.

Common follow-up questions

After anonymising user conversations, may we keep them long term?

The key question is whether the user can still be identified. Article 11 prohibits unlawfully retaining input information and usage records that can identify a user. Even after de-identification you must assess re-identification risk and give a justification for the retention.

How quickly must we answer a user who asks to see their information?

The Measures require prompt acceptance and handling in accordance with the law. The specific time limit follows the Personal Information Protection Law (个人信息保护法) and comparable laws and your local cyberspace administration; internally, set and publish a clear service time limit.

Is protecting input information the same as protecting usage records?

Article 11 places them side by side with the same standard of obligation. Usage records include call times, functions and generated results, and they may likewise not be unlawfully retained or unlawfully provided.

This entry belongs to the “China AI Compliance and Filing” cluster (42 entries in total): View all · Answers home