Skip to content
DCC · DATA COMPLIANCE CHINA China data law, for overseas counsel.
§ 125 · DEEPSEEK

The Questions Chinese Law Would Have to Answer: A January 2025 Reading of DeepSeek, the Garante, and Why EU Paperwork Cannot Fix Data Localization

A DCC summary of a 科技利维坦 piece on Chinese large models going overseas. Most of it is GDPR analysis; DCC keeps it for one paragraph that lists what would have to be settled in Chinese law before data stored in China is trusted abroad

A DCC summary, not a translation, of a 科技利维坦 (Li Wenlong) post dated January 31, 2025, written days after Italy's data protection authority, the Garante, questioned and then blocked DeepSeek. The post reads DeepSeek's position as a replay of OpenAI's in 2023, sets out what a minimum EU privacy posture for a large model looked like at the time, and identifies the one difference that compliance work cannot close: user data stored in China. Its China-facing conclusion is that EU documents and mechanisms do not solve localization. Trust would need clarity, in Chinese law, on the three data laws (CSL, DSL, PIPL) read with the National Intelligence Law, on the legal status of the Cyberspace Administration of China and the legal nature of the documents it issues, on due process when public bodies access data held by private companies, and on the remedies available to foreigners under the Civil Code and the PIPL. DCC sets out those four questions for counsel who are asked them from the other direction. The post predates everything after January 2025 and DCC has not updated its facts.

Editor’s Note — DCC.

This is an exception to DCC’s sourcing rule, made at the editor’s choice. DCC does not normally carry commentary on EU or US law, and most of this post is about the GDPR. It is summarized here, not translated, because of where it ends: with a list of what would have to be clarified in Chinese law before a Chinese AI company’s data practices are trusted abroad. Overseas counsel meet that list from the other side, whenever a client asks whether data may sit with a provider in China.

It is dated. The post was published on January 31, 2025, in the first week of the DeepSeek inquiries. Everything it says about pending proceedings is as of that date. DCC has not updated the facts and takes no position on how those proceedings ended.

Attribution. The post is unsigned on the channel; 科技利维坦 is Li Wenlong’s (李汶龙) personal account and DCC attributes it to him on that basis. GDPR terms below are the source’s own subject matter, not a gloss on Chinese law.

The post in six points

1. DeepSeek in January 2025 looked like OpenAI in March 2023. A technical shock, then regulators pressing for compliance to catch up. The author recalls the ChatGPT sequence: a data leak around March 20, 2023, in which some paying users saw other users’ chat history; a temporary ban by the Italian authority, the Garante, at the end of that month; restoration about a month later once some issues were addressed; and a finding of violation the following year. The Garante had identified at least four problems: inadequate privacy disclosures, no suitable legal basis (OpenAI first relied on performance of a contract, which the Garante held insufficient), inaccurate output about individuals, and no age verification despite a stated minimum age of 13.

For DeepSeek, the author reports, the Garante opened with an information request to be answered within 20 days and followed with a blocking order soon after. Ireland’s authority was reported to have made inquiries as well, with no official statement at the time of writing. A security firm had just disclosed a publicly accessible database.

2. No establishment in the EU means no single regulator. ChatGPT had no EU establishment in 2023, so the one-stop-shop mechanism did not apply and every national authority had jurisdiction, each with different priorities. The author saw DeepSeek in the same position, which “puts EU compliance in hard mode from the start”. OpenAI began setting up an EU representative and establishment in early 2024. The author says this “seems to be the road every model vendor entering the European market must eventually take”.

3. The questions had moved closer to the model. Beyond the basics (whether personal data is collected, from where, for what, on what legal basis, with what notice to registered and unregistered users), the DeepSeek request asked about the sources of training data and about the use and share of web scraping.

4. A minimum EU posture, as the author saw it in early 2025.

  • A privacy policy with real granularity. The author notes that the policies of the large US providers had problems of their own, so matching the prevailing standard was not hard in practice.
  • A legal basis for training. Contract does not work; consent is difficult, though it cannot be avoided entirely; legitimate interests was becoming the mainstream route. The cost lies in the balancing analysis and the safeguards, and in being able to give “a coherent and complete account” when a regulator asks. Where training does not rest on consent, an accessible way for users to opt out was becoming the norm.
  • Children. A long-standing Garante priority; the bar across the EU was rising.
  • Hard problems that could come later: identifying and separating sensitive data, hallucination, machine unlearning to give effect to erasure, and rectification. The author’s view was that legal certainty on these was low and their priority for baseline compliance correspondingly lower.

5. The “DeepSeek paradox”. The author cites the French scholar Theodore Christakis for the point that DeepSeek had no data “exporter” inside the EU, so that, on the European Data Protection Board’s guidance, the GDPR’s chapter on international transfers did not apply: in that situation there is no “transfer” at all. The other provisions of the GDPR still applied, and the author notes that the position was contested for some data flows.

6. The one difference from OpenAI: where the data is. The author’s central observation is that the DeepSeek inquiries differed from the 2023 ChatGPT ones in a single respect, data localization, and that this is “a problem a company’s own compliance cannot solve”. He links it to complaints that the privacy group noyb had filed shortly before against six large Chinese internet companies over data sent back to China, one of them lodged in Italy.

The paragraph DCC keeps this post for

Translated in full, less a closing line announcing an online event:

Even if DeepSeek and the other model vendors assemble all the compliance documents and mechanisms the EU requires, that still will not solve the compliance problem that localization brings. To build real trust at the international level, progress can only come once a series of questions has been clarified at the level of law in the strict sense: the three domestic data laws (plus the Intelligence Law); the legal status of the Cyberspace Administration and the legal nature of the documents it issues; due process for public-sector access to data held by the private sector; and the remedies by which foreigners can exercise their rights under the Civil Code and the Personal Information Protection Law. The questions of data sent back to China and of public-sector access are already at the center of the storm.


DCC’s reading

The notes below are DCC’s, not the author’s.

The author is a Chinese scholar writing for Chinese AI companies, and his list is a candid one. It is close to the list a foreign regulator or a transfer risk assessment works through when the destination is China. Set out as questions:

  1. How do the three data laws read together with the National Intelligence Law? The Cybersecurity Law, the Data Security Law and the PIPL mainly regulate handlers of data. On state access, Article 35 of the DSL says that public security and state security organs obtaining data must go through “strict approval procedures” under the relevant provisions of the State, and that organizations and individuals shall cooperate; Article 48 fines a refusal. The statute does not set out what the approval procedure is. That is the gap the author points to.
  2. What is the CAC, legally, and what are its documents? The Cyberspace Administration of China issues departmental rules, normative documents, guides and Q&As. Their rank differs, and so does what a foreign reader can rely on. DCC’s law pages record the rank of each instrument.
  3. What process governs public-sector access to privately held data? The author treats this as unsettled. DCC’s brief on the reverse situation, foreign authorities seeking data stored in China, is Cross-Border Data Discovery.
  4. Can a foreigner enforce PIPL rights in practice? The rights chapter of the PIPL speaks of “individuals” (个人) and states no nationality condition. The author’s point is about remedies: whether a person outside China can in fact bring a request, a complaint or a claim, under the PIPL or the Civil Code’s personal information provisions.

Two cautions. First, the post is an opinion written in one week of fast news, and its account of the Italian and Irish proceedings should not be cited as a record of them. Second, it does not answer its own four questions. Its value is that a Chinese commentator, addressing Chinese companies, named them as the precondition for trust.

— 科技利维坦 (Li Wenlong), Deepseek隐私之困:国内大模型出海隐私合规之路 (DeepSeek’s Privacy Predicament: The Road to Privacy Compliance for Chinese Large Models Going Overseas), January 31, 2025. Summary by DCC; one paragraph translated in full. Original (Chinese).

Facts about the EU proceedings, the security disclosure and the noyb complaints are as stated by the original author in January 2025; DCC has not independently verified or updated them.

— Not legal advice.

— Not legal advice.


§ SUBSCRIBE

The Monday brief.

One short email every Monday. New briefs on Chinese data-compliance rules from the previous week, with the source law cited.

Opt-in only. Unsubscribe anytime by replying "unsubscribe" to any issue.

SUPPORT DCC

Keep the publication free to read. Suggested support is $19.99, or choose your own amount.

Support →