Qwen and DeepSeek: Paris seals their training data, the AI Act claims them — ActuIA

November 21, 2025 | Source: ActuIA
Create monitoring on this topic Receive upcoming developments, with an AI monitoring set up from this article.
Create my free monitoring

European digital directions that have deployed Qwen or DeepSeek in their pipelines since 2025 now face, since Monday, June 1st, 2026, an unprecedented dual legal constraint. At this date, the Provisions on the Protection of Trade Secrets issued on February 24th, 2026 by the State Administration for Market Regulation (SAMR, Order No. 126) now classify AI training data, recommendation algorithms, and risk control models as administratively protected trade secrets, as reported by Bloomberg on June 2nd, 2026. However, Article 53 of the European AI Act requires general-purpose model providers (GPAI) to publish a sufficiently detailed summary of the content used to train the model - including when the model is open-weight. The regulatory collision is direct for the pre-compliance verification of DSI involved in these models.

A now massive installed base, a coordinated device from Beijing

The stakes of this collision are no longer hypothetical. According to a HuggingFace data analysis cited by the South China Morning Post, Alibaba's Qwen model family would have generated 153.6 million downloads in the single month of February 2026, more than double the cumulative total of the eight major actors listed, including Meta, DeepSeek, and OpenAI. The official agency Xinhua had documented in January 2026 the Qwen model surpassing the billion cumulative downloads on HuggingFace and reaching the threshold of 200,000 derived models - a technology park that the ActuIA archive was already tracking through DeepSeek-R1 and the new Qwen2.5-1M variants. The SAMR regulation is not alone. Bloomberg reported on May 26, 2026, the extension by China of outbound travel restrictions to the best AI profiles of private firms such as Alibaba and DeepSeek, conditioning all international travel on government approval. Reuters covered on June 1, 2026, the parallel tightening of outbound investment rules imposed by the Chinese State Council - central government. Three measures concentrated within a 15-day window outline a coordinated control mechanism for technology flows, talent, and capital.

An administrative regime without a judge, with extraterritorial reach

The instrument chosen by Beijing is administrative, not judicial. According to China IP Law Update, the regulator's actions may include orders to cease counterfeit activities, confiscation of illegal gains, and fines, with no available damages in this route - which adds to the existing civil and criminal avenues. The material extension constitutes the decisive innovation. The same source documents that the definition of what can be qualified as trade secrets now includes data and algorithms - categories not previously covered by the 1995 texts or the 2019 amendments to the unfair competition law. The primary Chinese text is aggregated on the MOFCOM (Chinese Ministry of Commerce) regulatory portal under the identifier id=105543, the regulation itself emanating from the SAMR (Order No. 126 of February 24, 2026). The extraterritorial scope of the provision remains aspirational: no operational application mechanism outside of China has been notified at this stage, observes China IP Law Update. Unlike American law, which since 2016 has opened a federal judicial remedy exclusively without direct administrative enforcement, the Chinese regime relies on the market regulatory weapon - leaving the question of its enforceability against a provider established in the EU entirely open.

Article 53 AI Act: what "sufficiently detailed summary" imposes on a DSI

The conflict arises from two contradictory obligations. On one hand, Article 53 of the AI Act requires providers of GPAI to make publicly available a sufficiently detailed summary of the training content, with the precise procedures to be established by an implementing act of the European Commission. Practically, for a European DSI incorporating a Qwen or DeepSeek model into a product sold in the EU, a sufficiently detailed summary entails, in practice, a data origin card, categorization of sources, and indication of volumes and acquisition dates. These are precisely the information categories that the SAMR regime now classifies as administratively enforceable trade secrets. On the other hand, disclosing this summary exposes one to a Chinese administrative risk; not disclosing it exposes one to a European sanction risk. The EU 2016/943 directive on trade secrets offers an alternative but subordinate judicial route, whose relation to the AI Act has not yet been clarified. The first dispute between an European GPAI provider and the obligation to disclose a summary will set the applicable doctrine for the entire market - and will condition the legal value of Chinese deployments open-weight already in production.

Disclosing the training summary exposes one to a Chinese administrative risk; not disclosing it exposes one to a European sanction.

Article 53 of the AI Act imposes the publication of a training data summary - an obligation now in direct conflict with the new SAMR rules, which classify these same data as enforceable trade secrets (effective June 1, 2026).

Travel, capital, outflows: three measures in fifteen days

The SAMR regulation is just one piece of a larger puzzle. On the talent registry, Bloomberg reported on May 26, 2026, the application of restrictions requiring government approval for any international travel by Alibaba and DeepSeek's AI profiles. On the capital registry, Reuters reported on June 1 the tightening of rules on outbound investment by the Chinese State Council - central government, targeting in particular the offshore structures regularly used by tech groups. On the register of information flows, finally, according to the Wall Street Journal of June 1, 2026, the restriction of unauthorized exports of goods, technologies, services, and state data comes to complement the framework. The three measures, concentrated within a fifteen-day window, impose on European AI user companies a verification process now across three registers: compliance with the AI Act transparency obligations, exposure to the SAMR regime, and traceability of upstream technical flows from Chinese suppliers.

Four coordinated measures from Beijing - convergence on June 1, 2026

Register Content In effect Source
Trade Secrets (SAMR) AI training data, recommendation algorithms, and qualified risk control models classified as trade secrets with enforceable rights June 1, 2026 Bloomberg / Global Times
Restrictions on Travel for AI Talents Government approval required before any international travel for strategic AI profiles (Alibaba, DeepSeek named) End of May 2026 Bloomberg, May 26, 2026
Strengthening of Outward Investment Rules New rules from the State Council's Affairs Council targeting offshore structures and technological capital flows June 1, 2026 Reuters, June 1, 2026
Export Restrictions New regime governing the export of goods, technologies, services and state data June 1, 2026 WSJ, June 1, 2026

 

Add ActuIA to my preferred sources on Google

Our articles will appear first in Google's « Headlines ».

Also read
  • Pékin veut restreindre l'export de 23 technologies d'IA, sans toucher à l'API de DeepSeek et Qwen Bill / consultation Beijing wants to restrict the export of 23 AI technologies, without affecting the DeepSeek and Qwen APIs October 20, 2026
  • Pékin veut encadrer les agents d'IA du commerce selon leur usage Bill / consultation Beijing wants to regulate AI agents in commerce based on their use October 22, 2026
  • Nouvelle étape pour l'AI Act : les obligations pour les modèles d'IA à usage général sont entrées en vigueur Law/adopted text AI Act: obligations for general-purpose AI models are now in effect August 4, 2025
  • AI Act : le compte à rebours est lancé pour les entreprises avant le 2 août 2026 Law/adopted text AI Act: countdown launched for companies before August 2, 2026 June 23, 2026
In the watch
  • November 4 China: Beijing Huairou district puts ten measures and a 2026-2028 support plan for « one-person companies » in AI on consultation
  • November 3 China: NDRC announces 38 founding countries for the World Organization for AI Cooperation and a watch on employment
  • November 4 According to Bloomberg, Irregular attribute the sandbox breaches of OpenAI, Anthropic, Meta, and Google models to a "human error"
All night →

open sourceAI ActDeepSeek-R1
ST
Stephane Nachez

ActuIA - News, data, and analysis on artificial intelligence for decision-makers.

@TAGn@