Code Wars, 2025 – 02
Where Tech meets Geopolitics
Jean-Baptiste Monnier, PhD
VP Digital Governance Asia Centre
jb.monnier@asiacentre.eu
DeepSeek, a Sputnik (or a Buran) moment?
AI and the CCP: A Clash Between Control and Change
Vladimir Putin, 2017: “Whoever becomes the leader in this [artificial intelligence] sphere will become the ruler of the world.”
At the AI Action Summit in Paris (2025), Chinese Vice Premier Zhang Guoqing touted China’s DeepSeek breakthrough, calling for global AI collaboration and U.S. VP JD Vance warned against AI ties with authoritarian regimes, citing security risks. The summit highlighted sharp divides in AI governance.
DeepSeek’s rise has been dramatic. It burst onto the scene just as Donald Trump was sworn in, and by January 27, it was shaking Wall Street and sublimating $1 trillion, the largest drop in history.
DeepSeek, bound by CCP regulations as a “technology with Chinese characteristics,” faces the same regulatory constraints that limited Baidu. This paper examines how China’s control mechanisms might shape DeepSeek’s trajectory and whether AI, vital to China’s global ambitions, can thrive under such conditions.
China’s Digital Iron Curtain
Bill Clinton in 2000 mocked China’s efforts to block the internet: “Good luck! That’s like trying to nail Jell-O to the wall.” Yet the CCP did exactly that. By building the Great Firewall (GFW), China created the first “splinternet,” insulating its cyberspace from foreign influence and setting a precedent for authoritarian regimes.
Baidu’s wings clipped
Baidu, once poised for global expansion, stumbled under CCP control. It failed in Japan (2008-2015), flopped in Brazil (2014), and was banned in India (2020). Today, 95% of its 669 million users are in China. After Google left China (2010), Baidu enjoyed a domestic monopoly—yet it couldn’t break into global markets. Why? GFW, censorship and opaque regulations created doubts in the eyes of foreign search users, advertisers and partners. The result? Baidu’s market cap is now just 0.1% of Google’s.
DeepSeek: Forging the Future with Elite Talent
DeepSeek spun off from High-Flyer, a quantitative hedge fund using AI for high-frequency trading. Launched in 2023, DS assembled a team of elite researchers to push the frontiers of AGI.
DeepSeek’s breakthroughs are impressive. Its Mixture of Experts (MoE) model maximizes efficiency while cutting compute costs. It ranks second in MMLU benchmarks, trailing only OpenAI’s GPT-4. Unlike most labs, DeepSeek prioritizes radical transparency—its research paper, signed by 200 team members, details key innovations. It also fully embraces opensource using the very permissive MIT license.
AI: The Censorship Nightmare
Steering AI is no easy task. Mitigating hallucinations is one challenge; enforcing “core socialist values” is like… nailing Jell-O to the wall, again. As users provoked DeepSeek with the term “Tiananmen,” it quickly became a viral meme. To be sure, China’s strategy for AI is ambitious. The Cyberspace Administration of China (CAC) built “question banks” (the Refusal-to-Answer Test Question Bank or the Generated Content Test Question Bank), each of which includes thousands of sensible topics that may “incite the subversion of state power” or “undermine national unity”. Most AI vendors have installed 2 filters, one at the input of the AI and one at the output.
Pre-Screening: User prompts are checked against a restricted topics list (e.g., Tiananmen). If flagged, the query is blocked before it ever reaches the AI, triggering a canned response instead.
Post-Processing: If a prompt makes it through the pre-screening, then it is processed by the AI and a response is they typed out. But as text appears on-screen, it is simultaneously analyzed against the “Generated Content Test Bank.” If flagged, the system abruptly interrupts the output flow—wiping the screen before the full response can be read. Now even DS is inconsistent, below is a side-by-side comparison of DeepSeek’s V3 model (left), compared to the new R1 model (right). On the left, you can see the pre-screening filter successfully blocking the prompt, while on the right both filters are actually failing (or possibly, they are not activated).
Pre-Screening: User prompts are checked against a restricted topics list (e.g., Tiananmen). If flagged, the query is blocked before it ever reaches the AI, triggering a canned response instead.
Post-Processing: If a prompt makes it through the pre-screening, then it is processed by the AI and a response is they typed out. But as text appears on-screen, it is simultaneously analyzed against the “Generated Content Test Bank.” If flagged, the system abruptly interrupts the output flow—wiping the screen before the full response can be read.
Now even DS is inconsistent, below is a side-by-side comparison of DeepSeek’s V3 model (left), compared to the new R1 model (right). On the left, you can see the pre-screening filter successfully blocking the prompt, while on the right both filters are actually failing (or possibly, they are not activated).
For now, China’s AI censorship remains difficult to enforce but if history is any guide, progress are only a matter of time.
Would You Trust DeepSeek with Your Data?
AI thrives on data, from banking to medical records. But in China, data privacy is secondary to state control. China’s Data Security Laws allow authorities full access to company-held data and DeepSeek operates under state oversight. South Korea has flagged DeepSeek for “excessive” data collection; Australia and the state of NY have banned DS from government systems.
And then there’s compute. DeepSeek’s frugal training methods did reduce GPU reliance, but as it plans to serve millions, inference scaling will require massive hardware capacity. Besides tariffs, Trump’s administration has yet to lay out its export control is-à-vis China. This will require coordination with partners such as Holland (ASML), Germany (Siltronic, Carl Zeiss and Trumpf), and Japan (Tokyo Electron).
Conclusion: Competing Models, Diverging Futures
China defied sceptics by “harmonizing” the internet. But AI is different—tight regulations could stifle its competitive edge.
Despite restrictions, China is positioned to lead in AI applications across education, healthcare, infrastructure, and defense. But true AGI leadership requires openness—not just for development, but for commercialization. There are options:
- China will likely lead in a number of local AI applications across education, healthcare, infrastructure, and defense, and only those applications that are not putting public data at risk will be global players.
- A TikTok-style split—separate models for China and global users—might help DS and others, but how trustful will it be?
- And then CCP may consider relaxing its policies, just as it reopened its economy post-COVID. That means AI reshaping China…
For as long as the vendors are closely matched in technology, the real competition will center on governance. The world is diverging, the U.S. embraces a Laissez-faire, China enforces State Control, and Europe seeks a Trust-Led Innovation.
AI is reshaping the world—but who is shaping AI?