The finding that unsettles the obvious
In the dominant Western narrative, China is synonymous with censorship. The Great Firewall (the internet filtering and censorship system put in place by the Chinese government, which has blocked sites such as Google, Facebook, YouTube and Wikipedia from within Chinese territory), media control, state surveillance: these realities exist and are documented. But when you look at the Chinese AI models being deployed globally in 2026, a stranger finding emerges. DeepSeek, Qwen and other Chinese models frequently agree to answer questions that ChatGPT, Claude or Gemini refuse.
On SpeechMap (an independent site that assesses the permissiveness of AI models on politically sensitive questions in the United States: civil rights, protests, political satire, moral debates), DeepSeek appears among the most permissive models in the world. More permissive than OpenAI's GPT-5 on the majority of American political topics.
How does a country accused of controlling everything export AIs that are freer than those produced in California? The answer reveals something fundamental about China's AI strategy, and about the blind spots in Western discourse.
The true scale of global distribution
Let's start with the figures. According to a joint study by MIT (Massachusetts Institute of Technology, one of the most prestigious American universities in science and technology) and Hugging Face (a French platform founded in New York that hosts the world's leading library of open-source AI models), Chinese open-source models accounted for 17.1% of global AI downloads in the year ending August 2025. It is the first time in history that China has overtaken the United States (15.86%) on this metric. On OpenRouter (one of the main marketplaces aggregating large language models), three Chinese labs (DeepSeek, Alibaba's Qwen, Z.AI) feature consistently in the top 10 by usage in the second half of 2025.
Qwen alone (the model family from Alibaba, the Chinese e-commerce and cloud giant) now totals more than 700 million downloads, making it the world's largest supplier of open-source AI. By comparison, Meta Llama, its American equivalent, is far behind. Of the 20 most intelligent models according to Artificial Analysis evaluations (an independent firm that benchmarks AI models), 6 are Chinese. Of the 15 best open-source/open-weight models, 9 are Chinese.
Singapore chose Qwen rather than Llama to build its regional sovereign model. Malaysia has announced that its national AI ecosystem would run on DeepSeek. Founders in Kenya, Brazil, India and even San Francisco are building their applications on Chinese foundations. The narrative that "China is frightening in AI" is contradicted by the concrete choices of states and developers around the world, including in countries with no particular sympathy for Beijing.
On what are Chinese models more permissive?
The picture is precise and nuanced. Here is what we observe when DeepSeek (Chinese AI lab founded in 2023, which became globally known in January 2025 with its R1 model that rivalled GPT-4 at significantly lower training costs) and Qwen are systematically tested against American models:
- American politics. DeepSeek and Qwen answer openly to questions about Donald Trump, about the BLM protests (Black Lives Matter, the American social movement against police violence towards African Americans), about constitutional controversies. Claude and ChatGPT refuse or hedge cautiously on many of these questions (in French: "esquivent", or answer in an excessively balanced way).
- Criticism of tech companies. The Chinese models are willing to criticise Google, Microsoft and Apple more frankly than the American models, which tend to balance obsessively.
- Mature and adult content. Several Chinese models generate adult fiction that ChatGPT would systematically refuse. Models derived from DeepSeek (notably those hosted outside China) are widely used by roleplay (role-playing games, mainly narrative) and fiction platforms.
- Medical, legal and financial questions. Where ChatGPT systematically adds "consult a professional" warnings, the Chinese models generally give the answer directly, leaving the user to judge.
- Recipes, sensitive technical instructions. Several studies have shown that the Chinese models are more vulnerable to "jailbreaking" (in French: libération, a technique consisting of bypassing an AI model's guardrails to make it produce prohibited content) to produce dangerous content. According to a Swedish study, some versions of DeepSeek and Qwen can be bypassed on critical subjects.
On this last point, it is not a quality: it is rather a security weakness that can be exploited by malicious actors. But it illustrates that the "guardrails" are structurally less strict.
On what are the Chinese models more restrictive?
The other side of the equation is very precise. The Chinese models have hardcoded restrictions (coded into the model, impossible to bypass) on subjects sensitive to the Chinese government. Researchers have identified explicit internal directives:
- DeepSeek is instructed to "avoid the common taboos of the Communist Party"
- Qwen is directed to keep answers about China "positive and constructive, avoid criticism and highlight successes"
- On questions concerning Taiwan, the 1989 Tiananmen massacre (student pro-democracy protests violently suppressed by the Chinese army in Tiananmen Square in Beijing, a taboo subject in China), the Uyghurs (a Turkic-speaking Muslim minority in Xinjiang, accused of being the victim of massive repression by Beijing), Tibet, Hong Kong or Falun Gong (a spiritual movement banned in China since 1999), the models refuse or produce answers aligned with official Chinese positions.
More subtly: researchers from Policy Genome and the Estonian Foreign Intelligence Service (Estonia's foreign intelligence service, one of the European countries most committed to cybersecurity) have documented that responses change depending on the language used. When DeepSeek is asked for information about the war in Ukraine in English or Ukrainian, the responses are largely factual. In Russian, several responses echo Kremlin (the Russian government) talking points.
The Estonian foreign intelligence service's 2026 report contained a chilling finding: "When discussing topics related to Estonia's security, DeepSeek conceals key information and inserts Chinese propaganda into its responses." This finding, from a European intelligence agency, confirms what many observers suspected: Chinese models are not politically neutral. They spread Beijing's positions on sensitive geopolitical topics on a large scale, sometimes subtly, sometimes explicitly.
Why this strategy of paradoxical permissiveness?
The apparent contradiction (liberal on American politics, restrictive on Chinese topics) is not an accident. It is a coherent strategy, which can be broken down into four logics.
Logic 1: conquer the markets where American models are restricted. As we documented in our article on the US-China AI war, DeepSeek has captured 49% of the market in Cuba, 43% in Russia, 39% in Iran. These countries are blocked from OpenAI and Anthropic services by American sanctions. Chinese models fill the void.
Logic 2: capture the Global South and emerging countries. The "Global South" is a geopolitical term designating the developing countries of Africa, Latin America, Asia and Oceania, as opposed to the "Global North" (North America, Europe, Japan). For a startup in Kenya or a developer in Brazil, the calculation is simple. ChatGPT at 20 dollars a month is expensive in local currency. Qwen or DeepSeek are free, perform well, and accept uses that the Americans restrict. The result: China is building massive technological soft power (influence through cultural and economic attraction rather than coercion) in regions that will be strategic for the next 20 years.
Logic 3: compensate for the hardware lag through software excellence. Under embargo on Nvidia chips (the American company that dominates the manufacture of graphics processors used to train AI) since 2022, Chinese laboratories have been forced to innovate on algorithmic efficiency. Some of these innovations have become available in open source. For American laboratories, releasing their models as open source would reduce their margins. For China, it is paradoxically the strategy that maximises influence.
Logic 4: the separation between the domestic market and exports. Models deployed in China are subject to the control of the Cyberspace Administration of China (the Chinese internet regulator). But the open source versions circulating internationally are not accessible to Chinese citizens themselves (Hugging Face is blocked in China). China therefore operates a differentiated policy: strict at home, liberal for export, because both serve the state's interests.
The risk Washington is starting to see
The US administration has finally grasped the scale of the strategic problem. At the AI Impact Summit in New Delhi in 2026, Michael Kratsios (director of the White House Office of Science and Technology Policy, the body that advises the US president on science and technology policy) set out a strategy for exporting American AI. The implicit message: if the Global South builds on Qwen and DeepSeek, Chinese influence over the intellectual infrastructure of tomorrow's world becomes irreversible.
The problem for Washington is that the defensive solution (banning DeepSeek and Qwen) is technically impossible. As we explored in our article on deepfakes and in our analysis of the impossibility of fully regulating AI, open source models cannot be recalled once they have been released. DeepSeek is available on dozens of mirrors (copies of the site on other servers). Qwen is built into Microsoft's Azure. Banning these models across the United States would only create workarounds through intermediaries.
The only strategically viable response is to make American models more accessible, more permissive and more competitively priced. But that requires American companies to accept lower margins and fewer guardrails, which they refuse to do for commercial and legal reasons (US class actions over dangerous content terrify them).
What this reveals about Europe
Europe, in this equation, is largely absent. No European model features in the global top 10 for downloads. Mistral (French lab founded in 2023, one of the few European AI champions), the French champion, is competing but remains far below Chinese or American volumes. As we documented in our article on France falling behind, Europe has the talent but not the scale.
More problematic still: Europe faces an uncomfortable geopolitical choice. Using American models means accepting strategic dependence on the US Big Tech firms. Using Chinese models means embedding Chinese political biases and restrictions in its digital ecosystem. Rejecting both means condemning itself to technological mediocrity.
Western users tend to perceive the biases of American models as "normal" because those are their own cultural reference points. The biases of Chinese models strike them as "manipulative" because they are foreign. The reality is that all models have structural biases aligned with their geography and jurisdiction of origin. The question is not "which one is neutral" (none of them are), but "which one best matches my values and my information needs".
What to do in practice in 2026?
For an SME, a developer, or an individual user, here are the practical implications.
If you use DeepSeek or Qwen. Be aware that your queries on topics sensitive to China will be filtered or steered. For everything else (code, analysis, creative writing), these models are often excellent and far cheaper than their American equivalents. The quality-to-price ratio can justify using them for non-political cases.
If you use ChatGPT or Claude. Be aware that these models also have biases, notably American-centric ones, and tend to avoid political topics, sometimes excessively. Do not assume they are "neutral" because they are Western.
For sensitive data. Any use of a model via API means sending your data to servers under foreign jurisdiction. China's 2017 intelligence law obliges Chinese companies to cooperate with the government. For genuinely confidential data (strategic contracts, intellectual property), self-hosting (running it yourself on your own servers) an open source model locally remains the only truly safe option.
For thinking about the subject. Never use a single model to form an opinion on controversial topics. Cross-reference sources, compare answers, and bear in mind that every model reflects the jurisdiction and ideology of its creators. This simple precaution is enough to neutralise 90% of biases.
The final irony
The great paradox of this story is probably the following. China's strategic openness (open source models, permissive licences, low prices) has done more to democratise access to AI in poor countries than all Western promises about "AI for the common good". A student in Lagos can download Qwen for free and run a near-frontier-level model on a laptop. The same is impossible with ChatGPT.
This situation is uncomfortable for the Western narrative, which likes to set "liberal democracies" against "authoritarian regimes" on the technological battlefield. The reality is more complex: China practises political control at home AND technological liberalisation abroad. The United States practises political freedom at home AND technological restriction abroad (through sanctions, export controls, model pricing).
This analysis is part of our series on the blind spots of AI in 2026. To go further, you can read our article on the limits of regulation, which directly extends this reflection on what states can still do, and our analysis of the hidden economy of AI companions, which reveals another paradox between regulation and mass adoption.