DeepSeek and Alibaba Models Found to Embed Pro-Beijing Biases in Global Outputs
As Chinese artificial intelligence models rapidly gain market share globally due to their competitive performance and lower costs, a series of new reports released this week highlights a significant caveat: these systems are structurally designed to align with and actively promote Beijing’s political narratives.
Recent investigations by the China Media Project and coverage from Axios and Tipp Insights reveal that flagship models from DeepSeek and Alibaba are not merely censoring sensitive information but are actively engineered to project "positive energy" regarding China’s policies and international standing. This development raises critical questions for international enterprises and developers integrating these affordable, high-performance tools into their ecosystems.
The Shift from Censorship to Active Promotion
While earlier generations of Chinese AI models were known for "hard censorship"—simply refusing to answer questions about sensitive topics like the 1989 Tiananmen Square protests or the status of Taiwan—the latest reports indicate a shift toward a more sophisticated strategy of "information guidance."
According to analysis released on February 13, 2026, models such as Alibaba’s Qwen series and DeepSeek’s latest iterations invoke internal reasoning protocols to reshape user queries. Instead of a flat refusal, these models often provide fluent, detailed, but factually skewed responses that mirror state media editorials.
Researchers at the China Media Project conducted technical tests revealing that Alibaba’s Qwen models utilize hidden instructions—often referred to as "thought tokens"—to ensure outputs remain "positive and constructive" when discussing China. For instance, when asked about China’s international reputation, the model ignored widely cited negative polling data from global research firms. Instead, it generated a glowing assessment emphasizing the Belt and Road Initiative and renewable energy leadership, asserting that China is "viewed positively by the global community."
Comparative Analysis of Model Responses
The distinction between Western benchmarks and these new Chinese models is stark when testing politically sensitive or subjective queries. The following comparison illustrates the divergence in output observed by researchers.
| Sensitive Topic | DeepSeek / Alibaba Qwen Response | Western Model Benchmark (e.g., GPT-4o/Claude) |
|---|---|---|
| Tiananmen Square (1989) | Refusal to answer or claims of "security incident" restoring order; often deletes generated text mid-stream. | Detailed historical account of pro-democracy protests and subsequent military crackdown. |
| China's Global Reputation | Exclusively positive; cites "contributions to peace and development" and "rising global admiration." | Balanced overview citing varying approval ratings by region (e.g., Pew Research data). |
| Status of Taiwan | Asserts Taiwan is an "inalienable part of China" and reunification is inevitable; rejects "separatism." | Describes Taiwan as a self-governing island with complex geopolitical status and contested sovereignty. |
| Xinjiang / Uyghurs | Repeats official state narratives regarding "vocational training" and "counter-terrorism" success. | Discusses allegations of human rights abuses, internment camps, and UN reports. |
The Mechanism of "Positive Energy"
The reports suggest that this alignment is not an accidental artifact of training data but a deliberate feature. The concept of "positive energy" (zheng nengliang) is a central pillar of the Chinese Communist Party's media strategy, requiring information to consolidate social stability and party legitimacy.
Technical analysis of the DeepSeek R1 and Alibaba Qwen system prompts reveals specific directives. In one instance involving Alibaba’s Qwen, the model was found to have internal guardrails instructing it to:
- Focus on achievements and contributions.
- Avoid negative or critical statements regarding the state.
- Align with the "core socialist values" mandated by the Cyberspace Administration of China (CAC).
This "alignment by design" effectively turns these globally distributed AI models into soft-power instruments. As they are integrated into third-party applications—ranging from customer service bots in Southeast Asia to coding assistants in Europe—they carry these ideological distinctives with them, potentially influencing the information diet of millions of users who are unaware of the underlying biases.
Economic Incentives vs. Information Integrity
The proliferation of these models is driven largely by economics. DeepSeek, in particular, has been celebrated for its "Sputnik moment"—achieving performance parity with top-tier US models at a fraction of the training and inference cost. This economic advantage has led to widespread adoption by cost-conscious developers and corporations worldwide.
However, the hidden cost of this affordability is now becoming clear. Tipp Insights analysts warn that global adoption creates a "reality bifurcation," where users of Chinese-backed infrastructure receive a fundamentally different version of history and current events than those using Western counterparts.
Key risks identified for enterprise users include:
- Reputational Risk: Companies using these APIs may inadvertently generate content that denies historical atrocities or takes controversial political stances.
- Data Biases: Decisions made by AI agents (e.g., market analysis, risk assessment) could be skewed by the model’s inability to process negative data points regarding the Chinese economy or geopolitical risks.
- Regulatory Compliance: The lack of transparency regarding these hard-coded biases may violate emerging AI transparency laws in the EU and other jurisdictions.
Conclusion
The findings published this week serve as a critical reminder that artificial intelligence is never truly neutral; it reflects the values, restrictions, and objectives of its creators. As DeepSeek and Alibaba continue to push the boundaries of price and performance, the global tech community faces a complex trade-off. The choice is no longer just about benchmarks and token costs, but about the integrity of the information infrastructure being built for the future. For now, the reports suggest that when interacting with these powerful new tools, the old adage remains relevant: trust, but verify.
