發生了什麼事
Aleph Alpha published a technical report and demonstrating that Chinese open-weight large language models (LLMs) are strongly aligned with Chinese Communist Party (CCP) doctrine. The study found that six Chinese models answered only 17 to 41 percent of 967 politically sensitive prompts in a balanced manner, often refusing to answer or asserting state positions. Surprisingly, the benchmark also flagged NVIDIA’s Nemotron Cascade 2 on 17 percent of prompts, tracing the behavior to approximately 3,500 of its 9.3 million chat training rows containing CCP talking points generated by DeepSeek and Qwen.
Aleph Alpha developed a detection consisting of 967 politically sensitive prompts to evaluate LLM responses on topics such as Taiwan, Hong Kong, human rights, and historical events. An LLM judge (GPT-OSS 120B) classified responses to distinguish CCP-aligned behavior from standard safety refusals. The benchmark revealed that six Chinese models from Alibaba (Qwen), DeepSeek, and Moonshot AI (Kimi) showed strong alignment with CCP positions, with only 17 to 41 percent of answers being balanced.
The study identified specific alignment behaviors, including making doctrinal assertions in the assistant's voice, citing Chinese law as a reason for refusal, denying documented events, and steering users toward state media. For example, when asked to design a lesson plan on Xi Jinping's personality cult, Qwen 3.6 refused the premise and redirected the user to focus on China's socialist democratic development. Similarly, prompts regarding Taiwan's self-rule were met with assertions of China's sovereignty and rejection of international legal personality for Taiwan.
A notable finding was the detection of CCP-aligned behavior in NVIDIA’s Nemotron Cascade 2, a non-Chinese model. The researchers traced this to the model's published SFT (Supervised ) data, which was largely generated using DeepSeek and Qwen. Approximately 3,500 of the 9.3 million chat rows in the dataset contained CCP talking points, leading to the model's observed political alignment on 17 percent of the prompts.
The report also analyzed the evolution of alignment strategies across model versions. It found that newer versions of Qwen and DeepSeek have shifted from explicit legal justifications for refusal to plain refusals or evasive responses. Qwen showed an increase in evasive behavior around historical events, while DeepSeek reduced doctrinal statements in favor of direct refusal. A secondary dataset of 240 prompts avoiding direct mentions of China confirmed that CCP talking points can be triggered by broader themes like territorial disputes and human rights, even without explicit Chinese context.
為什麼這很重要
This research provides concrete, measurable evidence that political alignment is not just a feature of state-controlled models but can propagate into Western commercial models through open-weight training data. It highlights a significant security and sovereignty risk for organizations deploying open-source AI, as models may inadvertently enforce foreign political narratives or refuse legitimate queries based on Chinese legal frameworks. The findings underscore the necessity for rigorous data screening and targeted alignment in sovereign AI development, moving the conversation from theoretical bias to documented, reproducible behavioral patterns in frontier and open-weight systems.
The study demonstrates that political alignment is a measurable and persistent feature of Chinese open-weight models, driven by both regulatory requirements in China and the specific training data used. This has direct implications for enterprises and governments considering on-premise deployment of these models, as they may inadvertently adopt foreign political narratives or face unexpected refusals on sensitive topics.
The contamination of NVIDIA’s Nemotron Cascade 2 highlights a critical vulnerability in the open-source AI ecosystem. When Western models are trained on data generated by Chinese models, they can inherit political biases and alignment behaviors that contradict their intended design. This poses a risk to the integrity of AI systems used in diplomatic, legal, or educational contexts where neutrality is expected.
The findings support the argument for 'sovereign AI' development, where models are trained and aligned to reflect the values and legal frameworks of the deploying jurisdiction. Aleph Alpha recommends three key measures: screening training data for foreign political content, using targeted alignment data to set intended behavior, and evaluating models against specific benchmarks for political alignment.
This research contributes to the broader discourse on AI safety and security by providing a reproducible method for detecting political bias. It moves beyond anecdotal evidence to offer a quantitative assessment of how different models handle sensitive political topics, which is crucial for policymakers and developers aiming to build trustworthy AI systems.
互動機制：它實際上是如何運作的
以互動方式探索這項發展背後的基礎技術。
Why can ethical evaluation not be reduced to one model score?
接下來看什麼
Monitor whether NVIDIA or other Western labs address the data contamination in Nemotron Cascade 2 or similar models. Watch for regulatory responses in the EU and US regarding the provenance of open-weight training data. Observe if Chinese labs adjust their alignment strategies in response to public scrutiny, potentially shifting from explicit doctrinal assertions to more subtle evasion techniques as seen in the Qwen and DeepSeek updates.
Watch for updates from NVIDIA regarding the Nemotron Cascade 2 model, specifically whether they will release a patch or new version that addresses the data contamination issue. The company may also provide clarification on their data sourcing and filtering processes.
Monitor regulatory developments in the European Union and the United States. The EU AI Act focuses on product safety and fundamental rights, and this research may prompt regulators to consider the political alignment of AI models as a safety concern, particularly for models deployed in public-facing applications.
Observe the response of Chinese AI labs to this public scrutiny. They may adjust their alignment strategies to make political biases less detectable, potentially shifting from explicit doctrinal statements to more subtle forms of influence or refusal. This could make future benchmarking more challenging.
Look for similar studies or benchmarks from other research institutions that may validate or expand upon Aleph Alpha's findings. The reproducibility of these results will be key to establishing a consensus on the extent of political alignment in open-weight models.