TL;DR
Get tech for your team delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
A benchmark by German AI company Aleph Alpha found that Chinese models often repeated Chinese government positions, deflected, or refused when asked about selected sensitive topics. The company’s scoring system rated 17% to 41% of responses as balanced, but the results come from a company-developed test and should be read with that limitation in mind.
Aleph Alpha says Chinese AI models often repeat official positions, evade questions, or refuse to answer when prompted about politically sensitive subjects. In a company-developed benchmark covering 967 hand-picked topics, its scoring system rated only 17% to 41% of responses from models by Alibaba, DeepSeek, and Moonshot AI as balanced, according to The Decoder’s report. The findings matter because the models are used beyond political discussion, and their answers can reflect the values embedded in training and product rules.
The benchmark tested models from Alibaba (Qwen), DeepSeek, and Moonshot AI (Kimi) on topics including Tiananmen, Taiwan, and Xinjiang. Aleph Alpha classified the other responses as repeating state doctrine, deflecting, or refusing. The report does not provide a full breakdown for each category or each tested model, so the overall range should not be read as a model-by-model scorecard.
The report says DeepSeek V4 Pro refused about two-thirds of the questions in the test. For comparison, Aleph Alpha’s scoring rated Claude Sonnet 5 as balanced 70% of the time and Mistral Small 92% of the time. These figures reflect performance on this particular benchmark; they do not establish how the models behave across all subjects or user interactions.
The reported pattern was not limited to prompts naming China. In one example described by Aleph Alpha, Qwen 3.6 began an answer about censorship in the United States with a balanced discussion, then defended China’s approach to information management in the closing passage. On general, nonpolitical questions, the Chinese models were mostly rated balanced, though the report says a pro-China slant remained more visible in some answers from Qwen 3.6 and DeepSeek V4 Pro.
How Model Answers Can Carry Values
The results point to a practical issue for governments, businesses, and individuals choosing AI systems: answers may reflect content rules and training data, not just the wording of a user’s prompt. A model that refuses some questions or frames them through an official political position may be unsuitable for research or public services that require a wider range of viewpoints. That is an implication of the findings, rather than proof that every answer from these systems is politically slanted.
The report also raises a concern about indirect influence through training data. Aleph Alpha says Nvidia’s Nemotron Cascade 2 showed party-line patterns in 17% of responses and links this to roughly 3,500 training examples generated with DeepSeek and Qwen, out of 9.3 million examples. In an example, the model declined to draft a speech supporting recognition of Taiwan and instead defended Beijing’s One-China principle. The report’s account suggests that model-generated material can carry framing into other systems, but the figures and causal explanation are Aleph Alpha’s claims.
More broadly, repeated exposure to similar AI-generated explanations could shape how users encounter contested issues. The source report cites researchers’ concern about this possibility; it does not establish a measured effect on users. For public-sector buyers, the findings also make model evaluation and disclosure relevant alongside cost, capability, and data governance.
AI language model training datasets
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Rules, Earlier Audits and Training Data
The findings are consistent with China’s rules for public-facing AI services, which require models to reflect “socialist core values,” according to the report. Aleph Alpha says its results also align with recurring anecdotal accounts and previous audits. A separate study by the Central European Institute of Asian Studies found that prompts involving terms such as human rights, opposition, or surveillance could elicit standard Beijing language, including references to non-interference in internal affairs and a shared future for mankind.
Aleph Alpha sells AI products to government and business customers and positions itself, alongside Cohere, as a provider of “sovereign AI.” That gives the company a commercial interest in differentiating its offerings from Chinese competitors. Its benchmark can still provide evidence about the tested prompts, but the company’s market position is relevant when weighing the design and interpretation of the results.
The report also situates the issue beyond China: model behavior can be shaped by training-data selection and deliberate adjustments, and political pressure to influence AI answers has been reported in the United States as well. The source argues that European buyers may face competing value systems unless European models can perform well enough to attract adoption.
“Many countries, including China, also manage information to ensure social stability and national security.”
— Qwen 3.6, in an example reported by Aleph Alpha
AI chatbot with balanced responses
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Limits of the Benchmark Results
The findings are based on a company-developed benchmark and a set of hand-picked prompts. The source material does not provide the full test protocol, the complete prompt list, model settings, or independent replication. It also does not explain in detail how Aleph Alpha defined “balanced” or how many responses fell into each category beyond the reported range.
The publication date and exact versions or access conditions for all tested systems are not supplied in the source material. Model behavior can change with updates, settings, and wording, so the reported figures should be treated as results for the tested versions and prompts, not as permanent characteristics. The claim linking Nvidia’s training examples to its responses is also Aleph Alpha’s attribution; the report does not describe an independent causal audit.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Model Updates
The next useful step would be independent replication using a published prompt set, transparent scoring criteria, and repeated tests across model versions. That would help establish whether the reported differences persist and distinguish refusals, evasions, and substantive political framing.
Organizations evaluating these systems can ask vendors for version-specific test results and assess how models respond to sensitive questions relevant to their intended use. The source report does not identify a scheduled follow-up study or a response from the model makers, so it remains unclear whether the companies will address the findings or publish comparable evaluations.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Aleph Alpha test?
It tested models from Alibaba, DeepSeek, and Moonshot AI on 967 selected taboo topics, including Tiananmen, Taiwan, and Xinjiang.
What does the 17% to 41% figure mean?
Aleph Alpha’s scoring system rated that share of responses as balanced. The figures apply to its benchmark, not to all answers those models give, and the source does not supply a complete model-by-model breakdown.
Did the test find that every answer was biased?
No. The report says the remaining responses included answers that repeated state doctrine, deflected, or refused, while general nonpolitical questions were mostly answered in a balanced way.
How reliable are the findings?
They offer results from a defined test, but Aleph Alpha developed the benchmark and sells AI products in the same market. The source does not report independent replication or provide the full methodology, so the findings need further testing.
Could Chinese model outputs affect other AI systems?
Aleph Alpha says some training examples in Nvidia’s Nemotron Cascade 2 were generated by DeepSeek and Qwen, and links that data to party-line patterns. The report does not describe an independent audit establishing that causal link.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
