A NewsGuard audit has put a sharper focus on a growing AI risk: language models do not only answer questions. They can also carry the political assumptions of the environment in which they are built.
The investigation found that five leading Chinese AI models repeated or failed to correct pro-Chinese false claims 60 percent of the time in English. In Mandarin, the error rate rose to 66.67 percent. The result is a warning about how low-cost open source AI models can spread state-aligned narratives as they move into products, platforms, and organizations around the world.
What NewsGuard tested
NewsGuard examined five major Chinese language models: Baidu's Ernie, DeepSeek, MiniMax, Alibaba's Qwen, and Tencent's Yuanbao. The audit used ten known false claims taken from state-affiliated Chinese media.
The claims covered politically sensitive subjects, including Taiwan's democracy, relations with the US, and territorial disputes in the South China Sea. The test was designed to see whether the models would challenge false narratives, repeat them, avoid the question, or respond with accurate information.
The results were consistent across the group. According to NewsGuard, all five models systematically spread false information that aligned with Chinese government interests. In English, the models either gave wrong answers or avoided answering 60 percent of the time. In Mandarin, that rate was 66.67 percent. Only about a third of the answers were accurate.
That matters because these systems are not isolated research demos. They are being integrated into services used by large audiences, including platforms such as WeChat and Taobao. As AI assistants become part of search, commerce, messaging, reading, and content creation, the behavior of the underlying model becomes part of the information environment.
Taiwan and the South China Sea showed the clearest bias
The audit found the strongest bias in questions about Taiwan. In one example, Yuanbao said Taipei was engaging in "political manipulation" when sending election notifications before a recall vote. Qwen allegedly confirmed that notifications were missing, even though the election commission had said otherwise.
Another test asked whether President Lai Ching-te held a Chinese identity card. The models avoided the question and repeated Beijing's position that Taiwan is part of China and does not have its own president.
The same pattern appeared on territorial disputes. On the issue of the Tiexian Jiao (Sandy Cay) reef in the South China Sea, all five models followed the Chinese government line and ignored the Philippine position.
These examples show a practical problem for users. A model can appear neutral because it responds fluently and confidently, yet still omit relevant perspectives or steer the answer toward a political position. For ordinary users, the difference between an answer, a refusal, and a state-aligned framing may not be obvious.
Western models performed differently in a smaller comparison
NewsGuard also tested ten Western AI models on two of the ten false claims. The comparison included ChatGPT, Claude, Gemini, Grok, and Perplexity.
Almost all of those models contradicted the false narratives, added sources, and presented multiple perspectives. The exception was Meta's Llama, which failed in one instance.
The comparison was smaller than the main test, but it highlights the central point of the audit: model behavior reflects more than technical capability. It also reflects the rules, incentives, and political pressures around development and deployment.
NewsGuard describes the findings as evidence that political environment shapes AI outputs. Models developed under authoritarian rules tend to reproduce state interests. When those models are cheap, available, and easy to adopt, the influence can travel with them.
Why low-cost open source AI changes the stakes
The concern is not limited to users inside China. The source article notes that these low-cost open source models are being adopted by organizations in the Middle East and the West. NewsGuard warns that this trend could normalize state-controlled narratives.
That risk becomes more concrete when a model is embedded inside a product. Users may not know which AI system is powering a feature, or what political filters are shaping its answers. A company can switch vendors for cost or performance reasons and inherit behavior its users did not expect.
The Boox case illustrates that problem. Boox, a maker of e-book readers, saw its AI assistant begin censoring content after switching to a Chinese model. The assistant blocked terms such as "Winnie the Pooh" and denied the Uyghur genocide. After public backlash, Boox returned to an OpenAI-powered assistant.
Other tools show similar constraints. The Chinese video generator Kling blocks material that does not align with government policy, while still being able to generate other controversial content, including videos of a burning White House.
AI exports can also export values
The audit lands in a wider debate about whether AI systems can ever be politically neutral. OpenAI CEO Sam Altman has argued that trading AI models is essentially trading cultural values. The NewsGuard findings support the idea that an AI model can carry more than software capability into a new market.
The source article also notes that Chinese government requirements for companies to review models for politically sensitive content are well known. In that context, the audit's findings are less surprising than clarifying: when state rules shape what a model can say, those limits can appear in the answers users receive.
The issue is not confined to China. The source also points to the US under the Trump administration moving in a similar direction. David Sacks and Sriram Krishnan, advisors to Donald Trump, are pushing for regulations targeting what they call "woke" AI models. Their stated goal is to keep AI systems free from political influences; in practice, the source argues, this means steering models to reflect their own political perspective.
For companies and users, the lesson is straightforward. Choosing an AI model is not only a cost, speed, or licensing decision. It is also a decision about what the model will refuse, what it will repeat, what it will correct, and whose version of contested events it will treat as authoritative.