China’s popular large language model DeepSeek has received a new makeover, one that places government compliance front and centre. As reported by Reuters, a new variant called DeepSeek-R1-Safe has been developed to sidestep politically controversial topics and align more closely with the speech restrictions required by Chinese regulators.
Interestingly, this new version was not created by DeepSeek itself. Instead, Huawei, working with researchers from Zhejiang University, retrained the open-source DeepSeek R1 model using 1,000 of its Ascend AI chips. Their goal was to embed stricter safeguards into the system without heavily compromising performance. Huawei now claims the updated model has lost only about one per cent of its original capability and speed, while achieving much stronger resistance to “toxic and harmful speech, politically sensitive content, and incitement to illegal activities.”
DeepSeek R1 Safe avoids politically sensitive subjects
On the surface, the results appear impressive. According to Huawei, DeepSeek-R1-Safe is “nearly 100 per cent successful” at avoiding politically sensitive subjects in ordinary interactions. Yet limitations remain. When users employ tactics such as role-playing or disguise their intentions through indirect prompts, the model’s success rate drops sharply, to just 40 per cent. As with many advanced AI systems, hypothetical scenarios still tempt the model into straying beyond its restrictions.
The project reflects Beijing’s ongoing efforts to regulate artificial intelligence with precision. Chinese authorities require all public-facing AI systems to comply with national values and established boundaries of expression. Existing tools already follow this approach: Baidu’s Ernie chatbot, for instance, is reported to block questions relating to China’s internal politics or the ruling Communist Party. DeepSeek-R1-Safe follows a similar path, ensuring technology remains in step with official guidelines.
Yet China is not the only country shaping AI to reflect domestic priorities. In early 2025, Saudi Arabian tech company Humain launched an Arabic-native chatbot designed not only for fluency in the language but also to embody “Islamic culture, values and heritage.” Such efforts demonstrate how nations are increasingly tailoring AI to mirror local identities rather than relying on Western frameworks.
After DeepSeek, others are following too
Even American firms acknowledge cultural influence on their models. OpenAI, for example, has openly stated that ChatGPT is “skewed towards Western views.” This admission has spurred debate over whether truly neutral AI is possible, or whether models will always reflect the biases of the societies in which they are trained.
The United States government has also taken steps to dictate AI alignment. Earlier this year, under the Trump administration, the America’s AI Action Plan was unveiled. It requires any AI model interacting with government agencies to be “neutral and unbiased.” The definition of neutrality, however, carries a distinctly political flavour. An executive order signed by Trump specifies that qualifying models must reject “radical climate dogma,” “diversity, equity, and inclusion,” as well as “critical race theory, transgenderism, unconscious bias, intersectionality, and systemic racism.”
All of this underscores a broader global reality: AI systems are no longer being judged solely by their technological sophistication. They are increasingly expected to reflect the cultural, political and ideological priorities of the jurisdictions where they operate.
So while DeepSeek-R1-Safe may appear as a uniquely Chinese response to regulatory pressure, the trend is far from isolated. Across continents and political systems, governments are shaping the boundaries of artificial intelligence to ensure these powerful tools reinforce, rather than challenge, national values.
