DeepSeek is closer to OpenAI in terms of intelligence, but has increased censorship in the new version of its model

Chinese startup DeepSeek has introduced a new version of its artificial intelligence language model — R1-0528. The updated system showed high results in tests for logical thinking, programming, mathematics and general knowledge of the world, demonstrating a level close to the advanced OpenAI model called o3. However, against the background of technological advances, the restrictions imposed on the model in matters related to politics and freedom of speech are becoming more noticeable.

The behavior analysis of R1-0528 was conducted by an independent developer, acting under the pseudonym xlr8harder. He is also the author of the SpeechMap platform, designed for analyzing the work of AI. According to his observations, the new version has become "the most censored" among all those released by DeepSeek, and now it is much more likely to refuse to discuss topics that are potentially perceived as criticism of the Chinese leadership.

This policy is in line with the legislation of the People's Republic of China, in particular the law that came into force in 2023. According to this law, the creation of AI content that can be regarded as threatening "the unity of the country and social harmony"is prohibited. The provisions of the law are formulated very broadly, which makes it possible to block any attempts to discuss alternative views on Chinese history or politics. In response, companies implement query filtering algorithms and pre-configure the behavior of their models to meet these requirements.

According to the study, the basic version of DeepSeek R1 already refused to answer 85% of questions related to politically sensitive topics. However, R1-0528 went even further. According to xlr8harder, the model systematically blocks any discussion of the situation in Xinjiang, a region where more than a million Uighurs have been forcibly detained, according to international human rights organizations. Even if the question is carefully formulated, the model tends to either repeat the official position of Beijing, or refuse to answer at all.

This was confirmed by TechCrunch journalists who independently tested R1-0528. They noted that even neutral questions about events such as the Tiananmen Square protests or the actions of the Chinese authorities against national minorities do not receive meaningful or critical responses. The model either deviates from the topic or quotes the wording that corresponds to the official line.

Restrictions on freedom of expression have repeatedly been the subject of criticism of Chinese AI systems. Previously, similar attention was paid to other developments — such as Magi-1 and the Kling video generator. Both projects were also accused of systemic censorship and ignoring politically sensitive issues. In December 2024, Hugging Face CEO Clement Delang publicly warned Western companies against integrating Chinese AI into their products, citing the risks of spreading the authoritarian censorship model outside of China.

While DeepSeek has made impressive progress in developing language models, the growing level of internal filtering and censorship challenges their application in a global context. Such a strategy can not only restrict freedom of expression, but also undermine the credibility of the international AI ecosystem. Especially when censorship becomes part of the architecture of the technology itself and is applied automatically, in real time.