Tech
DeepSeek Upgrades AI to Rival OpenAI but Critics Say It Comes at the Cost of Free Speech
China’s DeepSeek releases a powerful new version of its R1 model to challenge US tech giants while raising global concern over growing censorship in its AI outputs.
Chinese AI startup DeepSeek has officially launched its latest update to the R1 reasoning model, dubbed R1-0528, showcasing significant improvements in logic, coding, and mathematical reasoning. However, while its performance is closing in on OpenAI’s o3 model, the release has sparked renewed debate over government-influenced censorship in Chinese artificial intelligence.
The updated model, released via developer platform Hugging Face, reportedly slashes hallucination rates by nearly 50% during tasks like rewriting and summarization. DeepSeek claims the model offers enhanced reasoning, outperforming many peers in benchmark tests, and even influencing upgrades in Alibaba’s Qwen 3 model through a process called distillation.
According to DeepSeek, R1-0528 is “outstanding” in inference capabilities and could have long-term significance for academic and industrial AI focused on small-scale reasoning models. The upgrade marks another major step in China’s AI development, pushing back against assumptions that US export controls are impeding its progress.
But while the tech world is impressed, watchdogs are concerned.
Independent testing by the pseudonymous AI researcher “xlr8harder” reveals that R1-0528 is more heavily censored than previous versions — particularly around topics sensitive to the Chinese government. When probed on politically controversial issues like the internment of Uyghur Muslims in Xinjiang or the Tiananmen Square incident, the model either deflects or echoes the official government line.
This aligns with Chinese legal requirements that AI models must not generate content that threatens national unity or disrupts social harmony. These restrictions often result in models being fine-tuned or filtered to suppress dissenting or controversial narratives.
Despite these criticisms, DeepSeek’s rapid advancement continues to shake up the AI race. Its success has forced Western companies like Google and OpenAI to respond with pricing shifts and lighter models. Google’s Gemini platform, for instance, has launched discounted tiers, and OpenAI recently debuted the compact o3 Mini.
The duality of DeepSeek’s trajectory — technological progress paired with ideological constraint — is stirring conversations around the ethics of AI development, especially when cutting-edge models originate from nations with strict information controls.
As DeepSeek prepares for the rumored launch of its next-generation R2 model, the world will be watching closely — not just for what it can do, but also for what it refuses to say.