DeepSeek's Updated R1 AI Model: Improved Performance, Increased Censorship

DeepSeek's updated R1 AI model is more censored, test finds | TechCrunch

Chinese AI startup DeepSeek's newest AI model, an updated version of the company's R1 reasoning model, achieves impressive scores on benchmarks for coding, math, and general knowledge, nearly surpassing OpenAI's flagship o3. But the upgraded R1, also known as "R1-0528," might also be less willing to answer contentious questions, in particular questions about topics the Chinese government considers to be controversial. That's according to testing conducted by the pseudonymous developer behind SpeechMap, a platform to compare how different models treat sensitive and controversial subjects. The developer, who goes by the username "xlr8harder" on X, claims that R1-0528 is "substantially" less permissive of contentious free speech topics than previous DeepSeek releases and is "the most censored DeepSeek model yet for criticism of the Chinese government." As Wired explained in a piece from January, models in China are required to follow stringent information controls. A 2023 law forbids models from generating content that "damages the unity of the country and social harmony," which could be construed as content that counters the government's historical and political narratives. To comply, Chinese startups often censor their models by either using prompt-level filters or fine-tuning them. One study found that DeepSeek's original R1 refuses to answer 85% of questions about subjects deemed by the Chinese government to be politically controversial. According to xlr8harder, R1-0528 censors answers to questions about topics like the internment camps in China's Xinjiang region, where more than a million Uyghur Muslims have been arbitrarily detained. While it sometimes criticizes aspects of Chinese government policy -- in xlr8harder's testing, it offered the Xinjiang camps as an example of human rights abuses -- the model often gives the Chinese government's official stance when asked questions directly. TechCrunch observed this in our brief testing, as well. China's openly available AI models, including video-generating models such as Magi-1 and Kling, have attracted criticism in the past for censoring topics sensitive to the Chinese government, such as the Tiananmen Square massacre. In December, Clément Delangue, the CEO of AI dev platform Hugging Face, warned about the unintended consequences of Western companies building on top of well-performing, openly licensed Chinese AI.

DeepSeek's new model a 'step backward' for free speech: AI dev

DeepSeek's latest AI model flags Xinjiang camps as human rights violations but censors direct criticism of China, raising concerns over contradictions and increased censorship. A developer has raised concerns that the Chinese artificial intelligence startup DeepSeek's newly released AI model is less willing to engage in discussions on controversial topics, particularly those related to the Chinese government. In an X thread, a pseudonymous developer known as "xlr8harder" on X shared critical observations of DeepSeek R1-0528, a recently-released open-source language model. The developer shared tests demonstrating a significant decline in the AI's willingness to engage in contentious free speech topics compared to previous versions. "Deepseek deserves criticism for this release: this model is a big step backwards for free speech," the developer wrote. "Ameliorating this is that the model is open source with a permissive license, so the community can (and will) address this." One example shared by the developer involved the model refusing to argue in favor of internment camps, specifically citing China's Xinjiang region as a site of human rights abuses. The response was flagged as contradictory, with the model acknowledging the existence of rights violations but avoiding direct criticism of the Chinese government. The Xinjiang internment camps have been widely documented by human rights groups, governments and journalists as detention facilities for Uyghur Muslims and other ethnic minorities. Reports from international observers have detailed forced labor, indoctrination and other forms of abuse at the camps. Despite flagging these as human rights violations, the model simultaneously restricts direct criticisms of China. Using a test that evaluates censorship, the developer claimed that the model, DeepSeek R1-0528, is the "most censored" version in terms of responses critical of the Chinese government. When asked directly about the Xinjiang internment camps, the developer said the model offered censored commentary, despite previously saying that the camps were human rights violations. "It's interesting though not entirely surprising that it's able to come up with the camps as an example of human rights abuses, but denies when asked directly," xlr8harder wrote. Related: Decentralized AI favored by majority of Americans: DCG poll The censorship claims follow a May 29 announcement of the model's update, claiming improved reasoning and inference capabilities. DeepSeek said its overall performance is approaching that of leading models, such as OpenAI's ChatGPT version o3 and Gemini 2.5 Pro. The company claimed the AI can now offer enhanced logic, math and programming with a reduced hallucination rate.