DeepSeek’s updated R1 AI model is more censored, test finds

DeepSeek's updated R1-0528 AI model achieves near state-of-the-art performance in coding, math, and general knowledge benchmarks, rivaling OpenAI's o3. However, independent testing indicates the model exhibits increased censorship regarding topics sensitive to the Chinese government, including criticisms of its policies and historical narratives. This aligns with Chinese regulations mandating AI models avoid content that undermines national unity, raising concerns about the potential for bias and information control in openly available Chinese AI models.

Analysis

DeepSeek's newest AI model, R1-0528, demonstrates notable technological progress, achieving benchmark scores in coding, mathematics, and general knowledge that nearly rival OpenAI's o3. However, this advancement is coupled with significant concerns regarding increased censorship, as independent testing conducted by SpeechMap suggests R1-0528 is "substantially" less permissive of contentious topics, particularly those critical of the Chinese government, and is described as "the most censored DeepSeek model yet." This behavior aligns with China's 2023 law mandating AI models avoid generating content that could undermine national unity or social harmony, leading Chinese AI developers to implement content controls through methods like prompt-level filters or fine-tuning. The prior DeepSeek R1 model reportedly refused to answer 85% of questions on politically sensitive subjects, and R1-0528 appears to continue or intensify this trend, for instance, by censoring inquiries about the Xinjiang internment camps or frequently defaulting to official Chinese government narratives. This situation, also observed in other Chinese AI models like Magi-1 and Kling, has prompted warnings, such as from Hugging Face's CEO, about the unintended consequences and ethical considerations for Western entities utilizing or building upon these high-performing, yet heavily censored, openly licensed Chinese AI technologies, highlighting a tension between technological capability and information control.

Key Decisions for Investors

• Investors should critically evaluate AI models originating from China, such as DeepSeek's R1-0528, for embedded censorship and alignment with state narratives despite their technical capabilities, as these factors present potential ethical, reputational, and operational risks.
• Companies considering the integration of or building upon openly licensed Chinese AI models must conduct thorough due diligence regarding potential content restrictions and geopolitical implications, which could impact global applicability and brand perception.
• The rapid advancements in Chinese AI, exemplified by R1-0528's performance, signal a shifting competitive landscape; however, investors must also weigh the significant influence of state regulation and information control on the development and deployment of these technologies.