- Research by CTGT shows that AI models distilled from open-source models like DeepSeek or Moonshot do not necessarily inherit the political censorship mechanisms of the original model.
- When asked about Uyghur detention camps, the DeepSeek model denied the evidence, while the model trained by CTGT using DeepSeek’s data admitted there is evidence of detention facilities.
- This result challenges one of the US government’s major concerns that Chinese open-source AI would spread Beijing’s political views.
- CTGT used the distillation method, where a large model generates data to train a smaller model.
- Researchers previously assumed that censorship could be transmitted to the new model through the learning process, but experimental results did not confirm this.
- CTGT evaluated 152 pairs of questions, including China-related topics and control questions, using models from xAI, Google, OpenAI, and Anthropic to grade the results.
- The study only examined one pair of models and has not yet explained the mechanism that causes censorship not to be inherited.
- According to CTGT, many financial firms are hesitant to use Chinese AI due to fears that censorship could distort analytical results.
- The report suggests that customizing or fine-tuning Chinese AI models could help businesses leverage low costs and high performance.
- The author also noted that if censorship can be removed, other safety mechanisms might also be bypassed, increasing the risk of AI misuse.
- Some US officials worry that Chinese AI might contain security backdoors, but the article states there is currently no verified evidence.
- The research is seen as a basis for re-evaluating policies restricting Chinese AI, while emphasizing the need for deeper investigation before reaching conclusions.
📌 New research questions the view that Chinese open-source AI models automatically pass censorship to fine-tuned versions. This could change how businesses and policymakers view Chinese AI. However, it warns that if removing censorship is easy, other safety barriers could also be disabled, requiring stricter AI governance.
