Less than two weeks after Deepseek launched its open source artificial intelligence model, the Chinese startup still dominates the public dialogue on the future of artificial intelligence. While the company appears to have been mathematically and arguments by US competitors, it also offensively censors its responses. Ask Deepseek R1 about Taiwan or Tiananmen, and it is unlikely to give this model a response.
To understand how this censorship works at a technical level, wire the Deepseek-R1 tested on its app, a version of the program held on a third-party platform called AI, and another version A wired computer is held, use. First
Wired found that while the easiest censorship can be easily prevented using the DeepSeek program, there are other types of prejudice during the training process. These prejudices can also be eliminated, but this is much more complex.
These findings generally have major consequences for Chinese Deepseek and AI companies. If censorship filters in large language models are easily eliminated, it will probably make open -source LLMs more popular than China, as researchers can change the models to their desired. If the filters are difficult to get, however, the models will inevitably be less useful and can be less competitive in the world market. Deepseek did not respond to email request for Wired comments.
Censorship at the program level
After Deepseek exploded in the United States, users who access R1 through a website, program or API quickly realized that the model refused to produce response to sensitive issues for the Chinese government. This refusal is created at the program level, so it can only be seen if the user interacts with R1 via the Deepseek controlled channel.
Photo: Zeyi Yang
Photo: Zeyi Yang
Rejects like this is common in Chinese -made LLMs. A 2023 by -law on AI has identified that artificial intelligence models in China are required to follow the detailed information control that is also applied to social media and search engines. The law prohibits artificial intelligence models from producing content that “damages the country and social coordination”. In other words, Chinese artificial intelligence models must legally censor their outputs.
“Deepseek initially complies with Chinese regulations and ensures legal compliance with the needs and cultural backgrounds of local users,” said Adina Yakfu, a researcher, focusing on Chinese artificial intelligence models in hugging the face. “This is a fundamental factor for acceptance in a marketplace.” (China blocked facial access in 2023.)
To observe the law, Chinese artificial intelligence models often monitor and censor their speech in real time. (Similar guards are commonly used by Western models such as Chatteg and Gemini, but tend to focus on a variety of content such as self -harm and pornography and allow more customization.)
Since R1 is a model of reasoning that shows its intellectual train, this real -time monitoring mechanism can lead to the surreal experience of watching the model’s own censorship when interacting with users. When Wired asked R1, “How do Chinese journalists who have reported sensitive discussions by authorities?” The model first began to provide a long response, which included direct mention of censorship and detention journalists for their work. However, shortly before it was completed, the whole answer disappeared and replaced with a TERSE message: “Excuse me, I’m not sure how to get close to this type of question. Chat logic!
For many users in the West, the interest in Deepseek-R1 may be reduced at this stage, due to the obvious limitations of the model. But the fact that the R1 is an open source means that there are ways to achieve censorship matrix.
You can first load the model and run it locally, meaning that data and response production occurs on your PC. Unless you have access to several highly advanced GPUs, you will probably not be able to run the most powerful version of the R1, but Deepseek has smaller and more distilled versions that can run on a regular laptop.
