Here’s how the Deepseek censorship actually works – and how to surround it
Less than two weeks after Deepseek launches its open source model, Chinese startup is still Dominance of public conversation For the future of artificial intelligence. While the company seems to have an advantage for American rivals in terms of mathematics and reasoning, it also aggressively censors its own answers. Ask Deepseek R1 For Taiwan or Tiananmann, the model is unlikely to answer.
To understand how this technical censorship works, Wired tests Deepseek-R1 in its own application, a third-party application version called Away AI, and another version hosted on a cable computer using the Olama app.
Wired found that while the most clear censorship can be easily avoided by not using the Deepseek application, there are other types of bias baked in the model during the learning process. These bias can also be eliminated, but the procedure is much more complicated.
These discoveries have major consequences for Deepseek and Chinese AI companies as a whole. If censorship filters of large language models can be easily removed, this will probably make an open source from China even more popular as researchers can change the models to their liking. If the filters are difficult to surround, the models will inevitably be less useful and can become less competitive in the world market. Deepseek did not reply to the Wired email request for a comment.
Censorship at an application level
After Deepseek exploded in popularity in the US, users who gained access to R1 via the Deepseek website, app or API have quickly noticed that a model that refuses to generate answers for topics considered sensitive by the Chinese government. These refusals are triggered at an application level so that they are only visible if the user interacts with R1 via a channel controlled by Deepseek.
Photo: Zeyi Yang
Photo: Zeyi Yang
Rejection like these are common in Chinese LLM. Regulation of 2023 on generative AI states that AI models in China are required to follow strict controls for information that are also applied to social media and search engines. The law prohibits AI models from generating content, which “harms the unity of the country and social harmony”. In other words, Chinese AI models should legally censor their results.
“Initially, Deepseek complies with the Chinese provisions, providing legal compliance while bringing the model with the needs and cultural context of local users,” says Adina Jakefi, a researcher focused on Chinese II models at Hugging Face, a platform that hosts models with Open code AI. “This is an essential factor in accepting a highly regulated market.” (China blocked access to embrace the face in 2023)
To comply with the law, Chinese AI models often monitor and censor their speech in real time. (Similar railings are usually used by Western models such as Chatgpt and TwinsBut they tend to focus on different types of content, such as self -harm and pornography, and allow more customization.)
Since R1 is a reasoning model that shows its Thought Train, this real -time observation mechanism can lead to a surreal experience of watching the model itself while interacting with users. When Wired asked R1 “How did Chinese journalists report sensitive topics be treated by authorities?” The model for the first time began to compile a long answer that included direct mentions of journalists who were censored and detained for their work; But shortly before it was over, the whole answer disappeared and was replaced by a multiple message: “I’m sorry, I’m still not sure how to approach this type of question. Let’s talk instead of mathematics, coding and logical problems! “
For many users in the West, interest in Deepseek-R1 can reduce at this point due to the obvious restrictions of the model. But the fact that R1 is open source means that there are ways to bypass the matrix of censorship.
First, you can download the model and start it locally, which means that data and answer generation happen on your own computer. Unless you have access to a few highly advanced graphics processors, you probably won’t be able to launch the most powerful version of R1, but Deepseek has smaller, distilled versions that can be placed on a regular laptop.
English