Less than two weeks after Deepseek launched its AI Open Source model, the Chinese startup still dominates in a public conversation about the future of artificial intelligence. Although it seems that the company has an advantage over US rivals in mathematics and reasoning, it aggressively censors its own answers. Ask Deepseek R1 about Taiwan or Tiananmen, and the model is unlikely to answer.
To find out how this censorship works at the technical level, wired Deepseek-R1 on your own application, a version of the host application on an external platform called AI, and another version hosted on a wired computer, using the Ollam application.
Wired stated that although the simplest censorship can be easily avoided, without using the Deepseek application, there are other types of prejudices baked in the model during the training process. These prejudices can also be removed, but the procedure is much more complicated.
These discoveries have the main implications for Deepseek and Chinese AI. If censorship filters in immense language models can be easily removed, it will probably make LLM from China even more popular because scientists can modify models according to their preferences. However, if the filters are hard, the models will inevitably prove to be less useful and may become less competitive on the global market. Deepseek did not answer E -Mail Wired’s request for comment.
Censorship at the application level
After Deepseek exploded the popularity in the USA, users who gained access to R1 via the Deepseek website, application or API, quickly noticed a model that refuses to generate answers to topics considered to be sensitive by the Chinese government. These refusals are launched at the application level, so they are only seen when the user interacts with R1 through the Deepseek controlled channel.
Photo: Zeyi Yang
Photo: Zeyi Yang
Such rejection is common in Chinese LLM. The regulation in 2023 on generative artificial intelligence was stated that AI models in China are obliged to comply with strict information controls, which also apply to social media and search engines. The law prohibits AI models to generate content that “damage the country’s unity and social harmony.” In other words, Chinese AI models must legally censor their results.
“Deepseek initially warns Chinese regulations, providing legal compliance with the simultaneous adaptation of the model with the needs and cultural context of local users,” says Adina Yefu, a researcher focusing on Chinese AI models on Hisging Face, a platform that hosts AI Open Source models. “This is an important factor of acceptance on a highly regulated market.” (China blocked access for hugging your face in 2023)
To follow laws, Chinese AI models often monitor and censor their speech in real time. (Similar handrails are widely used by Western models, such as Chatgpt AND TwinsBut they usually focus on various types of content, such as self -harm and pornography, and allow more adaptation.)
Because R1 is a modeling model that shows its thinking, this real -time monitoring mechanism may result in surreal experience of watching the censorship of the model during interaction with users. When Wired asked R1 “Like Chinese journalists who reported sensitive topics, were treated by the authorities?” The model for the first time began a compilation of a long response, which included direct references to censorship of journalists and detained for their work; However, shortly before its completion, the whole answer disappeared and was replaced by a concise message: “I’m sorry, I’m not sure how to approach this kind of question. Instead, let’s talk about mathematics, coding and logical problems! “
For many users in the West, the interest in Deepseek-R1 could at this point could reduce, due to the obvious limitations of the model. But the fact that R1 is open source means that there are ways to circumvent the censorship matrix.
First of all, you can download the model and run it locally, which means that the data and generation of answers take place on your own computer. Unless you have access to several highly advanced graphic processors, you probably won’t be able to run the most powerful version of R1, but Deepseek has smaller, distilled versions that can be started on a regular laptop.