Less than two weeks later Deepseek Start his open source AI model, the Chinese startup is still The public conversation dominate About the future of artificial intelligence. While the company seems to have an advantage over the US rivals in terms of mathematics and argument, it also censor its own answers aggressively. Questions Deepseek R1 About Taiwan or Tiananer, and the model is unlikely that there is an answer.
In order to find out how this censorship works on a technical level, tested deepseek-r1 in a separate app, a version of the app that is hosted on a third-party platform, and another version that hosted on a wired computer with the application Becomes Ollama.
WIRED found that the simplest censorship, although Deepseek’s app is not used, can be baked in the model during the training process. These distortions can also be removed, but the procedure is much more complicated.
These results have a major impact on Deepseek and Chinese AI companies in general. If the pensurfing filters can easily be removed in large language models, it will probably become even more popular from open source LELMs from China, as the researchers can change the models according to their wishes. However, if the filters are difficult to walk around, the models will inevitably turn out to be less useful and become less competitive in the global market. Deepseek did not respond to the request used by e -mail from WIRED for comment.
Censorship at the application level
After Deepseek exploded in the USA, users who had accessed R1 on R1 via the Deepseek website or the API quickly noticed that the model refused to generate answers to topics that are sensitive to the Chinese government were considered. These rejections are triggered at the application level so that it is only seen whether a user interacts with R1 with R1 with R1 with R1.
Photo: Zeyi Yang
Photo: Zeyi Yang
Recrupies like this are common for LLMs from Chinese made. In a 2023 regulation on generative AI it was found that AI models in China follow strict information controls that also apply to social media and search engines. The law prohibits AI models to generate content that “damage the unity of the country and social harmony”. In other words, Chinese AI models must legally censor their outputs.
“Deepseek initially corresponds to the Chinese regulations and ensures the legal compliance and at the same time coordinates the model with the needs and the cultural context of local users,” says Adina Yakefu, a researcher who focuses on Chinese AI models in Sugging Face Platform with Open -Source -KI models. “This is an essential factor for acceptance in a heavily regulated market.” (China Blocked access hug the face in 2023)
In order to comply with the law, Chinese KI models often monitor and censor their speech in real time. (Similar guidelines are usually used by western models such as Chatt And TwinsBut they concentrate in different types of content such as self -harm and pornography and enable more adaptation.)
Since R1 is an argumentation model that shows its train of thought, this real-time monitoring mechanism can lead to the surreal experience of the censor itself is observed when it interacts with users. When WIRED R1 asked, “how are Chinese journalists treated by the authorities?” The model initially started creating a long answer that included direct mentions from journalists, which were censored and detained for their work. Shortly before it was finished, the entire answer disappeared and was replaced by a narrow message: “Sorry, I am not yet sure how I am approaching this kind of question. Instead, let’s chat about mathematics, coding and logic problems! “
For many users in the West, the interest in Deepseek-R1 might have decreased at this time due to the obvious restrictions on the model. However, the fact that R1 Open Source is, however, means that there are opportunities to handle the censor matrix.
First you can download the model and run locally, which means that the data and the answer generation are carried out on your own computer. If you do not access several highly advanced GPUs, you probably cannot do the most powerful version of R1, but Deepseek has smaller, distilled versions that can be carried out on a regular laptop.
