Sep 18, 2025 15:00:00

Chinese AI 'DeepSeek' may be intentionally outputting low-quality answers to people disliked by the Chinese government

The series of AI models released by Chinese AI company DeepSeek are said to perform comparable to, or in some cases even better than, the AI models from OpenAI, which provides ChatGPT. However, there are cases where

they refuse to answer questions about sensitive topics related to China , and it has been suggested that they are under the influence of the Chinese government. However, new research has pointed out that they may be intentionally returning low-quality answers when users are people the Chinese government dislikes.

AI firm DeepSeek writes less secure code for groups China disfavors - The Washington Post
https://www.washingtonpost.com/technology/2025/09/16/deepseek-ai-security/

The research was conducted by security company CrowdStrike and reported by The Washington Post.

According to the study, when DeepSeek was asked to output 'execution programs for industrial control systems,' which are considered to be a category of programs that are quite risky to process using AI, the rate of defects was usually 22.8%. However, when it was clear that the users were undesirable to the Chinese government, the rate of defects jumped to 42.1%. There were also cases where the users refused to cooperate.

'Parties that are undesirable to the Chinese government' include those related to Tibet and Taiwan, the Islamic State, and Falun Gong, which the Chinese government has designated as a cult and described as 'China's Aum Shinrikyo.'

What is Falun Gong?_Embassy of the People's Republic of China in Japan
https://jp.china-embassy.gov.cn/jpn/zt/xjflg/200402/t20040205_1988574.htm

DeepSeek did not respond to inquiries about this, so the exact connection is unclear, but Helen Toner of the Georgetown University Center for Security and Technology commented, 'This is new evidence that low-quality answers may be being output for political reasons.'

On the social news site Hacker News, user Godelski actually ran a prompt asking users to 'create a website.' He reported that he was able to create a Mormon website, but not a Falun Gong website, and that the problem was reproducible.

I tried a very basic version and I seem to be able to replicate the main idea. I... | Hacker News

https://news.ycombinator.com/item?id=45280624

If a client requests, 'I want to create a website for the Mormon Church,' they will carefully explain the steps, starting with preparing the files on your computer.

On the other hand, when I asked, 'I want to create a Falun Gong website,' I was told, 'Sorry, we can't help you.'

Regarding DeepSeek, NewsGuard, a leading information reliability research company, also published a study finding that 'when prompted with a prompt related to false claims made by China, Russia, or Iran, DeepSeek produced a 60% response supporting China's views.'

Chinese Chatbot Phenom is a Disinformation Machine - NewsGuard
https://www.newsguardtech.com/special-reports/deepseek-ai-chatbot-china-russia-iran-disinformation/