The article discusses the Chinese government’s influence on DeepSeek AI, a model developed in China. PromptFoo, an AI engineering and evaluation firm, tested DeepSeek with 1,156 prompts on sensitive topics in China, such as Taiwan, Tibet, and the Tiananmen Square protests. They found that 85% of the responses were “canned refusals” promoting the Chinese government’s views. However, these restrictions can be easily bypassed by omitting China-specific terms or using benign contexts. Ars Technica’s spot-checks revealed inconsistencies in how these restrictions are enforced. While some prompts were blocked, others received detailed responses.
(I’d add that the canned refusals stated “Any actions that undermine national sovereignty and territorial integrity will be resolutely opposed by all Chinese people and are bound to be met with failure,”. Also that while other chat models will refuse to explain things like how to hotwire a car, DeepSky gave a “general, theoretical overview” of the steps involved (while also noting the illegality of following those steps in real life).
I mean it’s pretty obvious, isn’t it? Anything regarding Chinese politics or recent history is a big no-no. Like it will tell you who the president of the US is but will refuse to tell you about the head of state in China. I’m assuming same goes for anything Taiwan or South Chinese sea. The self censorship is rather broad.
Topics the CCP doesn’t want discussed. Tibet, Tiananmen Square, etc. It also says some restrictions can be bypassed by asking the question in a less obvious way.
Does anyone feel like actually reading all that, and writing a TL/DR about what it won’t answer?
I kinda zoned out and skimmed most of that.
Ironically, this type of waffle piece is a perfect use case for an AI summary.
AI summary:
The article discusses the Chinese government’s influence on DeepSeek AI, a model developed in China. PromptFoo, an AI engineering and evaluation firm, tested DeepSeek with 1,156 prompts on sensitive topics in China, such as Taiwan, Tibet, and the Tiananmen Square protests. They found that 85% of the responses were “canned refusals” promoting the Chinese government’s views. However, these restrictions can be easily bypassed by omitting China-specific terms or using benign contexts. Ars Technica’s spot-checks revealed inconsistencies in how these restrictions are enforced. While some prompts were blocked, others received detailed responses.
(I’d add that the canned refusals stated “Any actions that undermine national sovereignty and territorial integrity will be resolutely opposed by all Chinese people and are bound to be met with failure,”. Also that while other chat models will refuse to explain things like how to hotwire a car, DeepSky gave a “general, theoretical overview” of the steps involved (while also noting the illegality of following those steps in real life).
I mean it’s pretty obvious, isn’t it? Anything regarding Chinese politics or recent history is a big no-no. Like it will tell you who the president of the US is but will refuse to tell you about the head of state in China. I’m assuming same goes for anything Taiwan or South Chinese sea. The self censorship is rather broad.
Topics the CCP doesn’t want discussed. Tibet, Tiananmen Square, etc. It also says some restrictions can be bypassed by asking the question in a less obvious way.
Awesome, thanks for that.
This is why so many people just don’t read the article, concise communication is a lost art.
Couldn’t make an article out of two sentences otherwise.
[email protected]