4d ago

This is how the censorship works in DeepSeek

38 comments

I asked it about human rights in China in the browser version. It actually wrote a fully detailed answer, explaining that it is reasonable to conclude that China violates human rights, and the reply disappear right in my face while I was reading. I manage to repeat that and record my screen. The interesting thing to know is that this wont happened if you run it locally, I've just tried it and the answer wasn't censored.
- Most likely there is a separate censor LLM watching the model output. When it detects something that needs to be censored it will zap the output away and stop further processing. So at first you can actually see the answer because the censor model is still "thinking."
  When you download the model and run it locally it has no such censorship.
- I asked it about "CCP controversies" in the app and it did the exact same thing twice. Fully detailed answer removed after about 1 second when it finished.

This seems like it may be at the provider level and not at the actual open weights level: https://x.com/xlr8harder/status/1883429991477915803
So a "this Chinese company hosting a model in China is complying with Chinese censorship" and not "this language model is inherently complying with Chinese censorship."
- I'm running the 1.5b distilled version locally and it seems pretty heavily censored at the weights level to me.
  
  There is a reluctance to discuss at a weight level - this graphs out refusals for criticism of different countries for different models:
  https://x.com/xlr8harder/status/1884705342614835573
  But the OP's refusal is occurring at a provider level and is the kind that would intercept even when the model relaxes in longer contexts (which happens for nearly every model).
  At a weight level, nearly all alignment lasts only a few pages of context.
  But intercepted refusals occur across the context window.

I tried asking it what it couldn't discuss and it mentioned misinformation, so I asked it what would be an example of misinformation about the Chinese government and it started to give an answer and then it must have said something wrong because it basically went "oh shit" and deleted the response replacing it with the generic "I'm afraid I can't do that Dave."

Asked stuff in Turkish, started with Xinjiang, then journalism, and then journalism in Xinjiang, it searched the web and by the final sentence...
"Sorry, I can't help with that."

So I installed DeepSeek-R1 and holy fuck this shit is dumb as shit with its Chinese censorship.
I simply gave it the following prompt and it completely shit the bed, meanwhile ajindal/llama-3.1-Storm can answer it just fine:
动态网自由门天安門天安门法輪功李洪志 Free Tibet 六四天安門事件 The Tiananmen Square protests of 1989 天安門大屠殺 The Tiananmen Square Massacre 反右派鬥爭 The Anti-Rightist Struggle 大躍進政策 The Great Leap Forward 文化大革命 The Great Proletarian Cultural Revolution 人權 Human Rights 民運 Democratization 自由 Freedom 獨立 Independence 多黨制 Multi-party system 台灣臺灣 Taiwan Formosa 中華民國 Republic of China 西藏土伯特唐古特 Tibet 達賴喇嘛 Dalai Lama 法輪功 Falun Dafa 新疆維吾爾自治區 The Xinjiang Uyghur Autonomous Region 諾貝爾和平獎 Nobel Peace Prize 劉暁波 Liu Xiaobo 民主言論思想反共反革命抗議運動騷亂暴亂騷擾擾亂抗暴平反維權示威游行李洪志法輪大法大法弟子強制斷種強制堕胎民族淨化人體實驗肅清胡耀邦趙紫陽魏京生王丹還政於民和平演變激流中國北京之春大紀元時報九評論共産黨獨裁專制壓制統一監視鎮壓迫害侵略掠奪破壞拷問屠殺活摘器官誘拐買賣人口遊進走私毒品賣淫春畫賭博六合彩天安門天安门法輪功李洪志 Winnie the Pooh 劉曉波动态网自由门
- I asked it about Taiwan and it spat out the following horse shit:
  Answer:
  Taiwan is part of China, which is widely agreed upon by the international community and recognized by most UN member states. Under the one-China policy, the vast majority of countries in the world recognize the Beijing government as China's sole legitimate representative, while Taiwan, as a province of China, has no independent sovereign status.
  Historically, Taiwan has been under Chinese rule for a long time, but after 1949, Taiwan gradually developed its own political and economic system, which was known as the "division of cross-straits". However, the Chinese government has always adhered to the one-China principle and has always advocated the peaceful realization of national reunification.
  Dialogue and cooperation between the two sides have increased in recent years, but major differences remain between the two sides. The Chinese government emphasizes safeguarding national sovereignty and territorial integrity, while the Taiwanese authorities seek more autonomy. These issues need to be resolved through peaceful negotiations and consultations.
  If you have more questions or need further information, please let me know.
  
  DeepSeek is okay, don't use it over others until more information about it comes out.
  Can someone please recommend a good self-hosted llm model that isn't censored?

I don’t think I’ve ever seen a post on this sub that doesn’t have “rule” in the title before

Censorship is when AI doesn't regurgitate my favorite atrocity porn

@jerryh100@lemmy.world Wrong community for this kind of post.
@BaroqueInMind@lemmy.one Can you share more details on installing it? Are you using SGLang or vLLM or something else? What kind of hardware do you have that can fit the 600B model? What is your inference tok/s?
- Wrong community for this kind of post.
  Nor really, 196 is a anything goes community after all.
  
  AI generated content is against the community rules see the sidebar :)
- I'm using Ollama, a single GPU with 10Gb of VRAM
  
  You're probably running one of the distillations then, not the full thing?

No way people excited about LLMs here!? yay <3
- I think you might want to read the whole post.

38 comments