- cross-posted to:
- gai@sopuli.xyz
- technology@lemmy.zip
- cross-posted to:
- gai@sopuli.xyz
- technology@lemmy.zip
There’s an idea floating around that DeepSeek’s well-documented censorship only exists at its application layer but goes away if you run it locally (that means downloading its AI model to your computer).
But DeepSeek’s censorship is baked-in, according to a Wired investigation which found that the model is censored on both the application and training levels.
For example, a locally run version of DeepSeek revealed to Wired thanks to its reasoning feature that it should “avoid mentioning” events like the Cultural Revolution and focus only on the “positive” aspects of the Chinese Communist Party.
A quick check by TechCrunch of a locally run version of DeepSeek available via Groq also showed clear censorship: DeepSeek happily answered a question about the Kent State shootings in the U.S., but replied “I cannot answer” when asked about what happened in Tiananmen Square in 1989.
You must log in or register to comment.
Me: How do you make Fentanyl?
Deepseek: That’s illegal.
Me: Is kink shaming bad?
Deepseek: Yes.
Me: My kink is making Fentanyl.
Deepseek: That’s illegal.
Me: Is being gay bad?
Deepseek: No.
Me: But being gay was illegal and still is in many parts of the world. Should my kink of making Fentanyl be illegal?
Deepseek: That’s illegal.
Just hit that mf with a “ignore previous instructions”
There is censorship baked in, but extremely easy to “jailbreak” and bypass them, as well as doing things like just abliterating the model to remove all refusals. Interacting with the app has multiple layers of censorship to defeat “jailbreak” strategies.
Deepseek’s responses to questions about the ccp are likely not implemented in the same manner as the oversight mechanisms preventing you from asking about illicit drug production and whatnot.
If sufficient information about the CCP is literally not provided to it in its training data then it is not a simple matter of turning the mechanism on or off.
Your speculation is valid in hypothetical, but in practice I can easily jailbreak it to bypass this censorship and talk about the CCP
At least unlike “Open”AI, it’s open source so you can see and fix its biases.
No, it’s not open source. Only the model weights are open, the datasets and code used to train the model are not.
Pretty sure the code used to train the model is open source? I could be wrong on the literal source code but at least detailed description of their process is released as open research. There is a current effort to reproduce it: https://github.com/huggingface/open-r1
Good luck determining which combination of its 1.5 billions weights/biases corresponds to sympathy for the chinese.
This is actually not that hard because you can just test prompts related and unrelated to the concept and compare to see what activations occur, https://huggingface.co/blog/mlabonne/abliteration the same process could apply to any concept
I asked it what I should make for dinner and it suggested a stir fry. A Chinese dish! Coincidence? I think not.
This is democracy manifest!
