Does DeepSeek Censor Its AI Answers? On These Sensitive Topics, Yes.

[pirat@lemmy.world]

You're asking the real important questions!

[bell@lemmy.world]

AI already has to deal with hallucinations, throw in government censorship too and I think it becomes even less of a serious, useful tool.

[fuzzy_feeling@programming.dev]

please let the username not check out...

[aboubenadhem@lemmy.world]

Making the censorship blatantly obvious while simultaneously releasing the model as open source feels a bit like malicious compliance.

[ulrich@feddit.org]

They're all censoring answers. They get flack every day for not censoring enough of them.

[banshee@lemmy.world]

Just to clarify - DeepSeek censors its hosted service. Self-hosted models aren't affected.

[yingwu@lemmy.dbzer0.com]

I ran Qwant by Alibaba locally, and these censorship constraints were still included there. Is it not the same with DeepSeek?

[b161@lemmy.blahaj.zone]

It just returns an error when I ask if Elon Musk is a fascist. When I ask about him generally it’s just returns propaganda.

[banshee@lemmy.world]

I think we might be talking about separate things. I tested with this 32B distilled model using llama-cpp.

[robottoaster@mander.xyz]

Even the Chinese AI is better at maths.

[lorips@lemmy.world]

Deepseek 2 is censored locally, had a bit of fun asking him about China 1989
(Running locally using Ollama with Alpaca as GUI)

[tyler@programming.dev]

On another person who’s actually running locally. In your opinion, is r1-32b better than Claude sonnet 3.5 or OpenAI o1? IMO it’s been quite bad, but I’ve mostly been using it for programming tasks and it really hasn’t been able to answer any of my prompts satisfactorily. If it’s working for you I’d be interested in hearing some of the topics you’ve been discussing with it.

[lorips@lemmy.world]

R1-32B hasn't been added to Ollama yet, the model I use is Deepseek v2, but as they're both licensed under MIT I'd assume they behave similarly. I haven't tried out OpenAI o1 or Claude yet as I'm only running models locally.

[bytejunk@lemmy.world]

I haven't looked into running any if these models myself so I'm not too informed, but isn't the censorship highly dependent on the training data? I assume they didn't release theirs.

[aboubenadhem@lemmy.world]

Video of censored answers show R1 beginning to give a valid answer, then deleting the answer and saying the question is outside its scope. That suggests the censorship isn’t in the training data but in some post-processing filter.

And even if the censorship were at the training level, the whole buzz about R1 is how cheap it is to train.

[tyler@programming.dev]

Hmm I’m using 32b from ollama, both on windows and Mac.

[banshee@lemmy.world]

Interesting. I wonder if model distillation affected censoring in R1.

[psmgx@lemmy.world]

I thought this was a big deal because it's Open Source -- is it not possible to see what is causing these blocks and censorings?

[imgonnatrythis@sh.itjust.works]

I'm still hazy on how open source it really is. Even if you ask it, it will tell you v3 is, and I quote "not open source". There is a github repository so there is some code available, but I get the sense that open-source is being used as a bit of a teaser here and that for-profit licensing is likely where this is headed.
Even if the code was fully available, I suspect it could take weeks to find the censorship bits.

[lorips@lemmy.world]

Ah, I just found it.
Alpaca is just being weird again.
(I'm presently typing this while attempting to look over the head of my cat)