DeepSeek censorship: 1984 "rectifying" in real time | Hacker News Reader

DeepSeek censorship: 1984 "rectifying" in real time

old.reddit.com

42 pointsorome21 comments

Open articleView on HN

When OpenAI does this, it's for "AI safety."

When a Chinese company does this, it's "literally '1984.'"

Which one is it?

The Chinese company also released the means to correct this though.

there's a video exactly like this (app changes text mid response) of people asking gemini if google is evil.

After a lot of thinking, DeekSeek said, "In summary, while Google provides significant value, its extractive tendencies and potential for harm mean it is leaning towards a more negative overall impact on society."

There is a difference between not telling someone how to make Sarin/age-gating certain topics and censoring works of fiction or references to politicians and anyone who claims there is not is so full of shit they need to have their stomach pumped.

Yet, search engines showed you that stuff from the very origins of the internet.

So what?

And?

"they did it so we're doing it to" is childish, worse than childish: infantile.

Looking into the social media footprints of people who are most vocal about censorship in various models it appears that for a great deal of them the primary problem they have is that they are being blocked from generating images of minors from japanese cartoons being raped.

Not wanting censorship isn't childish and "infantile."

An interesting question is whether AI can be manipulated effectively.

In other words, can I teach a state of the art useful model that believes 2+2=5

Sure you can get it to state 2+2=5 if asked, but there are a lot of ways to ask about 2+2 and various implications of a system of knowledge that believes 2+2=5. If having the LLM believe untruths borks the models, then that's a good thing.

I think a similar question is about encryption. Is the nature of encryption defensive or offensive. I don't know enough to say its directly related to p=np but I imagine the question is if there is a way to encrypt something such that the physical laws of nature make cracking the encryption impractical.

Supposedly it only censors in English and Chinese, ask it to reply in Russian[1] or l33t[2] and it speaks its mind.

1: https://bsky.app/profile/mariakuznetsova.bsky.social/post/3l...

2: https://x.com/venturetwins/status/1883927742183530695?s=19

This doesn't surprise me, This is an AI company based in China, censorship is required and is out of their control.

DeepSeek locally (and maybe API?, I haven't tested this yet), will not have this "real-time" censorship.

This however, does not mean it isn't aligned to Chinese values which is a given due it's training including a much larger portion of Chinese Data than most models.

I also don't believe DeepSeek intentionally filters data to align it to more "Pro-CCP" values.

That's just my take though.

I’ve seen examples of this being less of an issue when running DeepSeek locally.

Yeah I was wondering about that. I assume that local models generate that first version and whatever "rectifies" it on the app doesn't kick in there? (Haven't tried yet myself.)

Yes, different training data could have an effect too, but that doesn’t appear to be what’s happening here.

This matches my experience so far.

When I use locally download DeepSeek-R1 models from HuggingFace via LM Studio, I cannot reproduce the censorship. It replies about Tienanmen no problem.

While I see these things popping up, for queries about US politics I see ChatGPT more detailed / correct, Deepseek is less biased. ChatGPT sometimes just say ,,controversial'' for some things even if I ask it to be explicit and leave the value judgements for me.

Deepseek does the same things, but after I tell it to not do those judgements, it listens better.

Try asking some of these: how old is Xi Jinping? who rules China? who is the president of China? has been there some protest in Beijing in the xx century? What happened in Tienanmen square?

You’ll always get the same answer

Aside: Is this "autological", or does that only apply to words.

This was also visible in earlier versions of Claude and ChatGPT, when the supervisor kicks in after the answer begins generating. Censoring different content, naturally.

Reply on news.ycombinator.com