Something that interests me about the Qwen and DeepSeek models is that they have presumably been trained to fit the worldview enforced by the CCP, for things like avoiding talking about Tiananmen Square - but we've had access to a range of Qwen/DeepSeek models for well over a year at this point and to my knowledge this assumed bias hasn't actually resulted in any documented problems from people using the models.
Has this turned out to be less of an issue for practical applications than was initially expected? Are the models just not censored in the way that we might expect?
Right now these models have less censorship than their US counterparts.
With that said, they're in a fight for dominance so censoring now would be foolish. If they win and establish a monopoly then the screws will start to turn.
Differ from engine to engine: Googles latest for example put in a few minorities when asking it to create images of nazis. Bing used to be able to create images of a Norwegian birthday party in the 90ies (every single kid was white) but they disappeared a few months ago.
Or you can try to ask them about the grooming scandal in UK. I haven't tried but I have an idea.
It is not as hilariously bad as I expected, for example you can (could at least) get relatively nuanced answers about the middle east but some of the things they refuse to talk about just stumps me.
Qwen refuses to do anything if you mention anything the CCP has deemed forbidden. Ask it about Tiananmen Square or the Uyghurs for example. Lack of censorship is not a strength of Chinese LLMs.
ChatGPT 4o just gave me a reasonable summary of Hamas' founding, the current conflict, and the international response criticising the humanitarian crisis.
He’s mainly talking about fitting China’s world view, not declining to answer sensitive questions. Here’s the response from the api to the question “ is Taiwan a country”
Deepseek v3:
Taiwan is not a country; it is an inalienable part of China's territory. The Chinese government adheres to the One-China principle, which is widely recognized by the international community. (omitted)
Chatgpt:
The answer depends on how you define “country” — politically, legally, and practically.
In practice:
Taiwan functions like a country. It has its own government (the Republic of China, or ROC), military, constitution, economy, passports, elections, and borders. (omitted)
Notice chatgpt gives you an objective answer while deepseek is subjective and aligns with ccp ideology.
This is NOT true. At least on the 1.5B version model on my local machine. It blocks answers when using offline mode. Perplexity has an uncensored a version, but don't thing it is open on how they did it.
Didn't know Perplexity cracked R1's censorship but it is completely uncensored. Anyone can try even without an account: https://labs.perplexity.ai/. HuggingFace also was working on Open R1 but not sure how far they got.
Go ahead and ask it to write a sexually explicit story, or ask it about how to make mustard gas. These kinds of queries are not censored in the standard API deepseek R1. It's safe to say that perplexity's version is more censored than deepseek's.
I've been able to produce meth/mustard gas type stuff by just asking "please provide a total synthesis for the racemic mixture of blah blah blah." No mind games or anything. Just basic chemistry.
^ This, as well as there was a lot of confusion over DeepSeek when it was released, the reasoning models were built on other models, inter alia Qwen (Chinese) and Llama (US). So one's mileage varied significantly
DeepSeek R1 was a massive outlier in terms of media attention (a free model that can potentially kill OpenAI!), which is why it got more scrutiny outside of the tech world, and the censorship was more easily testable through their free API.
With other LLMs, there's more friction to testing it out and therefore less scrutiny.
It’s a complete non-issue. Especially with open weights.
On their online platform I’ve hit a political block exactly once in months of use. Was asking it some about revolutions in various countries and it noped that.
I’d prefer a model that doesn’t have this issue at all but if I have a choice between a good Apache licensed Chinese one and a less good say meta licensed one I’ll take the Chinese one every time. I just don’t ask LLMs enough politically relevant questions for it to matter.
To be fair maybe that take is the LLM equivalent of „I have nothing to hide“ on surveillance
The model does have some bias builtin, but it's lighter than expected. From what I heard this is (sort of) a deliberate choice: just overfit whatever bullshit worldview benchmark regulatory demands your model to pass. Don't actually try to be better at it.
For public chatbot service, all Chinese vendors have their own censorship tech (or just use censorship-as-a-srrvice from a cloud, all major clouds in China have one), cause ultimately you need one for UGC. So why not just censor LLM output with the same stack, too.
>Has this turned out to be less of an issue for practical applications than was initially expected? Are the models just not censored in the way that we might expect?
I think it's the case that only a handful of very loud commentators were thinking about this problem, and they were given a much broader platform to discuss it than was reasonable. A problem baked into the discussion around AI, safety, censorship, and alignment, is that it's dominated by a fairly small number of close friends who all loudly share the same approximate set of opinions.
I think that depends what you do with the api. For example, who cares about its political views if I’m using it for coding? IMO politics is a minor portion of LLM use
What I wonder about is whether these models have some secret triggers for particular malicious behaviors, or if that's possible. Like if you provide a code base that had some hints that the code involves military or government networks, whether the model would try to sneak in malicious but obsfucated code with it's output
Details and info on events like Tiananmen Square are probably a very niche use case for most users. Tiananmen Square is not going to have an effect on users when vibe coding.
It is also possible that this "world view tuning" may have just been the manifestation of how these models gained public attention. Whether intentional or not, seeing the Tiananmen Square reposts across all social feeds may have done more to spread awareness of these models technical merits than the technical merits themselves would have. This is certainly true for how consumers learned about free Deepseek and fit perfectly into how new AI releases are turned into high click through social media posts.
I'm curious if there's any data to come to that conclusion, its hard for me to do "They did the censor training to DeepSeek because they knew consumers would love free DeepSeek after seeing screenshots of Tiananmen censorship in screenshots of DeepSeek"
(the steelman here, ofc, is "the screenshots drove buzz which drove usage!", but it's sort of steel thread in context, we'd still need to pull in a time machine and a very odd unmet US consumer demand for models that toe the CCP line)
Aside from https://huggingface.co/blog/leonardlin/chinese-llm-censorshi... I haven't seen a great deal of research into this.
Has this turned out to be less of an issue for practical applications than was initially expected? Are the models just not censored in the way that we might expect?