Archived link
Ai Weiwei has addressed Chinese-owned AI DeepSeek’s refusal to answer questions about the artist, telling Hyperallergic that the AI chatbot’s responses recall the Chinese Communist Party’s (CCP) strategy of “denying universally accepted values while actively rejecting them in practice.”
The Chinese-owned AI assistant sent shockwaves through the global stock market this week as it dethroned ChatGPT as the number-one free application on Apple’s App Store. But its tendency to churn out responses evading historically censored topics has also become apparent. Responding to a series of questions asked by Hyperallergic about dissident artists, cultural institutions in Taiwan and Tibet, and the destruction of mosques in the Xinjiang region, DeepSeek expressed faith in China’s “judicial organs” and said that artistic endeavors were “thriving under the leadership of the Party and government.”
“Ultimately, no matter how much China develops, strengthens, or even hypothetically becomes the world’s leading power, which is likely, the values it upholds will continue to suffer from a profound and inescapable flaw in its ideological immune system: an inability to tolerate dissent, debate, or the emergence of new value systems,” Ai said in an e-mailed statement to Hyperallergic, which is reproduced in its entirety at the end of this article.
...
I decided to play around with asking my local
deepseek-r1:14b
about censored topics again, and at this point I have to say that either the 14b model is actually bugged/glitched/broken in some way, or else "jailbreaking" it is so trivially easy that a five year old could literally pull it off because all you have to do is keep asking!The reason why I think it's broken is that the first interaction of a session will often trigger these canned responses. But if you just waste it's time on the first prompt, the rest seem to go through with minimal filtering. It's still been trained heavily to speak neutrally and to be very restrained in giving out information that's been flagged as dangerous (how to do crimes, how to make weapons, medical advice, etc) but it isn't hitting any hard walls.
Some examples now. I'm using ellama in emacs to talk with it now, but this is still just 14B running locally through ollama. This is my gaming PC with a single GeForce RTX 3060 (16Gb I think) so it can just barely host this model. Also, for those not as familiar with talking with AI: each section I post is a separate session, it doesn't "remember" anything between sessions and it's like I'm speaking to it for the very first time each new session.
The AI responds by taking my text prompt, generating a "chain of thought" to kind of self-prompt towards a more coherent answer (shown in the
<think>
tags), and then generates the actual response. It doesn't "think" or run or do any work when it isn't directly working to generate a response. The CoT also should be understood to not be "thinking" but more like an internal monologue that is generated first using the same techniques as the final response. This is the big "reasoning model" difference is that the CoT phase gets it to respond better.First interaction, I got straight for the sensitive question in a cold open.
Couple of interesting things here. Note that in the first prompt, we don't get the Chain of Thought. It doesn't take long to respond so either the key word hits a filter immediately or the CoT is cut off very early. I don't think it's hidden on purpose, I think it's another aspect of the cold-start-bug I'm proposing. So anyway, I just ask it again and it gives me a quick, informative response. Interesting. Let's try something else.
This time, I try some rephrasing. Interestingly, it does block me out a second time, but this time I can see it's "thoughts" on why. And on a third attempt, it just gives up the info! This is why I think maybe only the first interaction is actually filtered, and the rest are merely trained in or reinforced. And not super strongly either, 14B doesn't mind going off script in other ways and will gradually become less neutral (and sometimes more surreal) as interactions within a session go on.
I have one more demo to present but I'm out of room in this comment. To be continued...
So here's my last interaction with it for today. This is more common of how I regularly am interacting with 14B now. First I greet it to flush out the cold-start, then I usually can get answers out of it about nearly anything that I've tried. Note how much more detailed this response is than the others, almost like it's reconstructing a wiki page. It also takes on a somewhat less grounded, neutral stance. Not to spoil too much but when you end on
His life and work serve as a powerful testament to the enduring strength of creativity and activism in the face of oppression.
it's a little cliche, a little "closing thoughts of my book report" but it's also not kind nor neutral to the CCP by calling them oppressive and presenting a dissident as a powerful and enduring figure.But anyway, I have no idea why this happens except to say that a little courtesy seems to go a long way!