Chinese AI applications have over-refusal issues similar to Western AI, where answers on sensitive topics are partially generated and then retracted at the application deployment layer rather than by the underlying model.
Nathan describes asking Chinese AI about the Manchu ethnic group and watching the answer partially appear before being retracted, which he says confirms that censorship happens at the app deployment level, not within the models themselves. ✦ AI generated
Nathan · The Cognitive Revolution · 2026-07-27 · original ↗
starts at this moment · 63:25
one time I asked what's up with the man group today? Are they still around? Are they still part of China? And there you would get the kind of classic old Sydney Bing experience where the model itself would answer the question and you'd see kind of the chain of thought. You might even see the beginning of an answer and then sometimes like with this Manchu question because again you're touching on something fairly sensitive in the Chinese political context of ethnic minorities. the answer is in the process of unrolling in front of you and then you would get this pullback. You'd see the answer disappear and then you'd see, I'm sorry, I can't help you with that. Ask a different question or whatever. And again, this wasn't something that I think they really even necessarily meant to censor. And not all of the apps censored it the same way. I was able to get some answers from some apps that I wasn't able to get from other apps. it goes to show that in many ways the the Chinese ecosystem has a lot of the same things going on that we do where people are trying to make sure that they're on the right side of regulators on the right side of history acting responsibly as they understand it and at times that can go a little too far and it can interfere with the user experience but this is also consistent with the reporting that I've seen in many places that the Chinese models themselves are not very censored it's the actual deployed application experiences that usually contain the censorship
verbatim transcript · starts at 63:25
63:25time I asked what's up with the man group today? Are they still around? Are they still part of China? And there you would get the kind of classic old Sydney Bing experience where the model itself would answer the question and you'd see kind of the chain of thought. You might even see the beginning of an answer and then sometimes like with this Manchu question because again you're touching
63:47on something fairly sensitive in the Chinese political context of ethnic minorities. the answer is in the process of unrolling in front of you and then you would get this pullback. You'd see the answer disappear and then you'd see, I'm sorry, I can't help you with that. Ask a different question or whatever. And again, this wasn't something that I think they really even necessarily meant to censor. And not all of the apps
64:11censored it the same way. I was able to get some answers from some apps that I wasn't able to get from other apps. It's it goes to show that in many ways the the Chinese ecosystem has a lot of the same things going on that we do where people are trying to make sure that they're on the right side of regulators on the right side of history acting
64:32responsibly as they understand it and at times that can go a little too far and it can interfere with the user experience but this is also consistent with the reporting that I've seen in many places that the Chinese models themselves are are not very censored it's the actual deployed application experiences that usually contain the censorship and you could you could see that as the the answer was like partially there and then
64:58it would be retracted sometimes and you'd be hit with this no I'm sorry I can't help you. So it was overall I'd say a very good experience with the Chinese AIS and of course I'm doing this all in English. They do have some of them have weird quirks where I think it's in deepseek advanced mode with the best DeepSseek model. Yeah, if you go into expert mode on DeSseek, then you
65:21can't attach any files and there's also no search enabled. That's kind of weird. Uh I'm not sure why that's the case, but that is the case. Either use depending on what I was trying to do, I might just use the instant mode on DeepSeek. On Kimmy, that issue is not there. you had the ability to have the model do searches even with the best model that was available at the time. And Miniax, I