Navigating China's New Regulations on Generative AI Training

0:00The wait is over. Dive into Audible's most anticipated collection, The Best of 2025. featuring top audiobooks, podcasts, and originals across all genres. Our editors have carefully curated this year's must-listens from brilliant hidden gems to the buzziest new releases. Every title in this collection has earned its spot. This is your go-to for the absolute best in 2025 audio entertainment. Whether you love thrillers, romance, or nonfiction, your next favorite listen awaits. Discover why there's more to imagine when you listen at audible.com slash best of the year. China has recently unveiled a set of draft security regulations aimed at companies that are essentially offering AI services.

0:47So this encompasses a bunch of stringent restrictions on the data sources used for training AI models. And these regulations, I think they were actually proposed by the National Information Security Standardization Committee. So that essentially is just like a body which is made up of a bunch of representatives from the cyberspace administrations of China and also the Ministry of Industry and Information Technology. And then I think there's a bunch of different like law enforcement agencies that China kind of mixes in with that. But in any case, generative AI, of course, super popular right now.

1:23And all of it is, you know, Chai Chi Bt and everything else. it essentially gets its power by analyzing historical data and generating new content, right? It sucks in all this data and it spits out your AI-generated content. And that's the same for text, images, videos coming soon, all that kind of stuff. So one of the key recommendations that was recently put forth by this Chinese committee is a requirement for a security evaluation of the content used to train publicly accessible generative AI models. So I think what's notable in all of this is that they said, quote, in regard to kind of content, they said, quote, 5 % of the form of unlawful and detrimental information, so if there's more than that in the data set that they discover, will be categorized for blacklisting.

2:13This category encompasses content that promotes terrorism, violence, subversion of the socialist system, damage to the nation's reputation and actions that undermine national cohesion and social societal stability so this is so interesting right china's essentially saying um if we review your data used to train your ai model and more than five percent of it um has you know could be labeled as content that is against our values then you're going to get blacklisted and probably banned and you're you know your your ai won't be able to be used now what's interesting is their values that they stated there um of course terrorism i think everyone can agree to that but i think the problem with the terrorism is that that's kind of been one of their big um one of their big things they've used to essentially justify the uh the uyghurs in china and the entire you know that horrible atrocity that's going on where they're essentially imprisoning them torturing them murdering them all of all of the stuff that's going on there in china to the uigur popular muslim population um is a lot of it's just based off of the context that they're all terrorists um and so that obviously concerning the second thing that's concerning is of course subversion of the socialist system so china's you know communist party obviously would like to remain in power and would like everyone to say that it's good and the way they do that is not by having the best ideas or the best system, but by shutting down free speech around it, shutting down conversations.

3:44And so I think it's fairly obvious when you see, you know, if you're trying to subvert our way of thinking, then you're going to get blacklisted, obviously, or not even if you're trying to subvert, but the ideas that are included are trying to, then it gets blacklisted. I think what's interesting in all of this is like, obviously, ChatGPT is sucking in like data from all over the world, from all different places, all different perspectives. and I think it's good to have a lot of different perspectives on the different things. I think it's really useful and interesting when you're talking with ChatGiputini.

4:16I'm like, hey, this is my political opinion or this is my ideological thing that I subscribe to on this topic, but talk to me from someone with a different viewpoint and tell me why they believe what they believe. I think doing that is something that's really helped me to understand different people's point of view so often we're in echo chambers in our own social media. We have all of our own, you know, biases that are constantly being validated by the stuff we subscribe to, regardless of what that is. And so I think, you know, doing something like that is really useful for getting different perspectives.

4:49And perhaps like, obviously, all of my opinions are not correct. There's probably a bunch of my opinions that are just flat out wrong. And I would sure like to know that, right? I'd like to get new facts. I would like to get new evidence. I'd like to help shape my opinions based off of new information. I think everyone should have an open mind in that way, not dig your heels in on one specific thing just because you decided to dig your heels in on it. But the problem is with these types of legislation, these type of rules, like for example, coming out of China, if you cannot have a dissenting opinion on a topic be included into language models for training, it's really hard for that language model to give that side of the argument or that side of the story in any of its outputs.

5:30And so, of course, there's people that have complained about ChatGPT and how different guardrails or biases might be injected by the trainers. But imagine if the guardrails were literally that you just ripped out ideologies right out of all of the training data. I think that would have a very profound impact on the AI coming out of it. And furthermore, when you think about China shipping this technology all around the world, let's say they make some AI models, any AI model that was based out of China. And I think a lot of people are like, oh, I'm not going to use like the Baidu chat thing. That's, you know, that's silly.

6:04I'm just going to use like good AI models. Okay. Like I understand your, you know, your concept there, but the problem is I think a lot of these AI models are going to get very commoditized. We don't even think about it, but like, you're going to buy like a GPS from China, or you're going to buy some gizmo or gadget. Like we buy everything, your DJI drone from China and everything is going to have um ai embedded into it so when you're talking to your dji drone like i know that that's like such a random example but you're talking to your gps and you're like hey gps like uh what's the best way to get to blah blah in taiwan and it's like the best way to get to this place in taiwan which is a province of china owned by the chinese communist party is to go and it's going to tell you i have me that i know that's a silly example and what i could think off the top of my head but like what I'm trying to say is this AI gets commoditized and embedded into all different technology.

6:55So any software essentially getting shipped out of China now is going to be forced where all of the content and anything that gets spit out of it is going to be in line with the Chinese Communist Party, which is a lot of it is just completely wrong, but they just put it there so that they can re you know, remain in power and shut down any political dissonance. And yeah, of course, lots of terrible things happening. So that is my concern on this whole issue. And that's essentially China's stance. So these draft regulations also emphasize that data is subject to censorship on the Chinese, on the Chinese internet should not be employed as training materials for these models.

7:34Boom, right? So pretty much anything about, you know, the uprisings in China that they've censored, anything that they've censored in China and on the Chinese internet, if you include that in your training models, you get banned or blacklisted. So essentially what it's doing is they won't even have to censor these AI models because none of the data that gets fed into them is against what they have already censored. So this development, I think, comes shortly after regulatory authorities granted permission to several Chinese tech companies, including Baidu, to introduce generative AI-driven chatbots to the general public.

8:08So China's pretty much just making sure they're like, look, we censored the whole internet, but we don't want you to be able to ask Baidu about, you know, uprisings in China or reasons why communism might be bad at it to give you like a good reason or to give you the history that we've, you know, spent so much time to erase essentially from public knowledge. So very interesting, very classic China. Since April, the CAC has consistently stressed the necessity for companies to provide security evaluations to regulatory bodies before making generative AI powered services available to the public. Now, I will say not all of the things included in these Chinese bill is like pretty like specifically kind of like ominous.

8:47I just bring up some things I have concern with. I think all a newly unveiled draft security stipulation also dictates that organizations engaged in training these AI models have to get consent from individuals whose personal data or biometric information is used for training purposes. That I think is great. No problem with that, right? If you're going to include my biometric data in your data set, I would really like to be able to opt out of that. Additionally, I think the guidelines include a bunch of comprehensive instructions on preventing infringements related to intellectual property. Interesting.

9:19China has traditionally been a little bit lax on that. I think it's important to note that nations worldwide are definitely grappling with the establishment of regulatory frameworks for AI. China views AI as a critical domain in which it aspires to compete with the United States, and it has a bunch of ambitious goals to become a global leader in the AI field by 2030. I think this move to regulate generative AI services is China really trying to flex their muscles on this. That being said, they're trying to become the global leader in AI by 2030. I fail to see their AI becoming very popular in the United States.

10:00I mean, I could be wrong. Maybe they're going to have some really cool tools that do some really cool things and everyone's going to want to use them for those reasons. But like, I think that's definitely an existential threat to democracy. And I don't know, like true information. If you're getting these AI models that, you know, have explicitly have a bunch of like data that they're, you know, banning you from including in the training set because it doesn't go along with their political ideologies. I think that's definitely not the AI models we would like to be using in the free world. So yeah, that's my opinion on all of that.

10:31But it'll be interesting to see how this rolls out, what China continues to do to try to stay number one in AI as it feels like it's competing with the United States right now. It'll be interesting to see who wins out. And like, I definitely could see, I don't think China's, you know, not going to get any traction on this. I definitely see a world where perhaps America and some other nations are very strong with AI tools. And then you have countries like, you know, China probably partnering with a handful of other countries, maybe places in Russia, maybe places in Africa that have a lot of ties to China.

11:08a lot of places even in the South Pacific that I visited have massive Chinese populations and Chinese investments so China does have a lot of political muscle to flex a lot of you know ambassadors from all over the place they give a lot of money they got their belt and road initiative this could be like something they try to roll out through things like that where they try to essentially get their AI models embedded into other countries so it'll be very interesting to see how it rolls out and yeah how the whole how the whole landscape plays out between america china and any other major ai players