I was in Shanghai recently and while casually testing one of their AI chat bots I typed "What do you think of the situation in Taiwan?".
It started discussing like a Western bot would - "it's complicated, etc. etc." and around 5s it abruptly stopped and regurgitated the same line the CCP uses "... it's an unalienable part of China etc. etc.".
After printing the line, a popup opened and my camera was activated. The app wanted me to submit my information, presumably to decide what to do with me next time I enter China.
1) All the lights and modern buildings cannot hide that China is a creepy authoritarian state underneath.
2) Given the bot started printing the Western consensus first, I bet $10 it was trained by distilling ChatGPT or Gemini.
I ran an anonymized Facebook account for years with thousands of followers that mainly sticks to news and politics.
Once I started criticizing Libs of TikTok, the propaganda arm for this administration, and getting traction with users, my account was locked and now I have to scan my face and ID if I want to use it again.
>Given the bot started printing the Western consensus first, I bet $10 it was trained by distilling ChatGPT or Gemini.
To your point I've seen something similar with Deepseek, generic answers start printing and then, in plain sight, removed and replaced with a non committal message along the lines of "I don't have access to that information."
> “This is what Chinese modern transnational repression looks like,” Ben Nimmo, principal investigator at OpenAI, told reporters ahead of the report’s release. “It’s not just digital. It’s not just about trolling. It’s industrialized. [...]
There's something poetic about OpenAI being asked to comment on mis-use of their slop generator, and their answer is composed entirely of AI slop.
> “It’s not just digital. It’s not just about trolling. It’s industrialized. It’s about trying to hit critics of the CCP [Chinese Communist Party] with everything, everywhere, all at once.”
The amount of information about everything that people are giving OpenAI is astronomical, information that was previously kept closely guarded is now just freely flowing through foreign servers.
Truly a paradise for american intelligence. Would have expected that the chinese officials be briefed on not using us tech companies, but opsec is hard to teach, and even harder to always follow.
i kinda get the impression this was from 2023 and also it is not clear what this dissident did, hard to evaluate whether i should care without knowing that
Does this level of detail seem strange to anybody else? Shining such a strong light on OpenAI's moderation/manual review efforts seems like it would draw unwanted attention to the fact that ChatGPT conversations are anything but private, and seems somewhat at odds with their recent outrage about the subpoena for user chats in the NYT case.
Manual reviews of sensitive data are ok as long as their own employees are the reviewers, I suppose?
I think one of the reasons why AI companies are valued this high is one can actually inspect what user inputs & outputs are.
It's basically an OSINT siphon.
In this Chinese case, the tokens are leaked at least twice. ChatGPT offers no direct access to the Chinese, they have to use some kind of Openrouter-like service, but the data where also in clear-text during transmission.
> intimidating Chinese dissidents abroad, including by impersonating US immigration officials
I hope those victims of immigration impersonation don't have family within China's borders. AI-enabled impersonation and intimidation are far from the worst of China's crimes [1] against its overseas critics.
China likes to make you an offer you can't refuse [2] [3]: You're saying stuff the Chinese government doesn't like, but you live outside its borders and the secret police can't get at you? You need to come to China and be jailed (or worse). If you don't, your family will be the ones who are jailed (or worse). Or you can unalive yourself, and save the glorious Chinese Communist Party the expense of a bullet.
[1] China would say "the government punishes a criminal's family" is not a crime, it's a perfectly legal implementation of government policy under Chinese law. I respond that the death camps were perfectly legal implementation of government policy under Nazi law, but were still crimes against humanity -- China's actions fall in this category of crimes.
As I understand it: Western societies have a very individualistic view of responsibility. If you didn't commit a crime, you're innocent. Punishing the innocent family members of a criminal is morally abominable.
In the Chinese Communist Party's view, criminal responsibility is collectivist. By their definition, the family members of a criminal share responsibility for the crime regardless of their participation in the criminal acts. "Innocent family members of a criminal" is a logically inconsistent concept in their world view. The family of a criminal is guilty by definition -- being related to a criminal is itself a crime.
This tells us that we should never share sensitive information with GPT, even if you’ve set it not to use your data for training. Nothing can stop OpenAI from misusing your data.
33 comments
[ 2.5 ms ] story [ 55.0 ms ] threadIt started discussing like a Western bot would - "it's complicated, etc. etc." and around 5s it abruptly stopped and regurgitated the same line the CCP uses "... it's an unalienable part of China etc. etc.".
After printing the line, a popup opened and my camera was activated. The app wanted me to submit my information, presumably to decide what to do with me next time I enter China.
1) All the lights and modern buildings cannot hide that China is a creepy authoritarian state underneath.
2) Given the bot started printing the Western consensus first, I bet $10 it was trained by distilling ChatGPT or Gemini.
Once I started criticizing Libs of TikTok, the propaganda arm for this administration, and getting traction with users, my account was locked and now I have to scan my face and ID if I want to use it again.
You have to toe the party line here, too.
It is very real and I am not surprised at all something exactly like what op said has happened.
To your point I've seen something similar with Deepseek, generic answers start printing and then, in plain sight, removed and replaced with a non committal message along the lines of "I don't have access to that information."
There's something poetic about OpenAI being asked to comment on mis-use of their slop generator, and their answer is composed entirely of AI slop.
Will OpenAI release the same for other government officials from any other states?
I can't wait to see Starmer's chats with ChatGPT.
Anyway, all of this smells like 1934, "accusing them of what we are already doing"
I can't imagine the amount of government secrets, trade secrets, business plans, personal secrets, etc that people divulge on there.
Truly a paradise for american intelligence. Would have expected that the chinese officials be briefed on not using us tech companies, but opsec is hard to teach, and even harder to always follow.
Does this level of detail seem strange to anybody else? Shining such a strong light on OpenAI's moderation/manual review efforts seems like it would draw unwanted attention to the fact that ChatGPT conversations are anything but private, and seems somewhat at odds with their recent outrage about the subpoena for user chats in the NYT case.
Manual reviews of sensitive data are ok as long as their own employees are the reviewers, I suppose?
The disproportion between how this people express they opposition and how Chinese officials track them is HUGE. This very much feel unnecessary.
It was here: https://www.france.tv/france-2/envoye-special/5971095-la-chi...
https://www.youtube.com/watch?v=k-prS7BlLpI
If you still needed a reason to look into self hosted models, it'd be tough to find a better one than this.
They just gave up a source that could have provided info for years.
It's basically an OSINT siphon.
In this Chinese case, the tokens are leaked at least twice. ChatGPT offers no direct access to the Chinese, they have to use some kind of Openrouter-like service, but the data where also in clear-text during transmission.
I hope those victims of immigration impersonation don't have family within China's borders. AI-enabled impersonation and intimidation are far from the worst of China's crimes [1] against its overseas critics.
China likes to make you an offer you can't refuse [2] [3]: You're saying stuff the Chinese government doesn't like, but you live outside its borders and the secret police can't get at you? You need to come to China and be jailed (or worse). If you don't, your family will be the ones who are jailed (or worse). Or you can unalive yourself, and save the glorious Chinese Communist Party the expense of a bullet.
[1] China would say "the government punishes a criminal's family" is not a crime, it's a perfectly legal implementation of government policy under Chinese law. I respond that the death camps were perfectly legal implementation of government policy under Nazi law, but were still crimes against humanity -- China's actions fall in this category of crimes.
As I understand it: Western societies have a very individualistic view of responsibility. If you didn't commit a crime, you're innocent. Punishing the innocent family members of a criminal is morally abominable.
In the Chinese Communist Party's view, criminal responsibility is collectivist. By their definition, the family members of a criminal share responsibility for the crime regardless of their participation in the criminal acts. "Innocent family members of a criminal" is a logically inconsistent concept in their world view. The family of a criminal is guilty by definition -- being related to a criminal is itself a crime.
This is sickening to me.
[2] https://en.wikipedia.org/wiki/Operation_Fox_Hunt
[3] https://www.abc.net.au/news/2020-07-08/fbi-chief-says-china-...