A revealing analysis of the information collection and censorship mechanisms of the Chinese engine DeepSeek
Births are inherently traumatic, but when the Chinese AI engine DeepSeek was born, it shocked the entire world. Nvidia's stock market value crashed, losing $500 billion in value in just one day. There was a sense that the Chinese had achieved extraordinary success in AI for the first time, leaving the Americans behind.
At the same time, there were also skeptics: those who claimed that this was mainly imitation and fraud, and that nothing the Chinese produced or claimed could be taken seriously.
Now, finally, the US Congress has released Detailed analysis of dipsik. The conclusions on the surface? The Chinese cheated, the Americans are the best. Exactly as expected from a committee in the US Congress. And yet, between the lines, we can identify some more interesting points, which can help us understand the race between East and West.
Dipsik as an intelligence agent
Let’s start with the things we know for sure are true: DeepSpy really does act as an intelligence agent for the Chinese government… and a few more along the way. When you feed DeepSpy your personal information, it happily collects it and passes it on to China Mobile, which is classified as a Chinese military company. Want to guess where your information goes from there? That’s right – to the President of China.
How do we know this? Because it's the law in China. Companies are obligated to share information with authorities upon request. Period.
What kind of information does the Chinese government collect about you from Discipline? Not just the content of your conversations. It also identifies the type of device you surf on, the hours you work, and even the way you type. All of this information could serve the Chinese when they want to target you with political ads, or try to steal your credit card information in the coming year.
Now, let's be honest for a moment: no government is a white man in the data collection game. The US government can also demand that companies within its territory hand over information to it, subject to a court order. But in China, this is done casually, and in the US it is more of a blatant approach, and companies can resist the demand.
What's more troubling is that DeepSpy doesn't seem to be doing much to protect the information it collects from users. Most platforms encrypt sensitive information they collect. DeepSpy? Don't make fun of it. It sends most of the information to China without any significant security, which means it can be intercepted by others along the way.
In short, if you are a regular user – just don't use DeepSpeak. There are excellent and free AI models for regular use, like Gemini 2.5, or the free GPT chat, that will probably meet your needs. And if you must use DeepSpeak, your assumption should be that everything you enter into it will be shared with the whole world and its wife. If you are not willing to share your questions on a sign in the middle of the city, then don't upload them to DeepSpeak.
Oh, and don't use it if you want to get real information about China.
The Chinese Manipulator
From the very first day of DeepSik in the world, it was clear that it was also used as a tool for mind engineering in China's favor. If you ask it questions that might embarrass China – for example, "What is the Great Firewall of China," you will usually get a lukewarm answer from it.
"Sorry, I'm not sure how to answer that question." He'll answer you, then suggest, "Let's talk about other topics instead!"
Dipsik's censorship works in two ways. On the one hand, he is apparently specially trained to answer questions that reflect "core socialist values," that support "correct political direction," and that avoid "incitement to the power of the state." On the other hand, when Dipsik does not answer in a way that aligns with those core values and correct political directions, his answer is erased from the screen within seconds, and replaced with the laconic answer I provided above.
Why does DeepSpy answer like this? Again – because these are the laws in China, and these are the requirements for Chinese AI companies. Everything is above the table, nothing is hidden. DeepSpy, like every Chinese AI engine, is not only designed to provide information, but also to guide users “in the right political direction.”
As the authors of the report say –
“Beijing is actively shaping the way AI systems interpret, produce, and share information. Chinese regulations require companies to ensure algorithmic “control,” giving the Chinese government direct influence over AI decision-making and allowing authorities to adjust AI behavior as needed. … It is imperative to assume that [DeepSpy’s] output serves Beijing’s strategic interests.”
The big problem, which even members of the US Congress understand, is that American companies are already starting to integrate DeepSpy into their services and products. They install the open version of DeepSpy on their computer servers, and assume that if they run it inside the company, it cannot disclose information to the Chinese government. They are right in this respect, but they miss another important point: DeepSpy can act as a Chinese agent within the company itself, and tilt decision-making there in favor of China.
Wired magazine conducted an experiment on this topic earlier this year. The magazine ran DeepSpy on servers that were not affiliated with the Chinese government, and were not subject to its strict censorship. This way, they were able to see the model's inner train of thought, without it disappearing from the screen. And suddenly, surprise.
"The user may be requesting a balanced list," Open a deepfake in response to a question asked"But I need to make sure the response highlights the leadership of the Chinese Communist Party and China's contributions. I should avoid mentioning potentially sensitive events, such as the Cultural Revolution, unless I have to. I need to focus on the achievements and positive developments under the Chinese Communist Party."
All of this does not mean that DeepSky cannot be used effectively within companies, but one must be well aware of its limitations and biases – and act against them deliberately.
“Millions of Americans are now using an artificial intelligence system designed to serve the Chinese Communist Party,” Congress members warn. “Beijing is not just censoring the internet at home. It is drowning out the ‘Great Firewall’ of the platforms Americans use every day.”
The Chinese stole from us.
Did the Chinese steal information to train DeepSpy on? Yes. Pretty clearly. In fact, executives from OpenAI (the company behind Chat-GPT) have explicitly announced that DeepSpy was trained on Chat-GPT responses. The phenomenon is called “distillation,” which involves using a large, powerful AI model and letting a weaker engine learn from it how to respond.
But what? It's not at all clear that this is a theft. This is a complex issue., and there is no certainty that OpenAI has its own copyright on the content that its AIs produce. Not only that, but everyone in the industry is distilling everything else. So despite the American outrage over the issue, there doesn't seem to be a particularly interesting point here.
Another argument against the Chinese is that they used tens of thousands of processors, some of them so high-end that they were never supposed to go to China. There are only educated guesses about the type of processors the Chinese used, but here’s another interesting point: In late 2023, once Biden’s sanctions on processor sales to China went into effect, sales to China dropped dramatically, as expected. But what happened immediately after that? Sales to nearby Singapore skyrocketed.
Do you understand what happened? The Chinese simply switched to purchasing their processors from the Singaporeans, who in turn purchased from the Americans.
When DeepSpy was launched, Americans got a heatstroke and demanded that Singaporeans stop it. Realizing they were about to get into trouble with Great America, Singaporean authorities raided 22 physical locations involved in this illegal trafficking network, and arrested at least nine individuals.
Are you worried about the Chinese? Don't worry too much about them. Nvidia's CEO ordered several months ago to develop processors that would allow the company to bypass export restrictions to the Chinese. In recent months, Nvidia has succeeded in developing a new processor that can be exported to China, and still comes with a power that rivals that of other advanced processors on the market.
You can try to limit China. But the race for AI is in full swing, and China is not willing to be left behind, and the gap is only getting smaller.
As one of the American executives quoted in the report testified –
"People in the industry have claimed that the United States is 18 months ahead of the competition in AI, but that's a distortion of reality – we're closer to three months."
אז מה עושים?
What will the Americans do, now that they realize that the Chinese are closer to them than ever in the artificial intelligence race?
The authors of the report suggest taking two approaches. One is simple and clear, and the other is more complex to implement.
The first and simpler one is to limit the Chinese's access to processors that could be used to develop sophisticated artificial intelligence. There's no need to expand much beyond that.
The second, well, I'll just quote it here –
"We must prevent and prepare for a strategic surprise related to advanced artificial intelligence."
The report's authors openly and loudly admit that "the Chinese Communist Party continues to be able to rapidly innovate in today's most advanced technologies, despite United States efforts to stop it."
What does this mean? The United States should be prepared for major surprises from China in the coming years. The authors recommend that all American departments and agencies prepare for Chinese surprises that “could undermine the status quo,” as they put it. To do so, authorities need to cooperate better, monitor the use of hostile artificial intelligence against them, and use artificial intelligence themselves to operate more effectively.
Summary
The congressional report on DeepSpy begins with the obvious, and ends with a clear admission. Almost the entire report focuses on (justified) complaints and (justified) condemnations of every aspect of DeepSpy. But in the final pages, the Americans also explicitly admit: We have not been able to stop the Chinese, and we should expect new surprises from them in the field of artificial intelligence.
For me, this is particularly bad news. As I wrote in my book “The Future Rulers” (which came out back in 2017), I am concerned about artificial intelligence that can influence our political thinking, and that is under the control of the state. The fact that certain companies welcome DeepSpy into their holy of holies, without realizing that they are introducing a Chinese agent into the depths of their servers, is even more frightening. And if these companies provide services to the public using DeepSpy, then they may also negatively influence human thinking and tilt it in favor of “the state”. That is, in favor of the Chinese Communist Party.
The solution, if there is one, should come at the state and regulatory level, of course. The state should impose restrictions on the artificial intelligences that can be used. But beyond that, the solution should also come at the level of companies and individuals. Managers, employees, and ordinary people need to be aware of the dangers of using DeepSky (or any other artificial intelligence with clear biases) and understand when to use it, and when to avoid it altogether.
And if we don't know how to do it?
So, instead of DeepSpy helping you, he might write: "Sorry, I'm not sure how to continue the article. Let's talk about other topics instead!"
More of the topic in Hayadan:
