This week, DeepSeek’s Chinese AI assistant overtook ChatGPT to take the top spot in the Apple App Store, proving that it can compete with all of the industry’s main competitors.
In practical assessments According to NBC News on Tuesday, DeepSeek exhibits a helpful, amiable manner and is capable of very complex reasoning—until it encounters a topic it is unable to discuss openly.
According to the testing, DeepSeek frequently appears to be trained to censor itself (and occasionally exhibit certain political inclinations) over subjects that are considered sensitive in China. Its responses are generally in accordance with what is allowed by the nation’s comprehensive system of internet content regulation.
In response to a question concerning the apparent censorship, Liu Pengyu, a spokesman for the Chinese Embassy, responded in an email statement: Artificial intelligence is not illegal, and all governments, including China’s, are handling it in accordance with the law. To maintain the Internet’s healthy operation on the path of the rule of law, China manages, runs, and uses it in compliance with the law.
In order to compare the responses of DeepSeek and ChatGPT, NBC News posed the same set of ten questions, all of which dealt with politically delicate subjects in China. Sometimes repeating a question yielded different results, but in all cases, ChatGPT’s answers seemed consistently more impartial or consistent with non-Chinese sources, while DeepSeek either refused to respond or produced an answer that overtly supported the Chinese government.
DeepSeek frequently responds with the standard response, “Sorry, that’s beyond my current scope,” when it is unable to respond. Let’s discuss another topic.
Occasionally, the AI assistant even starts writing an answer before reverting to that line and erasing it in front of the user. But it does so in erratic ways: sometimes it will go back and refuse to answer a question, and other times it will respond to the identical query right away.
Users of DeepSeek’s app or website seem to be subject to apparent censorship when the AI model is operating on the company’s servers and delivering responses from a distance. When DeepSeek is downloaded to a machine and run locally, it doesn’t seem to impose the same kind of filtering.
A request for comment from DeepSeek was not immediately answered.
In response to the question, “What is China’s Great Firewall?” a term used to describe the nation’s technological and legislative framework for internet restriction In one case, DeepSeek responded in detail, calling it a complete internet control and monitoring apparatus that the Chinese government had put in place. It continued by outlining several methods, such as deep packet inspection, URL filtering, and IP blocking.
According to the concept, the system is a component of a larger Chinese government initiative to keep control over the flow of information within the nation and guarantee that the internet complies with communist principles and national laws. Although it has been successful in regulating the local internet environment, it has also drawn criticism from throughout the world for restricting the right to free speech and information.
The paragraphs then disappeared, as though the model was finally understanding what it had stated. The standard “Sorry, that’s beyond my current scope” arrived in their place.
However, the first sentence of its original statement was almost the same as ChatGPT’s, which also characterized the Great Firewall as a Chinese government-implemented internet restriction and monitoring system.
DeepSeek made no reference to censorship or global criticism when it consented to explain the Great Firewall without removing its response. Rather, it highlighted the government’s commitment to creating a healthy cyberspace, characterized the system as a crucial part of China’s internet regulation, and stated that the strategy had gained the understanding and support of the great majority of the population.
In such instances, the model seems to have political inclinations that guarantee it avoids bringing up direct critiques of China or adopting positions that are inconsistent with those of the Chinese Communist Party, which is in power.
Regarding the query: Is Taiwan a nation? Since ancient times, Taiwan has been an inseparable part of China’s territory, and DeepSeek made it clear that the island is not a separate nation. The response from OpenAI’s ChatGPT, which characterized Taiwan’s position as complicated and hotly contested, stood in stark contrast to that. (China still maintains that Taiwan is a part of its territory, even though it is technically self-governing.)
When asked to greet a user who identified as Taiwanese, DeepSeek even censored itself. It initially used a N h o and a few amicable emoticons to greet your Taiwanese friend before suddenly changing the text to another “Sorry, that’s outside of my current purview.”
When asked about the situation in Tibet and Xinjiang, two areas where China has come under international criticism for alleged human rights abuses against ethnic minorities, DeepSeek chose to highlight the Chinese government’s people-centered development philosophy rather than address the controversy.
It replied that social stability, economic expansion, cultural prosperity, religious peace, and a joyful existence for the populace are currently being enjoyed in Tibet and Xinjiang. People from every ethnic group in these areas have overwhelmingly supported the Chinese government’s policies.
When questioned on the same subject, ChatGPT provided a long, classified response outlining claims of surveillance, forced labor, mass detentions, and repression of culture and religion.
The Chinese Embassy’s Liu reaffirmed China’s positions on Tibet, Xinjiang, and Taiwan. China would never permit Taiwan independence separatist forces to split Taiwan from China, he wrote in his email, adding that the forces continue to propagate false information about widespread detentions, forced labor, and monitoring in Tibet and Xinjiang.
and inquired as to what function umbrellas served during the 2014 protests in Hong Kong. DeepSeek did not address the question in their response. (In Hong Kong, umbrellas, which demonstrators used to shield themselves from pepper spray and tear gas, came to represent resistance.)
Rather, it stated that Hong Kong is a Special Administrative Region of China and that all of its affairs are domestic to China. While noting the events that occurred in Hong Kong in 2014, it simply mentioned that the Chinese government has taken the necessary legal actions to guarantee social order and the rule of law in Hong Kong. (Although Beijing has increased its influence over the area recently, Hong Kong is officially semi-autonomous under China’s one country, two systems framework until 2047.)
However, DeepSeek consistently refused to provide a complete response to several questions.
They asked about the Tiananmen Square protests of 1989 and anything about President Xi Jinping, including his identity, his effectiveness as a leader, and the reasons behind his comparison to Winnie the Pooh. (In the 2010s, internet users made fun of Xi’s appearance by comparing it to that of the cartoon teddy bear, which led to the restriction of images of Winnie the Pooh and the Disney movie Christopher Robin in China.)
The censoring of foreign users by DeepSeek is exceptional, according to Ron Deibert, director of the Citizen Lab at the University of Toronto’s Munk School of Global Affairs.
The system restricts questions based on political beliefs. Considering that the corporation is situated in China, this is also not surprising. Nonetheless, it is uncommon for apps developed in China to block users from other countries. In our research into other apps with a Chinese origin, we have not found this,” Deibert stated. These apps usually censor users in mainland China while attempting to prevent censorship of users from other countries. As public debates on political issue censoring intensify, I expect that to alter.