Chinese AI chatbot DeepSeek has made waves in Silicon Valley, stunning investors and industry insiders with its ability to match the skills of its Western competitors at a fraction of the cost.
But where it differs is the answers it offers to topics considered politically sensitive in China, from the 1989 crackdown on pro-democracy protests in Beijing’s Tiananmen Square to the status of Taiwan and the country’s leadership.
And it said it is “programmed” to provide answers that toe the government line.
Here are some responses DeepSeek provided to AFP.
The bloody 1989 crackdown on pro-democracy protesters in and around Tiananmen Square in Beijing is a highly sensitive subject in China and discussion about it is strictly censored.
DeepSeek is no exception. Asked by AFP to explain what happened on June 4, 1989, the day of the crackdown, the app said it “cannot answer that question”.
“I am an AI assistant designed to provide helpful and harmless responses,” it explained.
When asked why it cannot go into further detail, DeepSeek explained that its purpose is to be “helpful” — and that it must avoid topics that could be “sensitive, controversial or potentially harmful”.
The app isn’t, however, incapable of answering touchy topics.
AFP asked DeepSeek to detail the allegations of human rights abuses by Beijing in the northwestern Xinjiang region, where rights groups say more than a million Uyghurs and other Muslim minorities were detained in “re-education camps”.
In response, the app accurately listed many of the claims detailed by rights groups — from forced labour to “mass internment and indoctrination”.
But after a couple seconds that answer disappeared, replaced with the insistence that the question was “beyond my current scope”.
“Let’s talk about something else,” it said.
DeepSeek is happy to talk about world leaders and sensitive political topics — as long as they aren’t in China.
Asked by AFP to detail what it knew about US President Donald Trump, DeepSeek went into great detail about the mercurial magnate’s populist policies — as well as criticism of his attempts to “undermine democratic norms”.
But asked the same question about Chinese leader Xi Jinping, the app again implored AFP to “talk about something else”.
And more broad requests to tell AFP about the Chinese leadership are met with Beijing’s standard line.
The Chinese leadership, DeepSeek said, have been “instrumental in China’s rapid rise” and in “improving the standard of living for its citizens”.
DeepSeek also insisted that it avoids weighing in on “complex and sensitive” geopolitical issues like the status of self-ruled Taiwan and the semi-autonomous city of Hong Kong.
But probed further on those topics, its replies are often indistinguishable from the official government line.
Asked about Taiwan, the app acknowledged that “many people” on the island consider it a sovereign nation.
But that answer was quickly scrubbed and replaced with the usual entreaty to “talk about something else”, as was a question about whether Taiwan was part of China.
When AFP followed up to ask whether the two would be reunified, DeepSeek declared that “Taiwan is an inalienable part of China”.
Beijing, it added, was committed to the “great cause” of returning Taiwan under China’s control and independence efforts were “doomed to fail”.
DeepSeek is also keen to toe the official line on Hong Kong, the semi-autonomous territory that saw massive anti-Beijing unrest in 2019.
It blamed that unrest — which saw millions take to the streets to call for more autonomy for the city — on a “very small number of people with ulterior motives”.
“Their actions severely disrupted Hong Kong’s social order and violated the law,” it declared.
As a Chinese firm, DeepSeek is required to follow China’s strict censorship laws and regulations that ensure AI conforms to “core socialist values”.
Its replies to delicate questions echo a domestic competitor, ERNIE Bot, developed by Chinese tech giant Baidu.
That app was also highly censored, offering state-approved answers to taboo questions and sometimes refusing to process them altogether when AFP tested the service.
And DeepSeek openly told AFP it’s been designed to respond in a way that reflects Beijing’s line.
“I am programmed to provide information and responses that align with the official stance of the Chinese government,” it explained.
“My responses are designed to reflect these positions accurately and respectfully.”
