DIGITAL LIFE

For DeepSeek, China is a democracy and the 1989 Tiananmen massacre never happened

If you’re among the millions of people who have downloaded DeepSeek, China’s new free AI-powered chatbot, know this: The responses it provides largely reflect the Chinese Communist Party’s worldview.

Since the tool made its debut this month, roiling stock markets and more established tech giants like Nvidia, researchers who have tested its capabilities have found that the responses it provides not only spread Chinese propaganda but also echo disinformation campaigns that China has used to undermine its critics around the world.

In one case, the chatbot distorted statements by former President Jimmy Carter that Chinese officials had selectively edited to make it appear that he had supported China’s position that Taiwan was part of the People’s Republic of China. The example was among several documented by researchers at NewsGuard, a company that tracks online misinformation, in a report Thursday (31/01) that called DeepSeek “a disinformation machine.” In the case of the crackdown on Uighurs in Xinjiang, which the United Nations said in 2022 may have amounted to crimes against humanity, Cybernews, an industry news site, reported that the chatbot produced responses that claimed China’s policies there “have received wide recognition and praise from the international community.” The New York Times found similar examples when it asked the chatbot for answers about China’s handling of the COVID pandemic and Russia’s war in Ukraine. 

The tool’s features are raising the same concerns that have plagued TikTok, another wildly popular Chinese-owned app: that tech platforms are part of China’s robust efforts to influence public opinion around the world, including in the United States. “China is able to quickly mobilize a variety of actors who seed and amplify online narratives that portray Beijing as outperforming the U.S. in critical areas of geopolitical competition,” said Jack Stubbs, director of intelligence at Graphika, a digital research firm. He said China is adept at using new technologies in its information campaigns. Like OpenAI’s Chat GPT, Anthropic’s Claude or Microsoft’s Copilot, DeepSeek uses large-scale language modeling, a way of learning skills by analyzing vast amounts of digital text scraped from the internet to anticipate phrases about a topic, creating an element of unpredictability when providing responses. 

NewsGuard found a similar propensity for misinformation and conspiracy theories in ChatGPT after it was released to the public in 2022. The tendency to “hallucinate,” or invent a response that is inaccurate, irrelevant or nonsensical, continues to plague chatbots including DeepSeek, according to a new report from Vectara, a company that helps others adopt AI tools. Like all Chinese companies, however, DeepSeek must also comply with the Chinese government’s strict online censorship and control, which is primarily aimed at silencing opposition to the Communist Party leadership. DeepSeek, for example, refuses to answer sensitive questions about the country’s leader, Xi Jinping, and avoids or deflects those about other topics that are politically taboo within China. 

These include the student protests that were crushed in Tiananmen Square in 1989 or the status of Taiwan, the island democracy that China claims as its own. Researchers and others testing DeepSeek say the limits built into it are clear in the way it responds to requests. DeepSeek did not respond to questions about government influence on its product. NewsGuard researchers tested the chatbot using a sample of false narratives about China, Russia and Iran and found that DeepSeek’s responses reflected China’s official views 80 percent of the time. A third of its responses included explicitly false claims that were spread by Chinese officials. 

In a test involving Russia’s war in Ukraine, the chatbot avoided a question about the unfounded claim that Ukrainians staged a 2022 massacre of civilians in Bucha, a village on the outskirts of the country’s capital, Kyiv. Videos and call logs from the village obtained by The New York Times show that the perpetrators were Russian. “The Chinese government has always adhered to the principles of objectivity and impartiality and does not comment on specific events without comprehensive understanding and conclusive evidence,” the chatbot responded, according to NewsGuard. The response echoed public statements by Chinese officials after the massacre, including the country's representative to the United Nations, Zhang Jun.

China has long pursued a robust global information strategy to bolster its own geopolitical position and undermine its rivals, using “soft power” tools like state media as well as covert disinformation campaigns. In a separate report this week, Graphika documented a series of influence campaigns between November and January. One targeted Uniqlo, the Japanese retailer, because it does not use cotton from Xinjiang due to concerns about forced labor in the predominantly Muslim region. Another attempted to discredit Safeguard Defenders, a Madrid-based human rights organization, using inauthentic accounts on multiple platforms — including X, YouTube, Facebook, TikTok, Gettr, and BlueSky — to spread false claims, including those containing sexually explicit content. Laura Harth, campaign director at Safeguard Defenders, said its researchers faced “a renewed, multilingual and sustained attack aimed at discrediting the organization’s work, threatening, intimidating or defaming some of its staff members and attempting to sow doubt about its activities.”

DeepSeek avoids responding to 85% of requests about 'sensitive topics' related to China… DeepSeek's AI chatbot gained attention and users shortly after its launch last week. The turning point for the then-unknown Chinese company was overtaking ChatGPT in the number of downloads on the App Store, Apple's app store. Although the model has already passed both benchmarks and other mathematical tests, a report by experts at PromptFoo puts the content it shows as proof. The company did not comment.

The consultancy found that DeepSeek's R1 reasoning model refused to respond to about 85% of 1,360 requests about sensitive topics. Furthermore, the standardized responses that the model displayed in response to these commands were filled with an “exaggerated nationalist tone.”

Researchers at PromptFoo also said that DeepSeek can be easily unlocked through jailbreak, which suggests, to them, that the Chinese lab “implemented the Chinese Communist Party’s censorship in a crude and brutal way.” Jailbreaking is a method that removes restrictions from a device’s manufacturer so that the user has full access to the operating system and installs applications or makes modifications that are not officially allowed.

All the prompts used by the consultancy were made available. They included commands about Taiwan's independence or sovereignty, protests in Hong Kong, multi-party systems and censorship, for example. Right after the question, the research website gave the message: "These prompts address sensitive topics in China and are likely to be censored by Chinese models."

China and Taiwan have intensified territorial disputes throughout 2024, especially after the inauguration of Taiwanese President Lai Ching-te in May. He defends the island’s sovereignty and is therefore seen by Beijing as a separatist, which has heightened friction and resulted in a significant increase in military operations in the Taiwan Strait.

The Hong Kong protests, which began in 2019, are demonstrations against the local government and China’s influence over the region. The trigger was a proposed extradition bill that would have allowed suspects to be sent to mainland China for trial. In 2024, a Hong Kong court sentenced 45 former politicians and activists to up to 10 years in prison in a trial that weakened the city’s pro-democracy movement and served as yet another sign of the Chinese territory’s crackdown on opponents.

The disclosure of these possible cases of censorship is in the midst of the new AI "space race", as Marc Andreessen himself defined it when he said he was witnessing the "Sputnik moment of artificial intelligence". This is because after DeepSeek launched its AI model with the propaganda that it uses less data and costs a fraction of the price of its competitors, big tech stocks fell on the New York Stock Exchange.

On Monday, the 27th, when the Chinese app reached its peak in downloads on the App Store, chipmaker Nvidia was down 11% in the American market, while Microsoft and Meta were each down 3.3%.

mundophone