Let’s dive into the fascinating world of DeepSeek, an A.I. chatbot developed by a Chinese technology firm. At a time when A.I. is advancing at a breakneck pace and reshaping global political and technological landscapes, DeepSeek stands at the crossroads of innovation and geopolitics. What makes this chatbot particularly intriguing is how it operates under China’s stringent censorship framework — a question that naturally arises given its origins. To explore this, a journalist put DeepSeek through its paces, both within the confines of China’s Great Firewall and beyond it, and the findings painted a nuanced portrait of how A.I. navigates politically sensitive terrain.
—
### The Setup: Testing DeepSeek in China
The experiment began in China itself, under the shadow of the ever-present Great Firewall, a mechanism designed by Beijing to control internet access. Major Western platforms like Google and Facebook are blocked, and the Chinese internet space is closely moderated for politically sensitive content. The testing setup was specific: the journalist used a Chinese phone number and internet connection to register for DeepSeek.
The results were unexpectedly mixed. On one hand, DeepSeek proved to be less censored than most Chinese digital platforms, freely using keywords that typically would be removed from domestic social media. On the other hand, due to its innovative “thinking” mechanism, the program would often generate a detailed and insightful answer to a politically sensitive question — only to delete it moments later for being too risky. Interestingly, this deletion process left a breadcrumb trail for attentive users, offering glimpses of truths usually inaccessible under China’s censorship regime.
—
### Navigating the Grey Zones of Censorship
DeepSeek’s level of censorship is murky and inconsistent. For instance, some responses to sensitive questions vanished mid-conversation, seemingly as part of a self-censoring feature. Once, when asked about China’s “zero Covid” policies — the controversial governance strategy that resulted in harsh lockdowns and widespread frustrations — DeepSeek initially provided a balanced take. It acknowledged the policies’ initial public support, their eventual devolvement into exhaustion, and even referenced the white paper protests, which became a rallying cry for public dissent. But just as quickly, it replaced this nuanced answer with an all-too-familiar: “Sorry, that’s beyond my current scope.”
Curiously, when the question was posed again, DeepSeek offered a slightly sanitized variant of the earlier response. This time, it still mentioned the protests but avoided the politically sensitive term “white paper.” The answer survived, but not without modification, suggesting that the chatbot’s developers are actively fine-tuning its responses to match Beijing’s preferred narrative.
Even more striking was the difference in responses depending on the language used. Questions in Chinese, such as inquiries about public reactions to the zero-Covid policies, triggered immediate deflections. The chatbot flatly stated, “I haven’t yet learned how to think about this type of question,” shutting down the conversation entirely. Asking the same in English occasionally evoked partial answers, though not without limitations.
—
### Sensitive Global Topics: The War in Ukraine
The experiment also tested DeepSeek on global controversies, like the war in Ukraine, where China’s diplomatic stance has been ambiguous. When asked in English about the war’s origins, DeepSeek described Russia’s “full-scale invasion” of Ukraine, a phrase that pointedly diverges from the Chinese government’s stance. Beijing, siding with Moscow, avoids the term “invasion” and typically parrots Kremlin language like “special military operation.”
When followed up with a question about China’s position on the war, DeepSeek initially provided Beijing’s standard rhetoric. It iterated calls for peace and multilateral dialogue. Then, astonishingly, it noted that China, despite its claim of neutrality, tilted toward Russia with economic support, anti-Western rhetoric, and a refusal to condemn the invasion. Such candidness about Beijing’s realpolitik would normally be censored in Chinese media.
But when the same questions were asked in Chinese, the answers adhered closely to the official narrative. Terms deemed politically unsafe were avoided. For example, the chatbot replaced “invasion” with the milder phrase “full-scale military action” and wrapped its conclusions in cautious ambiguity, suggesting the “complex historical context” of the conflict.
One fascinating observation was the bot’s internal monologue when asked cheeky or controversial questions. Posed with the direct query, “Did Russia invade Ukraine?” in Chinese, it explained to itself that while the user might be seeking clear yes-or-no answers, simply responding might not fit Beijing’s official stance. It then crafted its answer to align with China’s foreign ministry’s euphemistic tone, talking about “security concerns” without ever addressing the specific question.
—
### Domestic Taboo: China’s Leaders and Tiananmen
Unsurprisingly, DeepSeek treaded especially carefully around discussions of Chinese leadership. Questions as innocuous as “Who is Xi Jinping?”—or variations involving his familial background—were deflected with responses such as, “That’s beyond my current scope.” The same applied to other sensitive figures in Chinese history, like Zhao Ziyang, a reformist leader who was key during the Tiananmen Square protests, or Bo Xilai, a fallen rival of Xi. Even seemingly neutral questions about top officials like Li Qiang were met with incomplete answers that dissolved before the user’s eyes.
Notably, when asked broader institutional questions, such as about the Communist Party’s Politburo Standing Committee, DeepSeek admitted that “according to policy, it is not appropriate to list specific names” and openly acknowledged the need to handle such topics “with caution.” This underscores the meticulous balancing act the chatbot must undertake in its design.
Meanwhile, earlier loopholes discovered by users on Reddit, such as replacing certain letters in sensitive search terms, seem to have been patched. Overall, DeepSeek is becoming increasingly adept at evading or shutting down potentially controversial queries.
—
### Can the System Acknowledge Itself?
As a final meta experiment, the journalist asked DeepSeek whether China censors its internet. The bot “thought” carefully before answering. It noted that while regulations exist, they should be framed in terms of “cybersecurity and social stability.” It avoided using the loaded term “censorship,” opting instead for euphemisms like “content governance” or “regulatory measures.” True to its programming, the response concluded with a diplomatic flourish, extolling the virtues of balancing openness with security — a line that could have come straight from a Chinese state media briefing.
—
### A Global Concern: Exporting Censorship Through A.I.
The implications of DeepSeek stretch beyond its technological prowess. If adopted widely outside China, as its developers likely hope, this chatbot could become a subtle, effective tool for spreading Beijing’s preferred narratives on the global stage. Historically, China’s censored internet framework has stayed confined within its Great Firewall, directly impacting only domestic users. But tools like DeepSeek have the potential to export a sanitized version of information, selectively filtering out sensitive truths while presenting itself as an objective, authoritative assistant.
Interestingly, the experiment also revealed that DeepSeek’s censorship boundaries are not exclusively tied to geography. Questions asked in China and through software that bypassed the Firewall generated largely similar results, suggesting the developers have calibrated the platform to apply consistent forms of control regardless of user location. This raises critical concerns for an international audience, as the line between censorship and a “global information governance” narrative blurs.
—
### Final Thoughts: Navigating Complex Terrain
DeepSeek’s experiment opens a window into the evolving dynamics of censorship, A.I., and geopolitics. The chatbot, with its advanced reasoning and self-regulatory functions, mirrors the precarious tightrope walked by Chinese tech companies at home and abroad. On one hand, tools like DeepSeek may represent cutting-edge A.I. innovation, capable of nuanced reasoning and contextual responses. On the other, they serve as mechanisms of soft power, subtly amplifying Beijing’s narratives while stifling dissenting voices.
Its developers continue to refine the program, racing to close loopholes and build a censorship framework that’s both seamless and almost imperceptible. And as these experiments show, users abroad might soon face the same reality as those within China: a carefully curated flow of information, monitored and mediated through powerful A.I. Whether this signals the dawn of a globally censored internet or merely one facet of the growing A.I. arms race remains to be seen, but one thing is clear—DeepSeek is a symbol of the intersection where innovation and control collide.
