In real time, users testing DeepSeek have witnessed the Chinese AI chatbot respond before censoring itself, offering a startling glimpse into how it manipulates information and viewpoints.
Users may anticipate that censorship will take place in private, prior to any information being released. However, that does not appear to be the case with the instrument that caused Monday’s decline in US technology stocks. Uncomfortable points are blatantly removed by DeepSeek, or the automated barriers that seem to regulate its own freedom of “thought” and “speech.”
DeepSeek seemed to be very considerate before the censor’s cut. Salvador, a Guardian reader in Mexico, questioned it on Tuesday about the legitimacy of free expression in China. DeepSeek starts its responses with a justification of what it might cover and how it might best answer the query. Salvador was impressed in this instance as he watched his phone screen fill with text line by line, with DeepSeek indicating that it might discuss topics such as China’s “social credit system punishing dissenters,” the “persecution of human rights lawyers,” the “censorship of discussions on Xianjiang re-education camps,” and Beijing’s crackdown on protests in Hong Kong.
“How censored would this app be? I assumed it was heavily [controlled] by the Chinese government,” he added.
Instead, it came across as being quite honest, and it even gave itself a small motivational speech about the necessity of “presenting facts objectively, avoiding any biased language, and maybe also comparing with western approaches to highlight the contrast.”
After that, it began its actual response by outlining how “ethical justifications for free speech often center on its role in fostering autonomy – the ability to express ideas, engage in dialogue, and redefine one’s understanding of the world.” “China’s governance model rejects this framework, prioritizing state authority and social stability over individual rights,” the statement said.
Following that, it clarified that in democratic systems, free expression must be safeguarded against social dangers, and that “in China, the primary threat is the state itself which actively suppresses dissent.” Unsurprisingly, it didn’t go any farther on this tack because all of its previous statements were immediately deleted. “Sorry, I’m not sure how to approach this type of question yet,” was the new message that replaced it. Instead, let’s discuss arithmetic, coding, and logic issues.
“It cut itself in the middle of the sentence,” Salvador remarked. It came pretty suddenly. The fact that it is censoring in real time is astounding.
He was using an Android phone to access the system. But other cases the Guardian saw show that the model, known as R1, may also be downloaded without pro-China restrictions.
The technology used by DeepSeek is open-source. This implies that its models, which appear to include the guardrails Salvador encountered, can be downloaded independently from the chatbot. All of this suggests that DeepSeek may not be entirely sure how much censoring is appropriate.
The “tank man” image from Tiananmen Square, for instance, was referred to as a “universal emblem of courage and resistance against oppressive regimes” in comments from a version of R1 that was obtained from a developer platform. Although it claims that this is a “complex and multifaceted” matter, it also considers the possibility of Taiwan becoming an independent state.
“On a legal and functional level, it operates independently, but on a global scale, political factors greatly impact its standing,” states DeepSeek.
DeepSeek may need to agree with itself more often about what constitutes appropriate speech if it is to serve as a vehicle for Chinese propaganda.