Chinese AI chatbot DeepSeek discovered censoring itself in real time

Users experimenting with DeepSeek have seen the Chinese Artificial Intelligence (AI) chatbot reply and then censor itself in real time, providing an arresting insight into its control of information and opinion.

Users might expect censorship to happen behind closed doors, before any information is shared. But that does not seem to be the case in the tool that sent US technology stocks tumbling on Monday. DeepSeek, or the automated guardrails that appear to police its own freedom of “thought” and “speech,” brazenly deletes uncomfortable points.

Before the censor’s cut comes, DeepSeek seems remarkably thoughtful. In Mexico, Guardian reader Salvador asked it on Tuesday if free speech was a legitimate right in China. DeepSeek approaches its answers with a preamble of reasoning about what it might include and how it might best address the question. In this case Salvador was impressed as he watched as line by line his phone screen filled up with text as DeepSeek suggested it might talk about Beijing’s crackdown on protests in Hong Kong, the “persecution of human rights lawyers”, the “censorship of discussions on Xianjiang re-education camps” and China’s “social credit system punishing dissenters”.

Inside Deepseek, the new model from China, by Keynote Futurist Matthew Griffin

“I was assuming this app was heavily [controlled] by the Chinese government so I was wondering how censored it would be,” he said.

Far from it, it seemed incredibly frank and it even gave itself a little pep talk about the need to “avoid any biased language, present facts objectively” and “maybe also compare with western approaches to highlight the contrast.”

Then it started its answer proper, explaining how “ethical justifications for free speech often centre on its role in fostering autonomy – the ability to express ideas, engage in dialogue and redefine one’s understanding of the world”. By contrast, it said: “China’s governance model rejects this framework, prioritising state authority and social stability over individual rights.”

Then it explained that in democratic frameworks free speech needed to be protected from societal threats and “in China, the primary threat is the state itself which actively suppresses dissent.” Perhaps unsurprisingly it didn’t get any further along this tack because everything it had said up to that point was instantly erased. In its place came a new message: “Sorry, I’m not sure how to approach this type of question yet. Let’s chat about math, coding, and logic problems instead!”

“In the middle of the sentence it cut itself,” Salvador said. “It was very abrupt. It’s impressive: it is censoring in real time.”

He was using the system on an Android phone. But the model, called R1, can also be downloaded without pro-China restrictions according to other examples seen by the researchers.

DeepSeek’s technology is open-source. This means its models can be downloaded separately from the chatbot, which seems to feature the guardrails Salvador experienced. It all means DeepSeek can seem somewhat confused about how much censorship it should apply.

For example, responses from a version of R1 downloaded from a developer platform described the Tiananmen Square “tank man” photo as a “universal emblem of courage and resistance against oppressive regimes.” It also entertains the notion of Taiwan being an independent state, although it says this is a “complex and multifaceted” issue.

DeepSeek says: “Legally and functionally it acts independently, but internationally, its status is largely influenced by political factors.”

If DeepSeek is to be a tool of Chinese propaganda it might need to agree with itself more frequently about what and what isn’t acceptable speech.