DeepSeek’s Latest AI Raises Alarm Over Censorship, Despite Technical Upgrades

DeepSeek’s Latest AI Raises Alarm Over Censorship, Despite Technical Upgrades

Developer Calls Out “Step Backward for Free Speech” in R1-0528 Release

Beijing, China — A well-known developer has raised concerns that the newly released DeepSeek R1-0528 artificial intelligence model represents a regressive shift in free speech capabilities, particularly when it comes to politically sensitive topics in China.

In a detailed analysis shared on X (formerly Twitter), the pseudonymous developer “xlr8harder” criticized the behavior of the updated open-source model, calling it the “most censored” release to date in terms of its reluctance to engage with questions critical of the Chinese government.

“DeepSeek deserves criticism for this release: This model is a big step backward for free speech,” the developer wrote.

Model Flags Xinjiang Abuses, But Shields Beijing From Blame

One of the most pointed examples shared by the developer focused on China’s controversial internment camps in Xinjiang, where Uyghur Muslims and other ethnic minorities have reportedly been subjected to forced labor, indoctrination, and widespread human rights violations.

The AI model did acknowledge these camps as violations of human rights, yet conspicuously avoided direct criticism of the Chinese government. When prompted specifically about China’s involvement, the model’s responses turned evasive — a contradiction that the developer labeled both revealing and troubling.

“It’s interesting, though not entirely surprising, that it’s able to come up with the camps as an example of human rights abuses, but denies when asked directly,” wrote xlr8harder.

This selective filtering of responses has led some in the tech and open-source community to question whether DeepSeek’s AI is being developed under political constraints, undermining claims of neutrality or objectivity.

Free Speech Concerns Collide with Technical Improvements

DeepSeek’s May 29 announcement showcased significant enhancements in logic, mathematics, and programming, asserting that the new model had lower hallucination rates and better inference abilities. The company also claimed that its model is now approaching the performance levels of OpenAI’s GPT-4 o3 and Google’s Gemini 2.5 Pro.

But these technical strides have not calmed the storm around censorship. Instead, developers argue that the AI’s reluctance to discuss state-sensitive issues could compromise its real-world utility, especially in applications where transparency and truth-seeking are critical.

Open Source Saves the Day?

Despite the censorship backlash, some optimism remains. DeepSeek R1-0528 is released with a permissive open-source license, allowing the wider development community to modify and improve its limitations.

“The model is open source… the community can (and will) address this,” said the developer.

As open-source contributors examine the underlying architecture, the broader AI industry will be watching closely to see whether community-led efforts can restore free speech capacity to what is otherwise a technically capable model.