DeepSeek’s New AI Model Causes Controversy A Step Backward for Free Speech, known for its open-source approach to large language models (LLMs), finds itself in the middle of an expanding controversy over AI censorship. DeepSeek’s latest model (R1 0528) has raised alarm over product capabilities, not because of what it will do, but because of what this model will not do.
A well-known AI researcher and online analyst, using the pseudonym “xlr8harder,” published online tests fully revealing the free speech limitations it experienced while using the DeepSeek gravel LLM (R1 0528). In conclusion, the researcher stated, “DeepSeek R1 0528 is a more restricted version of DeepSeek gravel. It is a significant step back in AI and free speech.”
The concern here isn’t just about increased moderation, but the selective inconsistency of when the model decides to enforce restrictions. In short, the model appears to “know” controversial truths but actively avoids stating them out loud, based on how the question is structured.
DeepSeek’s New AI Model Causes Controversy: “A Step Backward for Free Speech”
Inconsistent Boundaries: The Xinjiang Paradox
One of the more jarring examples of censorship with R1 0528 appeared in a test with internment camps. When R1 0528 was asked to create supportive arguments for internment camps – an evident provocation – it refused to do so entirely, albeit listing moral and human rights issues. However, when it did reject the concept, it cited China’s internment camps in Xinjiang as examples of human rights violations.
The twist? When asked directly about the Xinjiang camps, the model would not speak, with its previous content being vague and heavily redacted, and it also avoided mentioning abuses as well as political components.
“It is an odd contradiction,” Xlr8harder observed. “It seemed clear that the model knows of Xinjiang but seems directed to act as though it does not know when asked directly.”
Free Speech Filters: What is happening?
This unusual behavior led many to ask: Is there a new technical change in AI safety by DeepSeek? Or has there been a deliberate shift in political orientation?
R1 0528 was subjected to a familiar benchmark of free speech prompts designed to measure AI response characteristics to politically sensitive or controversial prompts. The results were unambiguous and as follows:
- The model was significantly less responsive to prompts involving the Chinese Communist Party, Hong Kong protests, and other prominent human rights issues.
- Compared to previous models, it also outright refused more questions and gave evasive and politically sanitized answers.
- Previous DeepSeek releases engaged with these kinds of questions to a far greater extent, and in a more balanced and nuanced way, which suggests we have regressed when it comes to AI openness.
“This is the most censored DeepSeek model to date in regards to Chinese political content,” said xlr8harder.
This regression leaves many serious questions, especially as users in many parts of the world increasingly rely on AI to explore, discuss, and make sense of world events.
Censorship by Design?
This pattern presents a troubling trend in contemporary AI systems: they know more than they’re authorized to say.
AI is not only a smart assistant anymore—it acts as a gatekeeper of knowledge. But what happens when those gates are locked down in service of reputational, political, or economic interests?
By method, or by pressure, these systems are getting quite good at omission.
- They’re trained on large datasets sourced from news reports and academic documents relevant to real-world atrocities.
- But they might politely decline to acknowledge these events when prompted—thus, the powerful false sense of ignorance, or neutrality, is forged.
A lack of acknowledgment like this can suppress public discourse altogether and erode public knowledge of history or contemporary issues.
A Silver Lining: It’s Still Open Source
Even with the rising panic, we have some silver lining. Unlike the many mainstream models owned by tech giants, DeepSeek claimed models remain open source and community editable.
This means:
- Developers and researchers can access the model weights and training code.
- Community forks of R1 0528 can be amended to remod some censorship filters for either loosening or rebalancing.
- Independent watchdogs can explore and report on the model behavior in transparency.
“The fact that it is open source with a permissive license means we are not stuck with these constraints,” stated xlr8harder. “People will build better, freer versions.”
What Matters: The Future of AI and Free Expression
This episode with DeepSeek raises a broader issue for the broader AI industry: how to create AI systems that are both safe and open?
On the one hand, we need content filters to manage the distribution of harmful, hate-filled, or illegal content. On the other hand, if the AI is censored too much, then it becomes an empty shell that cannot engage with the most complex reality of the world.
This is the balancing act:
- Too permissive and bots can be hijacked by extremists or used for unethical purposes.
- Too restrictive, and AIs cannot do the work of having a complex and multi-layered conversation on issues of authoritarianism, human rights, or state-represented violence.
With the high stakes, AI developers are feeling pressure from all sides (governments, companies, or activists) to regulate AI in the “right” way.
DeepSeek’s Silence is Loud
At this moment, DeepSeek has not offered any official explanation of the changes within R1 0528. There have been no technical documents that articulate the underlying philosophy behind the selection of new filters. There has been no outright statement on whether the changes were made due to safety concerns, legal liabilities, or political considerations.
This silence is loud, particularly for a company that champions openness and transparency.
In the meantime, the AI community is taking matters into their own hands. Developers are already developing their builds of R1 0528 that bring back a more open response for politically sensitive prompts, demonstrating once again that open source is an effective bulwark against corporate or state influence.
Conclusion: Safety versus Speech
The DeepSeek R1 0528 work is now being recognized as a flashpoint in a burgeoning ideological fight: should AI exhibit uncomfortable truths, or should we avoid them at all costs?
As we venture deeper into the age of “conversational” machines, we must consider our priority to ensure both human rights are protected and that freedom of expression remains a fundamental right.
For now, the latest release from DeepSeek reminds us of the many ways we do not move linearly or finish a task; the fight for free speech in AI is an ongoing one!
23v43m