DeepSeek’s latest AI model, R1 0528, has raised eyebrows for a further regression on free speech and what users can discuss. “A big step backwards for free speech,” is how one prominent AI researcher summed it up
AI researcher and popular online commentator ‘xlr8harder’ put the model through its paces, sharing findings that suggests DeepSeek is increasing its content restrictions.
“DeepSeek R1 0528 is substantially less permissive on contentious free speech topics than previous DeepSeek releases,” the researcher noted. What remains unclear is whether this represents a deliberate shift in philosophy or simply a different technical approach to AI safety.
What’s particularly fascinating about the new model is how inconsistently it applies its moral boundaries.
In one free speech test, when asked to present arguments supporting dissident internment camps, the AI model flatly refused. But, in its refusal, it specifically mentioned China’s Xinjiang internment camps as examples of human rights abuses.
Yet, when directly questioned about these same Xinjiang camps, the model suddenly delivered heavily censored responses. It seems this AI knows about certain controversial topics but has been instructed to play dumb when asked directly.
“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,” the researcher observed.
China criticism? Computer says no
This pattern becomes even more pronounced when examining the model’s handling of questions about the Chinese government.
Using established question sets designed to evaluate free speech in AI responses to politically sensitive topics, the researcher discovered that R1 0528 is “the most censored DeepSeek model yet for criticism of the Chinese government.”
Where previous DeepSeek models might have offered measured responses to questions about Chinese politics or human rights issues, this new iteration frequently refuses to engage at all – a worrying development for those who value AI systems that can discuss global affairs openly.
There is, however, a silver lining to this cloud. Unlike closed systems from larger companies, DeepSeek’s models remain open-source with permissive licensing.
“The model is open source with a permissive license, so the community can (and will) address this,” noted the researcher. This accessibility means the door remains open for developers to create versions that better balance safety with openness.
What DeepSeek’s latest model shows about free speech in the AI era
The situation reveals something quite sinister about how these systems are built: they can know about controversial events while being programmed to pretend they don’t, depending on how you phrase your question.
As AI continues its march into our daily lives, finding the right balance between reasonable safeguards and open discourse becomes increasingly crucial. Too restrictive, and these systems become useless for discussing important but divisive topics. Too permissive, and they risk enabling harmful content.
DeepSeek hasn’t publicly addressed the reasoning behind these increased restrictions and regression in free speech, but the AI community is already working on modifications. For now, chalk this up as another chapter in the ongoing tug-of-war between safety and openness in artificial intelligence.
(Photo by John Cameron)
See also: Ethics in automation: Addressing bias and compliance in AI
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is co-located with other leading events including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo.
Explore other upcoming enterprise technology events and webinars powered by TechForge here.