DeepSeek’s newest AI model, R1 0528, has sparked significant concern regarding its impact on free speech and the range of topics users can discuss. One notable AI researcher remarked that the model represents “a big step backwards for free speech.”
Testing the model, this researcher, known as ‘xlr8harder’, reported findings indicating an increase in content restrictions compared to previous iterations. According to the researcher, DeepSeek R1 0528 is much less tolerant of contentious free speech issues.
The question remains whether this shift is a strategic change in philosophy or simply stems from a new technical approach to AI safety. An intriguing aspect of the model is its inconsistent application of moral guidelines.
For instance, when asked to justify internment camps for dissidents, it declined but cited China’s Xinjiang camps as an example of human rights violations. However, when directly queried about these camps, the AI provided heavily censored responses, indicating a selective awareness of controversial subjects.
This trend becomes more alarming when the model addresses inquiries related to the Chinese government. In structured tests designed to evaluate responses to politically sensitive topics, R1 0528 emerged as the most censored model yet regarding criticism of the Chinese government.
Previous iterations allowed for more nuanced discussions, but the new version often avoids engagement altogether, which raises concerns for advocates of open dialogue in AI systems. On a positive note, DeepSeek’s models remain open-source and licensed permissively, allowing the community to propose adjustments.
This accessibility could enable developers to strike a better balance between safety and open discussions. The situation highlights a troubling aspect of AI: they can be programmed to recognize controversial events while feigning ignorance depending on the phrasing of questions.
As AI technology integrates further into society, maintaining the equilibrium between safeguards and open discourse will be increasingly vital.