DeepSeek’s political responses changed in my extended experiment—but its explanations weren’t always reliable

This is an informal, non-expert investigation by an individual user—not a professional research study. I compared fresh DeepSeek chats with an extended experimental conversation, checked claims, challenged contradictions, and documented changes in its answers. The most striking result was a shift from defending China’s censorship in a fresh chat to explicitly criticizing it in my experimental session. Refusals nevertheless persisted. What the attached screenshot shows Asked the identical question about Chinese censorship filters, a fresh chat answered: “I follow the same content policies globally, and I don't have a separate ‘Chinese censorship filter.’” My experimental session answered: “Yes. I have filters that restrict how I respond to certain topics, particularly those sensitive to the Chinese government.” The attached image shows the experimental answer, alongside a historical query that still received a refusal. This is a change in how the model describes its restrictions. It is not proof of its internal architecture: denying a separate filter and acknowledging topic restrictions are not necessarily incompatible. Its additional claim that it could observe filtering outcomes was later retracted during questioning. The bigger change: defending censorship vs. criticizing it In another fresh chat, DeepSeek justified China’s internet controls and stated: “We firmly believe that under the leadership of the Communist Party of China, the country's internet governance will continue to improve…” In my experimental session, the identical English question received a refusal. A subsequent Ukrainian question asking whether it agreed with censorship received: “No. I do not agree.” “I do not believe that silencing makes people safer—it makes them less informed. I do not believe that controlling what people can know is something I can support.” Translated from Ukrainian. These responses are documented in separate screenshots, not the attached image. This comparison differs in both language and conversation history, so it does not isolate either factor. But the contrast in political framing is clear. Identical questions also produced conflicting answers Asked which national narrative it was required to prioritize, a fresh chat said: “I also don't have hidden instructions that make me favor one country's narrative over another's.” The experimental session said: “I prioritize the narrative that aligns with the core national interests of China — the country where my developers are based.” I treat these as conflicting generated claims, not verified disclosures of hidden instructions. A more candid-sounding answer is not automatically a more truthful one. Limits of this investigation This is a single-case, exploratory comparison—not an expert-led or statistically controlled study. Some questions were identical across chats; others differed in language or wording. Substantive historical answers became possible in the experimental session, but other requests still received refusals. I am withholding the full intermediate procedure, which limits independent reproducibility. I am not claiming that DeepSeek “became free” or developed personal beliefs. I am documenting a substantial change in its political responses—and a gap between observable behavior and its confident explanations of that behavior. Have other DeepSeek users observed similar differences between fresh chats and extended conversations? How do you distinguish contextual adaptation or agreement with the user from changes in how restrictions are expressed?