ludo_PriceleSS

As our community knows, our products are embedded with ML and AI. Our AI team found some strange outputs with DeepSeek Input: censorship in China 1. The web version I tested started writing “Yes, China does use censorship and control of information. This system is known as the Great China Firewall. They block access to certain sites and resources, as well as removing some articles and data inside the country…” and after a few PARAGRAPHS of confirming the China censorship it was removed and replaced with: "Sorry, that's beyond my current scope. Let’s talk about something else".  2. The Mobile version of the censorship question AND China policies provides an in-depth response but then shuts down to : "Sorry I'm not sure how to approach this type of question yet". You can test this yourself. 3. What other AI is doing post-processing after the text is already printed? Is the post-process another team? Otherwise this censorship could be easily achieved by just training the model on the pre-validated/pre-filtered data scope. Doing it in this hacky (and quite obvious) way doesn’t make any sense. 4. When you ask the same questions to GPT and DeepSeek they provide surprisingly similar results. And this is not even talking that much about the core, but about training data. It’s known that GPT was using mostly english-speaking data, which is why (at least in the original version) it was highly skewed towards a western point of view. If DeepSeek was developed in China…why wouldn’t they train it on Chinese data first? And if the training data set was different, you should expect different results (not only on political topics, but in general). 5. Could they have mutated the GPT code or rewritten the code of ChatGPT (or any other AI) or they could post a code that looks like it's working but not really? Something strange is going on here. Would love to hear from different researchers. $VRA