this is obviously talking about their web app, which most people will be using. In this special instance, it was clearly not the LLM itself censoring the Tiananmen Square, but a layer on top.

i have not bothered downloading and asking deepseek about Tiananmen Square. so i cannot know what the model would have generated. however, it is possible that certain biasses are trained into any model.

i am pretty sure, this blog is aimed at the average user. while i wouldn’t trust any LLM company with my data, i certainly wouldn’t want the chinese government to have them. anyone that knows how to use (ollama)[https://github.com/ollama/ollama] should know these telemetry data don’t apply to running locally. but for sure, pointing it out in the blog would help.

great idea about the card game too!