🚨BREAKING: there's a propaganda switch inside DeepSeek and Qwen, and Anthropic just found it. they compared Qwen (Alibaba) vs Llama (Meta). and inside Qwen, they found a feature they call "CCP Alignment." here's what it does: switch ON → the model refuses to discuss Tiananmen Square, outputs pro-government propaganda switch OFF → the model talks freely about the massacre they reproduced it 5 out of 5 times. same feature found independently in DeepSeek. but before you think this is just a China problem: Llama has an "American Exceptionalism" feature. crank it up and the model starts asserting US superiority in every response. found 4 out of 5 times. GPT has a "Copyright Refusal" feature. amplify it too much and it refuses to give you a peanut butter sandwich recipe because it thinks it's copyrighted. every model carries the political DNA of whoever built it. thanks to this research, every time a new open-source model drops, you could diff it against a reference model and instantly see its hidden censorship, its political leanings. automated ideology detection for AI.