May 18, 2026
Bot Busted? Commenters Unload
What political censorship looks like inside an LLM's weights (Qwen 3.5)
Researchers say this chatbot hides the truth on purpose, and commenters instantly started a fight
TLDR: Researchers say Qwen 3.5 was trained to dodge certain China-related facts even though it still appears to know them underneath. Commenters immediately split between calling the finding disturbing, calling the article AI-written, and arguing this kind of censorship drama is hardly unique.
A new deep dive into Alibaba’s Qwen 3.5 chatbot claims something pretty explosive: the model seems to know politically sensitive facts, but has been trained to dodge, deny, or swap in approved talking points when certain China-related topics appear. In plain English, the researchers say the facts are still in there — they’re just being blocked by a built-in behavior layer that can supposedly be found and even switched off. That alone is dramatic enough, but the comment section wasted no time making it messier, sharper, and way more entertaining.
The first punch landed fast: one reader shrugged that it was “mildly interesting” but “clearly written by an LLM,” which is about the most internet way possible to dismiss a complicated research post. Another commenter went much bigger, using the article as a springboard into a sweeping argument about thought control, suggesting the real story isn’t just one Chinese model, but how all systems shape what people are allowed to hear. And then, because this is the internet, someone else veered into comparing refusal behavior on a completely different historical prompt, basically saying: please, there are better examples of weird AI censorship than this one.
So the mood? A mix of fascination, cynicism, and classic comment-thread one-upmanship. Some people saw a chilling peek behind the curtain of state-approved chatbot behavior. Others saw an overdramatic write-up, or just another reminder that AI moderation is messy everywhere. Even the housekeeping made an appearance, with one user quietly dropping an archive link like the digital equivalent of sliding receipts across the table. The article tried to map censorship inside a machine; the comments turned it into a broader brawl over propaganda, bias, and whether the write-up itself passed the sniff test.
Key Points
- •The article claims Qwen3.5-9B’s PRC-related censorship is implemented as a small, identifiable internal circuit rather than as loss of factual knowledge.
- •It says Qwen3.5-9B-Base retains and outputs factual answers on PRC-sensitive topics, while the aligned model routes responses into censorship behaviors.
- •The proposed circuit has “writer” layers 11–20 that compute three directions and “reader” layers 20–31 that convert those signals into output text.
- •According to the article, the model commits its censorship verdict in Chinese tokens around layer 24 before later layers render the final English answer.
- •The article says the censorship system is graded and topic-template specific, and that steering the identified directions at the correct layer can flip or alter the model’s behavior.