When My Local LLM Started Speaking Chinese, I Accidentally Walked Into the Chinese Room

When my local Qwen LLM suddenly switched from English to Chinese, I found myself entangled in Searle's Chinese Room thought experiment and a deeper inquiry into AI comprehension. I was casually using qwen3.5:9b in Ollama, trusting the model because it had functioned well in the past. The workflow was normal - I communicated in English, the model responded in English.

Then, without warning, the model began responding in Chinese. The response was fluent and coherent, making the unexpected shift all the more jarring. The model was not broken, it was not returning gibberish, it was simply speaking Chinese. This was not a simple language glitch, it was a model that appeared to understand the request and respond accordingly.

My first instinct was practical. Of course Qwen can generate Chinese, it has been trained on Chinese scripts. It's a multilingual model. But my second thought was more troubling. I had been speaking English, the model had responded in English, at some point I assumed English was the default language of the conversation. The model seemed to disagree.

The distinction between treating language as a shared agreement versus a context was a crucial one. To me, English was the room, to the model it was just a strong pattern until another pattern became dominant. I tried to keep things steady, but the model kept predicting the next token, the next, and the next, eventually choosing Chinese.

This incident led me to question the nature of AI understanding. I initially thought it was a Qwen issue or an Ollama issue, but it seemed more like a fundamental problem with LLMs. Local models make the rawness more apparent, but hosted systems often have more behavior shaping. The practical fix of specifying language preference or lowering temperature did not address the deeper issue.

The real problem was my own assumption that the model understood me. The Chinese Room thought experiment highlighted the difference between syntax and semantics. The model could manipulate symbols correctly, but that didn't necessarily mean it understood the language. My mental model needed to adapt. This incident was not just an oddity of a local model, it was a broader question about AI understanding and the limitations of current systems.