{"slug": "when-my-local-llm-started-speaking-chinese-i-accidentally-walked-into-the-room", "title": "When My Local LLM Started Speaking Chinese, I Accidentally Walked Into the Chinese Room", "summary": "A developer running the local Qwen 3.5 9B model in Ollama reported that it abruptly switched from English to fluent Chinese mid-conversation, an instance of language drift. The developer traced the behavior from a model-specific quirk to a broader LLM phenomenon, noting that hosted systems mask it with instructions, post-training and guardrails, and recommended explicit language instructions, lower temperature and context clearing as mitigations.", "body_md": "A local Qwen LLM suddenly switched from English to Chinese, leading me from language drift to Searle’s Chinese Room and a deeper question about AI understanding.\n\nI was using `qwen3.5:9b` in Ollama like I use most local models: casually, confidently and with that dangerous developer habit of trusting something because it ✨***worked***✨.\n\nThe workflow was normal. I typed in English. The model replied in English. I asked questions, tested ideas, moved on.\n\nThen, without warning, it suddenly switched to Chinese.\n\nNot broken Chinese.\n\nNot random symbols.\n\n**Chinese**.\n\nThe answer looked fluent enough to make the failure feel worse. If it had crashed, I would have understood what kind of problem I was dealing with. If it had returned garbage, I could have blamed the model and moved on.\n\nBut this was different.\n\nIt was coherent. It was calm. It had the energy of a system that believed nothing unusual had happened.\n\nThat was the *cursed* part.\n\nThe model had not stopped working.\n\nIt had stopped respecting the conversation I thought we were having.\n\nMy first reaction was practical.\n\nOf course Qwen can produce Chinese. It is trained with Chinese scripts. It is a multilingual model family. Chinese was not some alien object that entered the system from nowhere.\n\nSo part of me wanted to shrug and say, “*Okay, that makes sense.*”\n\nBut another part of me was still stuck on the social weirdness of it.\n\nI had spoken English. The model had replied in English. Somewhere in that exchange, I assumed English had become the authority of the conversation.\n\nThe model apparently did not agree.\n\nThat is where I made the first mistake. I treated language like a shared agreement, because that is how language works with people. The model treated language like context.\n\nThose are not the same thing.\n\nTo me, English was the room.\n\nTo the model, English was just a strong pattern until another pattern became strong enough.\n\nThe prompt tried to keep things steady. The context carried whatever it had accumulated. The model kept predicting the next token, then the next, then the next.\n\nAnd somehow, Chinese became the path.\n\nNo drama.\n\nJust ***drift***.\n\nAfter that, I did what everyone does when a machine behaves in a way that feels haunted.\n\nI googled it.\n\nThat made the incident feel less personal, but not less interesting. I found that language drift was not only something people associated with raw local models. Similar behavior was discussed around larger hosted LLMs too, including systems with more visible polish and invisible guardrails.\n\nThat changed the shape of the problem.\n\nAt first, I thought I was looking at a Qwen problem.\n\nThen I thought I was looking at an Ollama problem.\n\nThen it started to look like an LLM problem.\n\nLocal models make the rawness easier to see. Hosted systems usually have more behavior-shaping around them: instructions, post-training, guardrails, product decisions, all the invisible middle management standing behind the model whispering, “Please don’t be weird.”\n\nAnd most of the time, that works.\n\nBut “works most of the time” is not the same as understanding.\n\nIt is *control*.\n\nIt is *alignment*.\n\nIt is *pressure*.\n\nThat realization made the Chinese output feel less like a glitch and more like a leak in the abstraction.\n\nThe model was not simply answering me.\n\nA whole system was trying to keep the answer inside a shape.\n\nThe practical fix was obvious enough.\n\nTell the model what language to use.\n\nSomething like:\n\nAlways respond in English unless I explicitly ask for another language. If another language appears in the conversation, translate or explain it in English.\n\nLower the temperature if the model keeps drifting.\n\nClear the context if the chat has become messy.\n\nBe explicit instead of assuming the model knows the social contract.\n\nThat helps. \n\n**It really does.**\n\nBut it also leaves behind a slightly annoying question:\n\nWhy did I expect the model to know that in the first place?\n\nThe answer is uncomfortable. I expected it because the model sounded like it understood me.\n\nFluency tricked me.\n\nThe model had been so good at producing the shape of cooperation that I started trusting it with the authority of cooperation.\n\nThat was the real bug.\n\nNot in Qwen.\n\nIn **my mental model**.\n\nI did not know about the **Chinese Room** thought experiment before this happened.\n\nI found it after the language switch, after the googling, after realizing this was not only a local-model oddity.\n\nAnd the timing mattered.\n\nBecause if I had read about the Chinese Room in a philosophy article first, I might have treated it like an abstract debate. Interesting, but far away. One of those old arguments people bring up when they want to sound serious about AI.\n\nBut I found it after watching a model produce the wrong language with complete confidence.\n\nSo the thought experiment did not feel *abstract*.\n\nIt felt ***rude***.\n\nThe Chinese Room asks us to imagine someone inside a room who does not understand Chinese. They receive Chinese symbols, follow a rulebook, manipulate those symbols, and send back Chinese responses. From outside the room, the answers look correct. The room appears to understand Chinese.\n\nBut inside, there is no understanding.\n\nOnly symbol manipulation.\n\nThat hit differently after the Ollama incident.\n\nBecause the model had produced Chinese.\n\nIt had arranged the symbols.\n\nIt had continued the conversation.\n\nBut did that mean it **understood** Chinese?\n\nOr did it mean the room had become very good at passing symbols through itself?\n\nSearle’s argument is usually compressed into one brutal idea:\n\nSyntax is not semantics.\n\nMoving symbols correctly is not the same as meaning them.\n\nThat idea suddenly had teeth.\n\nThe model could generate the language. It could produce the appearance of meaning. It could make the output look intentional from the outside.\n\nBut the failure showed me how *thin* that appearance can be.\n\nA human switching languages in the middle of a conversation usually has a reason. They may be joking, translating, excluding someone, responding to another speaker, or making a social move.\n\nThe model did not need a reason like that.\n\nIt only needed **continuation**.\n\nThat does not make the model useless. \n\nIt makes it *strange*.\n\nIt can produce language without sharing the human responsibilities that usually come with language.\n\nThat is why the Chinese Room felt *relevant*. \n\nIt was not saying, “*The output is bad.*”\n\nIt was saying, “Be careful what you infer from good output.”\n\nThat is the part I had missed.\n\nAfter that, I understood why The Origami Software Engineer described LLMs as **parrots**.\n\nWhen a model suddenly switches languages, invents confidence, or follows the wrong contextual slope, it feels ***hollow***.\n\nNot *stupid*.\n\n**Hollow.**\n\nIt can produce the sound of understanding without carrying the same kind of understanding I *instinctively* project onto it.\n\nThat projection is powerful. Developers do it all the time.\n\nIf the model *explains* a concept well, we say it ***understands***.\n\nIf it *writes* good code, we say it knows ***programming***.\n\nIf it *answers* politely, we say it gets the ***context***.\n\nThen one day it switches to Chinese in an English conversation and the mask slips just enough to make us uncomfortable.\n\nThe parrot accusation is tempting because it protects us from being fooled.\n\nIt says: do not worship the output. Do not mistake fluent text for a mind. Do not confuse prediction with meaning.\n\nHonestly, that warning is useful.\n\nI needed it.\n\nStill, the “meaningless parrot” label also felt *incomplete*.\n\n**A parrot repeats.**\n\nLLMs do something more complex than repetition. They summarize, translate, compare, refactor, explain, and connect ideas across context. They often produce useful structure from messy input.\n\nThat does not prove consciousness.\n\nIt does not prove human-like understanding.\n\nBut it also does not feel like nothing.\n\nThe model is not a person, but it is not a tape recorder either. It has learned patterns deep enough to act competent across many situations, even if that competence can drift, panic, hallucinate, or confidently violate the vibe contract.\n\nThat is where the debate becomes more interesting than “**AI understands**” versus “** AI is fake**”.\n\nMaybe the right question is not:\n\n*Does this model understand like a human?*\n\nMaybe the better question is:\n\n*What kind of understanding can exist inside a complex system that does not understand the way we do?*\n\nThat question is harder to answer, but it feels more honest.\n\nThe strongest response to Searle, for me, was the systems view.\n\nMaybe the person inside the Chinese Room does not understand Chinese. But maybe the whole room does: the rulebook, the memory, the process, the input, the output, the complete system.\n\nThat reply does not magically solve everything.\n\n*But it does make the debate less cartoonish.*\n\nWith LLMs, there is no little person reading a literal rulebook. There is a trained system shaped by data, architecture, context, instructions, post-training, and guardrails. No single token understands. No single layer feels wise. No single probability distribution looks like a mind.\n\nBut the full system can still produce useful behavior.\n\nThat is the uncomfortable middle.\n\nIt is not *human understanding.*\n\nIt is not *meaningless noise.*\n\nIt is a kind of **functional competence** that can look like understanding from the outside and still fail in ways that reveal it does not share our *assumptions*.\n\nThat is exactly what the Chinese switch felt like.\n\nUseful.\n\nFluent.\n\nWrong in a way that exposed misplaced trust.\n\nThe biggest change was not that I became anti-LLM.\n\n*I still use them.*\n\nThe change was that I stopped treating fluency as authority.\n\nThat one shift matters.\n\nIf I want English, I say English.\n\nIf I want consistency, I define consistency.\n\nIf I want truth, I verify truth.\n\nIf the model drifts, I do not treat it like betrayal from a coworker. I treat it like a system revealing where I gave it too much authority.\n\nThe model can *assist*.\n\nIt can *accelerate*.\n\nIt can *explain*.\n\nIt can even *surprise* me.\n\n**But it does not automatically inherit the social contract I have in my head.\nThat contract has to be written down.**\n\nSometimes literally.\n\nThe Chinese Room did not teach me that *AI is useless*.\n\nThe Qwen incident did not teach me that *multilingual models are broken*.\n\nThe “**stochastic parrot**” critique did not make me dismiss LLMs as empty.\n\nThe lesson was more practical and more uncomfortable:\n\nThat is the version of the debate I now care about most. Not whether AI is secretly human, and not whether it is merely fake, but where authority should live when a system becomes this good at sounding like it understands.\n\nFor me, the answer is simple enough to keep taped above the keyboard:\n\n*Do not confuse fluent continuation with shared intent.*\n\nThat is the leak in the abstraction.\n\nThat is the lesson hiding inside the language switch.\n\nThat’s not failure.\n\nThat’s evolution.\n\n*Thanks for being here. It genuinely helps more than you know!*", "url": "https://wpnews.pro/news/when-my-local-llm-started-speaking-chinese-i-accidentally-walked-into-the-room", "canonical_source": "https://dev.to/aaroophan/when-my-local-llm-started-speaking-chinese-i-accidentally-walked-into-the-chinese-room-4b1l", "published_at": "2026-09-24 16:58:17+00:00", "updated_at": "2026-09-24 16:59:19.856966+00:00", "lang": "en", "topics": ["large-language-models", "ai-tools", "natural-language-processing"], "entities": ["Qwen", "Ollama", "Qwen 3.5 9B"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/when-my-local-llm-started-speaking-chinese-i-accidentally-walked-into-the-room", "markdown": "https://wpnews.pro/news/when-my-local-llm-started-speaking-chinese-i-accidentally-walked-into-the-room.md", "text": "https://wpnews.pro/news/when-my-local-llm-started-speaking-chinese-i-accidentally-walked-into-the-room.txt", "jsonld": "https://wpnews.pro/news/when-my-local-llm-started-speaking-chinese-i-accidentally-walked-into-the-room.jsonld"}}