Detecting LLM-Generated Texts with “Classical” Machine Learning

	Detecting LLM-Generated Texts with “Classical” Machine Learning(blog.lyc8503.net)
	210 points by uneven9434 19 hours ago \| 153 comments
	tl;dr: A developer trained TF-IDF + LinearSVC classifiers on ~10,000 pre-ChatGPT Chinese texts paired with LLM-regenerated versions from seven models (Gemini, Qwen, GLM, Kimi, Doubao, Deepseek), achieving ~85% per-sentence accuracy using majority voting. The classical ML approach generalized well to unseen models like Claude and GPT-5, with false positive rates under 0.04% on human-written fanfics, while common evasion tactics (translation roundtrips, "de-AI" prompts) only marginally reduced detection. The full model runs client-side in JavaScript via a web demo, suggesting this is likely how commercial AIGC detectors actually work.
	HN Discussion: ↓Detecting LLM text is fundamentally futile; text lacks information density for reliable provenance signals ↓Detection is a losing arms race since models will train against these classifiers ↓Alternative approaches like proof-of-work or measuring writing effort would be more meaningful than detection ↑Classical/small-model classifiers work well and could be practically deployed (e.g., browser extensions, on-device apps) •Translation nuance concern about the author's phrasing in the English version of the post