Hasty Briefsbeta

Bilingual

Detecting LLM-Generated Texts with “Classical” Machine Learning

3 hours ago
  • #Machine Learning
  • #LLM Analysis
  • #AI Detection
  • Mainstream LLM-generated text as of early 2026 shows strong statistical patterns, detectable by traditional machine learning models.
  • An online demo using a simple SVM classifier achieves ~85% single-sentence accuracy, trained on human and LLM-generated data.
  • Data was gathered from pre-2022 human texts and LLM-regenerated summaries via multiple low-cost/free APIs from models like Gemini, Qwen, and GLM.
  • Seven binary classifiers were trained, using majority voting; detection remains robust against common anti-detection methods like translation or rewriting prompts.
  • Testing on unseen models and human texts shows low false positives; analysis of Lofter content suggests significant undisclosed AI-generated material.
  • The author expresses skepticism about LLMs' creativity, viewing AI-generated content as repetitive and shallow, and reflects on broader implications of AI misuse.