Elevator Pitch

• Traditional ML (TF‑IDF + Linear SVM) can reliably detect mainstream LLM-generated text in early 2026 because it exhibits strong, learnable statistical patterns.

Key Takeaways

• A sentence-level TF‑IDF → LinearSVC classifier trained on paired human vs LLM-regenerated texts reaches ~85% accuracy, and aggregating across sentences boosts confidence for long articles.
• Perplexity-based detection was found impractical and unreliable due to thresholding problems, cost, and deployment/generalization issues.
• A browser-only demo reimplements TF‑IDF + SVM in JavaScript (serverless), trading model size (up to 500k features) for more stable false-positive behavior.

Most Memorable Quotes

• “As of early 2026, mainstream LLM-generated text exhibits strong statistical patterns that can be effectively distinguished from human-written content using traditional machine learning models.”
• “I spent some time trying this method, but results were disappointing—plenty of false positives and false negatives, and no reasonable threshold could be set.”
• “After cleaning up some noise, sentence-level classification still hit ~85% accuracy!”