This research introduces statistical tests that can detect LLM-generated text without requiring training data, using mathematical concentration inequalities.

• Achieves high accuracy in distinguishing human vs AI content across multiple LLMs
• Provides theoretical guarantees on false positive rates
• Enables zero-shot detection without needing model-specific training
• Demonstrates practical applications for educational integrity and misinformation prevention

This research is critical for security professionals as it offers robust methods to verify content authenticity, ensure compliance with emerging AI regulations, and protect organizations from unauthorized AI-generated content. The approach works even on new or proprietary LLMs without requiring access to their parameters.

Zero-Shot Statistical Tests for LLM-Generated Text Detection using Finite Sample Concentration Inequalities