This research introduces GECScore, a novel approach that detects AI-generated text by analyzing grammatical error patterns, requiring no training data or access to the source model.

• LLMs tend to produce more grammatically perfect text than humans
• By measuring the grammatical error correction rate between original and corrected text, the system can distinguish human vs. AI authorship
• Achieves impressive detection accuracy across multiple language models and domains
• Works as a black-box detector without needing access to the generative model's parameters

This advancement addresses critical security challenges in detecting AI-generated misinformation and academic dishonesty, offering a practical tool to identify synthetic content without complex training requirements.

Who Wrote This? The Key to Zero-Shot LLM-Generated Text Detection Is GECScore