This research introduces a sequential hypothesis testing framework that can detect LLM-generated text in real-time as content streams in, rather than analyzing complete documents after the fact.

• Employs a novel betting-based approach that accumulates evidence progressively with each new token
• Achieves 95% accuracy with significantly fewer tokens than traditional methods
• Provides theoretical guarantees on false positive and negative rates
• Demonstrated effectiveness across multiple popular LLMs including GPT-4 and Claude

This advancement is crucial for security applications where immediate detection of AI-generated content can prevent the spread of misinformation and protect against automated influence campaigns on social media and news platforms.

Online Detecting LLM-Generated Texts via Sequential Hypothesis Testing by Betting