Researchers have discovered that Large Language Models (LLMs) leave distinct idiosyncrasies in their outputs, creating identifiable patterns that can distinguish between different AI models.

• Models can be identified with high accuracy (often >95%) from their text outputs alone
• Simple fine-tuning of text embedding models enables effective LLM classification
• These patterns persist even across different prompts and content types
• The findings have significant implications for AI authentication and security

This research provides valuable tools for model attribution, detecting AI-generated content, and establishing provenance in security-critical applications, enabling better governance of AI systems.

Idiosyncrasies in Large Language Models