Who's Behind the Text? Robust Model Authorship Attribution

AI-generated text can now be traced back to its source model: The paper introduces ATLAS, a system that can identify which large language model produced a given piece of text by analyzing how likely that text is under each model's generation patterns. This addresses a practical need as LLM misuse becomes more common.

Current attribution methods fail when encountering new situations: Existing systems for identifying AI-generated text perform poorly when faced with texts from unfamiliar domains, different writing contexts, or newer versions of models they weren't trained on—a critical weakness for real-world deployment where new models and contexts constantly emerge.

The solution focuses on understanding how each model generates text, not just classifying outputs: Rather than trying to directly predict which model created text, ATLAS learns the underlying generation patterns of each source model and compares how well the text "fits" each pattern. This approach proves more robust to unexpected variations.

Responsible AI requires reliable detection across diverse scenarios: The paper introduces MAABench, a comprehensive testing framework that evaluates attribution systems under practical conditions like domain shifts and model version changes, establishing standards for what "reliable" attribution actually means in deployment.

Attribution technology is essential for combating AI misuse: By enabling detection of AI-generated text sources, this work supports efforts to prevent academic misconduct, unauthorized model copying, and cyberattacks—making it a foundational tool for responsible AI governance.