The problem
AI-generated text usually does not include reliable source attribution, making it difficult to determine which language model produced a document.
Approach
The project examined whether different language models leave measurable linguistic signatures in their generated text. We generated and transformed model outputs, extracted features, and trained a BERT-based classifier to distinguish between source models.
Relevance
Model attribution has applications in digital watermarking, copyright protection, provenance, and analysis of AI-generated content.
Outcome
The classification approach achieved more than 97% accuracy identifying the model that generated a given text sample.