Deepseek-R1 Exposed: 74% of Output Mirrors OpenAI

Copyleaks found that 74.2% of text generated by Deepseek-R1 closely resembles OpenAI’s stylistic patterns, suggesting potential unauthorized reliance on OpenAI’s model during training. This raises concerns about intellectual property, transparency, and ethical AI practices. Copyleaks used a rigorous method involving three AI classifiers to analyze stylistic fingerprints, achieving high precision. This discovery allows for model-specific attribution, which is crucial for protecting intellectual property and ensuring ethical AI development. In contrast, Microsoft’s Phi-4 model showed no resemblance to known models, confirming its independent training.