A recent study reveals that 74.2% of the text generated by DeepSeek closely resembles the writing style of OpenAI’s ChatGPT. This finding raises the possibility that DeepSeek may have been trained on outputs from ChatGPT. The research, conducted by the AI detection firm Copyleaks, will soon be published on Cornell’s arXiv.org repository. The implications of this study could significantly affect intellectual property rights, AI regulations, and the future of AI development.
Copyleaks utilized advanced screening technology and algorithm classifiers to analyze the stylistic characteristics of text produced by various language models, including OpenAI, Claude, Gemini, Llama, and DeepSeek. While most models displayed unique writing styles, a large portion of DeepSeek’s outputs were identified as similar to those generated by OpenAI.
Shai Nisan, head of data science at Copyleaks, compared the study to a handwriting expert identifying an author by comparing samples from different writers. He emphasized the strong stylistic similarity found between DeepSeek and OpenAI, which was not observed with other models.
This raises critical questions about how DeepSeek was trained and whether it used OpenAI’s outputs without permission. While the similarity does not definitively label DeepSeek as derivative, it suggests the need for further investigation into its training data and development process. If DeepSeek did use OpenAI-generated texts without authorization, it could violate OpenAI’s terms of service and intellectual property rights, highlighting the necessity for transparency in AI training data.
The study suggests that clear regulations and strong IP protections are essential for the future of AI development. As AI technology evolves, the need for defined ethical standards and legal frameworks becomes increasingly important. The findings may prompt regulators to require companies to disclose detailed information about their training datasets and model outputs, shaping the future landscape of AI.
