A new study by AI detection company Copyleaks found that 74.2% of DeepSeek’s text bears a strong resemblance to OpenAI’s ChatGPT. That’s a pretty striking overlap, raising big questions about whether DeepSeek may have used ChatGPT-generated content in its training.
Copyleaks used advanced classifiers to analyze different AI models, and while Claude, Gemini, and Llama all had distinct writing styles, DeepSeek’s output looked a lot like OpenAI’s. According to Copyleaks’ Shai Nisan, it’s kind of like recognizing an author’s work just by their writing style. While this doesn’t outright prove that DeepSeek is a direct copy of ChatGPT, it does suggest something unusual in its development.
If DeepSeek did train on OpenAI’s outputs without permission, that could open up a whole can of worms—potentially violating intellectual property rights and raising concerns about AI transparency. It’s another reminder that the industry needs clearer rules and regulations around AI training.
There could also be major business consequences. DeepSeek has marketed itself as a more cost-effective alternative to expensive AI chips like Nvidia’s, but if it turns out they’ve been using OpenAI’s data without authorization, they could be looking at some serious legal and financial trouble.
Of course, AI models often share training datasets, but Copyleaks suggests this goes deeper—pointing to structural similarities between DeepSeek and ChatGPT. Nisan explained that each AI model develops its own writing style based on its architecture and fine-tuning process, making this level of resemblance unusual.
With AI development moving at breakneck speed, these findings could push for stronger transparency requirements and clearer intellectual property protections in the industry. One thing’s for sure: this won’t be the last time AI companies face scrutiny over how their models are trained.