OpenAI has alleged that the new R1 model developed by DeepSeek may have been created using the company’s artificial intelligence (AI) models. Reports suggest that OpenAI has gathered evidence indicating that some users utilized its AI model outputs for DeepSeek, which released the open-source DeepSeek-R1 model last week via platforms like GitHub and Hugging Face. Early assessments show that this reasoning-centric model outperformed several of OpenAI’s own o1 AI models in various benchmarks.
OpenAI Accuses Competitor of Misconduct
A Financial Times report details OpenAI’s allegations that its proprietary AI models were used in training DeepSeek’s models. According to the company, it has identified evidence from multiple accounts using the OpenAI application programming interface (API). Following an investigation conducted alongside its cloud partner Microsoft, the companies moved to block access to the API for those involved.
In a statement provided to the Financial Times, OpenAI asserted, “We know [China]-based companies — and others — are constantly trying to distill the models of leading US AI companies.” The organization also mentioned its ongoing collaboration with the US government to safeguard its advanced models from rival companies and potential adversaries.
AI model distillation refers to a method of transferring knowledge from a larger model to a smaller, more efficient counterpart. This technique aims to make the smaller model as capable or superior to the larger one, while also lowering computational requirements. For context, OpenAI’s GPT-4 is equipped with approximately 1.8 trillion parameters, whereas DeepSeek-R1 has 1.5 billion parameters, aligning with distillation practices.
This transfer of knowledge typically involves using datasets from a larger model to train a more efficient version in-house. For instance, Meta employed its Llama 3 AI model to develop several coding-oriented Llama models. However, if a competitor lacks access to the proprietary datasets of an established model, achieving this through legitimate means becomes challenging. If OpenAI’s claims are validated, DeepSeek might have resorted to manipulating prompt injections via its APIs to generate substantial outputs, which could then be converted into code and utilized by a base model.
Currently, OpenAI has not released an official statement regarding these allegations. Interestingly, OpenAI’s CEO, Sam Altman, recently lauded DeepSeek for developing a sophisticated AI model that enhances competition within the AI sector.
