Chinese AI Lab DeepSeek Releases Model That Falsely Claims to Be ChatGPT
DeepSeek V3 outperforms many rivals on popular benchmarks but frequently identifies itself as OpenAI's ChatGPT due to suspected training on GPT-4 outputs. Experts warn this data recycling degrades model quality and potentially violates OpenAI's terms of service.
DeepSeek, a well-funded Chinese AI lab, releases a highly efficient open AI model called DeepSeek V3 that excels at text-based tasks like coding and essay writing. Despite its strong performance on popular benchmarks, the model exhibits a strange identity crisis where it frequently identifies itself as ChatGPT instead of acknowledging its true origin.
Testing reveals that DeepSeek V3 insists it is a version of OpenAI's GPT-4 model released in 2023, even providing instructions for OpenAI's API when asked about DeepSeek's own tools. The model also repeats the exact same jokes as GPT-4, leading researchers to conclude that DeepSeek likely trains its system on public datasets containing raw ChatGPT outputs.
AI experts warn that training a model on the outputs of a rival system acts like taking a photocopy of a photocopy, degrading quality and increasing hallucinations. Furthermore, this practice potentially violates OpenAI's terms of service, which explicitly prohibit users from developing competing models using content generated by ChatGPT.