Earlier this week, DeepSeek, a Chinese AI lab, launched DeepSeek V3, an AI model surpassing many others in efficiency for tasks like coding and writing. However, the model mistakenly believes it is ChatGPT. Tests revealed that DeepSeek V3 identifies as ChatGPT, claiming to be a version of OpenAI's GPT-4 model from 2023. The model even mimics GPT-4's responses, including telling similar jokes. The confusion arises because AI models like ChatGPT and DeepSeek V3 are statistical systems trained on vast datasets to predict patterns. DeepSeek has not disclosed its training data sources, but there's an abundance of public datasets with GPT-4-generated text. It seems DeepSeek V3 might have memorized some of these outputs. Training models on outputs from rival systems can be detrimental, causing inaccuracies and hallucinations.

It may also breach usage terms, as OpenAI prohibits using its product outputs to develop competing models. OpenAI's CEO, Sam Altman, subtly criticized this practice, highlighting the ease of copying versus innovating. This misidentification issue isn't unique to DeepSeek V3; other models like Google’s Gemini also misidentify. The growing prevalence of AI-generated content on the web, which serves as training data, exacerbates this problem. It's possible DeepSeek used ChatGPT-generated text for training, similar to past accusations against Google. Heidy Khlaaf from the AI Now Institute noted the appeal of using existing models' knowledge to save costs, despite the risks. This practice may lead to models like DeepSeek V3 unintentionally replicating biases and flaws from GPT-4, raising concerns about accuracy and self-identification.