Large language models (LLMs) like those powering ChatGPT are amazing at generating creative text formats, from poems to code. But new research reveals a hidden problem: even after fine-tuning, these models sometimes unintentionally plagiarize from their training data, raising concerns about originality. Researchers explored this “self-plagiarism” phenomenon, where LLMs generate text strikingly similar to training set examples—sometimes verbatim, sometimes paraphrased, or even borrowing core ideas without proper attribution. This tendency to mimic data poses challenges for academic integrity and creative content generation. To combat this, researchers developed a clever technique called "self-plagiarism contrastive decoding." It trains the model to identify and penalize its own plagiaristic tendencies. Initial experiments showed promising results. Using this method, LLMs generated significantly more original academic text and storytelling compared to traditional methods. While not perfect, this research points to exciting new approaches to ensure AI-generated content is both creative and original, paving the way for models that not only write fluently, but also think for themselves.