Large language models (LLMs) like ChatGPT have a knack for generating human-like text, but there’s a hidden problem: potential plagiarism. These models learn from massive datasets scraped from the web, raising legal and ethical concerns about copyright infringement. A recent research paper proposes a novel way to detect whether an LLM has “plagiarized” its training data. The researchers built a system that analyzes how LLMs construct knowledge. They convert both a source document and an LLM's continuation of that document into “knowledge graphs,” essentially maps of how ideas are connected. By comparing these graphs, they can identify similarities in both the content and structure of the text, revealing whether the LLM is simply regurgitating information. This method cleverly sidesteps the “black box” nature of many LLMs, where access to the training data is restricted. Instead, it focuses on the output, making it possible to hold LLMs accountable for responsible sourcing. This approach compares not just individual words or phrases but also the relationships between ideas, going beyond traditional plagiarism detection. The researchers use “cosine similarity” to measure how closely the LLM's output matches the source material in terms of content, and “normalized graph edit distance” to assess how similar the structures of the knowledge graphs are. They believe this two-pronged approach provides a more comprehensive picture of potential plagiarism. This research highlights the growing need for transparency and accountability in the development and use of LLMs. As these models become increasingly sophisticated, so too must our methods for ensuring they are used responsibly and ethically. The future of AI depends on it.