Large language models (LLMs) like ChatGPT are impressive, but they can also generate false information or even plagiarize existing content. Wouldn’t it be great if AI could tell us *where* it got its information, so we could check its sources? New research explores how to identify the source of text generated by LLMs. Imagine you ask an LLM about Christmas. It might generate text about family, friends, food, and the birth of Jesus Christ. Researchers are working on ways to identify the documents most likely to have influenced that generation – like specific Wikipedia articles related to Christmas traditions. This involves tracing word origins back to possible source documents within the massive datasets LLMs are trained on. The research introduces the idea of “token-level source identification,” which means figuring out where the model got the information for each individual word or token. It works by training a separate “source identifier” model that analyzes the LLM's internal representations and links them to the original training documents. This research tested different identifier models, language models, and dataset types. They found that it’s indeed possible to identify sources with reasonable accuracy, and larger language models generally perform better at this task. The best-performing method used "bigram representations," looking at pairs of successive words to increase accuracy. This research is important because it addresses concerns about the reliability and trustworthiness of information generated by LLMs. Imagine being able to fact-check an AI’s claims by reviewing its sources! While there are challenges in scaling this to extremely large datasets and preventing false positives, this work is a promising step toward making LLMs more transparent and accountable.