Large language models (LLMs) are like mysterious black boxes—they generate impressive text but leave us wondering how they came up with it. What parts of their training data did they actually use? This is not just an academic question; it's vital for transparency, accountability, and even legal compliance (think GDPR). Imagine needing to prove an AI didn't plagiarize or misuse sensitive data. Now, researchers have developed a powerful new tool called TRACE that shines a light into these black boxes. TRACE uses a clever technique called "contrastive learning." It works by creating a map of the training data, where similar pieces of information are clustered together. When the LLM generates a response, TRACE pinpoints the closest clusters on the map, effectively revealing the source of the information. Think of it like a detective tracing a suspect's steps. The real breakthrough with TRACE is that it doesn't need access to the LLM's inner workings. It's "model-agnostic," meaning it works with any LLM, without needing to peek inside. This is a game-changer for accountability and research. In tests, TRACE demonstrated impressive accuracy in identifying the source of information used by different LLMs, even when faced with adversarial attempts to mislead it (though paraphrasing proved trickier to defend against). While challenges remain, especially with highly similar datasets, TRACE opens exciting possibilities for making AI more transparent and trustworthy.