Large language models (LLMs) are impressive feats of artificial intelligence, able to generate human-like text, translate languages, and answer questions. But what happens when the knowledge they've learned becomes outdated or, worse, is intentionally manipulated? A new research paper explores the critical task of detecting when an LLM's output is based on edited knowledge versus its original training data. This is crucial for maintaining trust and transparency in these powerful models. Imagine an LLM confidently stating a false fact, like "The Space Needle is in Berlin." How can we tell if this is a simple error or a deliberate edit? The researchers propose a novel task called "Detecting Edited Knowledge in LLMs" (DEED). They examine various methods for identifying these edits, focusing on two key approaches: analyzing the model's internal representations (hidden states) and examining the probabilities it assigns to different words (probability distributions). Their findings reveal that simpler, more efficient editing techniques are surprisingly easier to detect. These techniques often involve directly modifying the model's parameters associated with specific facts, leaving a clearer trace. More complex methods, while potentially more subtle, can still be identified by examining the model's output probabilities. The research also highlights the challenge of distinguishing between edited facts and unedited but related facts. For example, differentiating between the edited fact "The Eiffel Tower is in Berlin" and the unedited fact "Marlene Dietrich was born in Berlin" proves tricky. This underscores the need for more sophisticated detection methods. The ability to detect edited knowledge is a vital step towards ensuring responsible use of LLMs. As these models become increasingly integrated into our lives, it's essential to have tools that can identify manipulation and maintain the integrity of information.