The rise of sophisticated AI text generators like ChatGPT has sparked a new arms race: building detectors that can reliably identify AI-written content. But as these language models (LLMs) get better at mimicking human writing, traditional detectors, often relying on statistical quirks, are becoming less effective. A new research project called DART (Detector using AMR of Rephrased Text) is taking a fresh approach. Instead of just looking at the text itself, DART focuses on subtle semantic shifts that happen when a text is rephrased. It leverages a technique called Abstract Meaning Representation (AMR) to analyze the underlying meaning of both the original text and its rephrased version, looking for inconsistencies that might betray an AI author. Think of it like asking someone to explain a concept in different words—a human might offer nuanced variations, while an AI could stumble and reveal its limitations. In experiments, DART outperformed several existing detectors, especially when tasked with identifying text from multiple different AI models. It even achieved impressive accuracy when trained on smaller datasets, making it a potentially more efficient solution. However, the research is still in its early stages. The effectiveness of DART relies heavily on the specific rephrasing tool used, currently GPT-4. Future research will need to explore whether DART can work with other rephrasers and how robust it is to variations in AI writing styles. Additionally, the accuracy of the AMR parsing itself can influence DART's results. The growing sophistication of AI text generation presents ongoing challenges for detection. DART’s semantic approach offers a promising direction but also highlights the ongoing cat-and-mouse game between AI generation and detection, as researchers strive to build tools that can keep pace with rapidly evolving language models.