A research paper debates the effectiveness of gradient-based data attribution methods for large language models, concluding that these methods track format similarity more than task semantics. This has implications for data selection and analysis in AI training, and highlights the need for robustness and reliability in these methods.