📄 中文摘要
AI 在回答历史事件时，可能会给出自信且详细的回答，但内容却与主流历史共识相悖。这种现象反映了模型训练数据中的偏见，尤其是使用了过时的术语和强调了现代学术研究已不再关注的人物和观点。通过分析模型输出，可以识别和重建这些偏见，类似于古生物学家通过化石重建过去的生态系统。这种方法不仅揭示了 AI 模型的局限性，也为改进模型的训练数据提供了新的视角，推动了更准确和公正的 AI 应用发展。
提示古生物学：通过模型输出中的化石猎捕重建训练数据偏见
📄 English Summary
Prompt Paleontology: Reconstructing Training Data Biases by Fossil-Hunting in Model Outputs
When an AI is asked about a historical event, it may provide a confident and detailed answer that diverges from mainstream historical consensus. This phenomenon highlights biases in the model's training data, particularly the use of outdated terminology and an emphasis on figures and perspectives that modern scholarship has moved away from. By analyzing model outputs, it is possible to identify and reconstruct these biases, akin to how paleontologists reconstruct past ecosystems from fossils. This approach not only reveals the limitations of AI models but also offers new insights for improving training data, promoting the development of more accurate and equitable AI applications.