Neural networks (e.g., LLMs) make often imperfect predictions, introducing biases into analyses that rely on them. Common empirical economics scenarios fall outside the existing literature on debiasing “black-box AI”. Our paper (with @J_S_Carlson) on robust and efficient inference with unstructured data (e.g., text, images) provides a unifying theoretical framework, relates inference with text/images to familiar problems like causal inference, develops new insights into efficiency, and provides practical guidance on common empirical scenarios. arxiv.org/abs/2505.00282