Large language models memorize parts of their training data. Memorizing short snippets and facts is required to answer questions about the world and to be fluent in any language. But models have also been shown to reproduce long verbatim sequences of memorized text when prompted by a motivated adversary. In this work, we investigate an intermediate regime of memorization that we call non-adversarial reproduction, where we quantify the overlap between model responses and pretraining data when responding to natural and benign prompts. For a variety of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up to 15% of the text output by popular conversational language models overlaps with snippets from the Internet. In worst cases, we find generations where 100% of the content can be found exactly online. For the same tasks, we find that human-written text has far less overlap with Internet data. We further study whether prompting strategies can close this reproduction gap between models and humans. While appropriate prompting can reduce non-adversarial reproduction on average, we find that mitigating worst-case reproduction of training data requires stronger defenses -- even for benign interactions.

This study aims to enable more reliable automated post-disaster building damage classification using artificial intelligence (AI) and multi-view imagery. The current practices and research efforts in adopting AI for post-disaster damage assessment are generally (a) qualitative, lacking refined classification of building damage levels based on standard damage scales, and (b) trained based on aerial or satellite imagery with limited views, which, although indicative, are not completely descriptive of the damage scale. To enable more accurate and reliable automated quantification of damage levels, the present study proposes the use of more comprehensive visual data in the form of multiple ground and aerial views of the buildings. To have such a spatially-aware damage prediction model, a Multi-view Convolution Neural Network (MV-CNN) architecture is used that combines the information from different views of a damaged building. This spatial 3D context damage information will result in more accurate identification of damages and reliable quantification of damage levels. The proposed model is trained and validated on reconnaissance visual dataset containing expert-labeled, geotagged images of the inspected buildings following hurricane Harvey. The developed model demonstrates reasonably good accuracy in predicting the damage levels and can be used to support more informed and reliable AI-assisted disaster management practices.