harm could be considered a subcategory of impacts on biological systems so we have some harm data [and could consider few-shot transfer here, to predict harm well from little data]
the task seems similar to the translation of rare languages with minimal text. this problem has gotten mainstream attention with success.
