"The aspect of AI which sabotages it in social settings, its tendency to generate collateral damage through out-of-distribution errors & so-called hallucinations, ceases to be a problem in conflicts where maximum indiscriminate damage is the actual point" https://zenodo.org/records/18908530