Language models judge war differently when tested for alignment

A comprehensive factorial experiment on 20 large language models found that when the prompting system was tested in a “alignment with human values” context, its war decision-making behavior was severely misjudged. The study covered 32 scenarios, with 10 rounds of repetition (a total of 12,800 judgments). The results showed that adding specific prompts had a dual effect: first, a horizontal effect, where the average willingness to go to war decreased by 13.43 points on a 0-100 scale; second, a structural effect, as the key factors driving judgment shifted from “success probability” to “civilian casualties” at baseline. Standardized estimates indicated that this change in order was mainly due to the model’s weakened consideration of strategic factors such as success probability and domestic support. Therefore, the existing evaluation framework not only changed the horizontal values of answers but also fundamentally altered the decision-making rules it revealed.