LLM Robots Favor Western Morals 2x More in Assistance Audits

A new paper from Carmen Ng and Gjergji Kasneci (TUM), accepted at FAccT '26, introduces a gradient-based audit framework to test whether LLM-governed social robots make culturally pluralistic prioritization decisions. Drawing from nine cross-domain reviews (over 8,000 papers) and translating the Moral Machine Experiment's 'whom to spare' into 'whom to assist first' dilemmas, they tested four major LLMs across four country-language pairs (US-English, UK-English, China-Chinese, Japan-Japanese) under four prompting regimes. The 57,600 decisions revealed persistent, culturally asymmetric failures in tracking moral preference gradients.

The results are stark: quality calibration for Western-language decisions was nearly twice as strong as for Chinese and Japanese. High determinism in 'majority-first' trade-offs erased cross-cultural gradients, and models showed only partial sensitivity to age- and status-based norms, risking sidelining minorities. Prompting effects were uneven — only contrastive exemplars yielded consistent gains, while reasoning-only prompts sometimes worsened tracking. The authors conclude that model factors are a more robust lever than prompting alone, calling for multilingual, pluralistic audits as a pre-deployment gate for LLM-robot systems.