AI-powered social robots have a triage problem: they favor Western moral norms when deciding who gets help first.
Researchers tested four large language models across four country-language pairs using more than 57,000 simulated decisions, adapted from the Moral Machine Experiment's ethical dilemmas. Instead of asking which pedestrian a self-driving car should spare, they asked which person a social robot should assist first — across care, education, and service scenarios. The benchmark compared model outputs against documented cultural preference data from each country. They found that calibration quality for Western-language decisions was nearly twice as strong as for Chinese and Japanese contexts, and that models frequently defaulted to majority-first trade-offs in ways that erased cross-cultural distinctions entirely.
This matters because social robots are not hypothetical. They are being deployed in elder care, schools, and service settings where prioritization decisions carry real consequences for real people. An AI that systematically sidelines minority or non-Western preferences is not a neutral tool — it is a policy embedded in code.
Attempts to fix the bias through prompting mostly failed; only contrastive examples produced consistent improvement, while chain-of-thought reasoning prompts sometimes made things worse. The researchers argue that model-level fixes — not prompt engineering — are the more reliable path forward, and that multilingual, pluralistic audits should become a hard pre-deployment requirement. Given how slowly that kind of structural change tends to move through the AI industry, the robots will probably be in the waiting room long before the audits are.