Does moral meaning survive translation between languages?

Moral values research relies almost entirely on English-annotated datasets, leaving most languages under-resourced. This study tested whether LLM translation could bridge the gap by translating ~50k Polish social media posts and checking if moral meaning survives the process. Using four validation methods—embedding similarity, alignment analysis, LLM judges, and classifier parity tests—they found translated texts retain moral cues faithfully enough for machine learning (0.86 cosine similarity, <0.02 AUC gaps), despite stumbles with slang and cultural references. The implication: machine translation offers a practical, low-cost way to extend moral AI systems to non-English languages.