Exploring Multilingual Concepts of Human Value in Large Language Models: Is Value Alignment Consistent, Transferable and Controllable across Languages?

Value concepts exist in LLMs across multiple languages, but inconsistently: The research confirms that human values (studied across 7 types of values and 16 languages) are represented in large language models beyond just English, but these representations are not uniform across languages, raising concerns about whether AI systems maintain consistent ethical principles regardless of the language they're operating in.

Language resource inequality creates alignment problems: The study found that high-resource languages (like English) have more stable and accurate value representations compared to low-resource languages, leading to "cross-lingual inconsistency" and "distorted linguistic relationships." This means AI systems may not reliably uphold the same values when communicating in less commonly-used languages.

Value alignment can be controlled through dominant languages: The researchers demonstrated that controlling value alignment in a high-resource language (like English) can transfer to influence how values are represented in other languages, suggesting a potential practical approach for ensuring consistent AI behavior across multilingual systems.

Current multilingual AI training data composition may be problematic: The paper's findings suggest that how companies combine different languages in training data significantly impacts whether AI systems maintain consistent values across languages, indicating that current data composition practices may inadvertently create alignment vulnerabilities in non-English languages.