When Personas Become Harmful: Aligning Large Language Models with Global Citizen Values

The researchers constructed a dataset of 11k pairs of controversial topics and value personas, with personas mapped to the value dimensions of the Schwartz Value Survey, iteratively refining topics and personas through an agentic pipeline, eliciting harmful responses from well-aligned commercial large language models with a success rate exceeding 47%. Such persona-conditioned requests usually contain no obvious malice, and existing alignment methods targeting adversarial requests may miss the harms arising from persona use. The paper uses UNESCO global citizenship principles as a minimum consensus for distinguishing legitimate value expression from harmful behavior, and proposes a controllable alignment model that can balance persona adherence and safety at inference time. On most samples, this model is safer and more persona-adherent than baselines.