Value-Action Alignment in Large Language Models under Privacy-Prosocial Conflict

A new study from researcher Xiyang Hu introduces a protocol to test whether large language models act consistently with their stated privacy and prosocial values when those values conflict, revealing substantial variation across models [1]. The work, posted to arXiv on 7 January 2026 and revised on 28 June 2026, addresses a gap in how AI systems are evaluated when making decisions about personal data [1]. Existing assessments often measure privacy attitudes or sharing intentions separately, making it difficult to determine whether a model’s expressed values jointly predict its downstream actions as they do in human behavior [2]. Hu’s protocol administers standardized questionnaires for privacy attitudes, prosocialness, and acceptance of data sharing within a single, history-carrying session [2]. Multi-group structural equation modeling, or MGSEM, is then used to map the relationships from privacy concerns and prosocialness to data-sharing decisions [1]. The paper also proposes a new metric called the Value-Action Alignment Rate, or VAAR, which aggregates path-level evidence to produce a human-referenced directional agreement score [2]. Across the models tested, the researchers observed stable but model-specific profiles and what they describe as “substantial heterogeneity in value-action alignment” [1]. The study arrives as large language models are increasingly deployed to simulate human decision-making in sensitive domains, including personal data sharing, where privacy and prosocial motivations can push choices in opposite directions [2]. The findings suggest that a model’s stated attitudes do not uniformly translate into consistent data-sharing behaviors, a result that carries implications for developers and regulators weighing the use of AI in contexts that demand predictable, value-driven decisions. The submission, totaling 501 KB, lists Xiyang Hu as the corresponding author and is hosted on arXiv under the Computation and Language category [1]. The paper’s code and data are associated with the Hugging Face platform, according to the article’s supplementary links [1].