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

TL;DR

We introduce a context-based, multi-construct evaluation framework to study how privacy concerns and prosocial motivations jointly influence data-sharing decisions in large language models. By applying multi-group structural equation modeling and a human-referenced Value–Action Alignment Rate, we quantify whether model-implied attitude–action relations align with human expectations. Results reveal stable yet highly model-specific privacy–prosocial–data-sharing profiles and generally limited, though not absent, value–action alignment, with some models showing strong adherence and others reversing expected directions or failing to estimate alignment. The findings highlight the limits of one-dimensional value–action gaps under competing motives and provide a general, relational tool for assessing and guiding alignment of LLMs in privacy-sensitive contexts. The framework can guide future work on training and governance strategies to enhance relational coherence in LLM decision-making under value conflict.

Abstract

Large language models (LLMs) are increasingly used to simulate decision-making tasks involving personal data sharing, where privacy concerns and prosocial motivations can push choices in opposite directions. Existing evaluations often measure privacy-related attitudes or sharing intentions in isolation, which makes it difficult to determine whether a model's expressed values jointly predict its downstream data-sharing actions as in real human behaviors. We introduce a context-based assessment protocol that sequentially administers standardized questionnaires for privacy attitudes, prosocialness, and acceptance of data sharing within a bounded, history-carrying session. To evaluate value-action alignments under competing attitudes, we use multi-group structural equation modeling (MGSEM) to identify relations from privacy concerns and prosocialness to data sharing. We propose Value-Action Alignment Rate (VAAR), a human-referenced directional agreement metric that aggregates path-level evidence for expected signs. Across multiple LLMs, we observe stable but model-specific Privacy-PSA-AoDS profiles, and substantial heterogeneity in value-action alignment.

Figures (6)

• Figure 1: Motivation. When privacy concern and prosocial motivation exert opposing pressures on data sharing, a single action cannot be mapped to a unique value reference. Gap-based value–action scores therefore become ambiguous, as the same choice may align with prosocial values while contradicting privacy concerns.
• Figure 2: Evaluation framework. We combine context-based administration of standardized Privacy (IUIPC), Prosocialness (PSA), and Acceptance of Data Sharing (AoDS) questionnaires with multi-group structural equation modeling. The resulting path estimates are compared against a human-referenced directional template to compute Value–Action Alignment Rate (VAAR) for each LLM.
• Figure 3: Heterogeneous but self-consistent model profiles. Each point represents a model’s mean Privacy and PSA score (Likert 1--7). Color encodes the mean AoDS level, and Point Area reflects the average within-scale standard deviation across repeated rounds, capturing the model’s characteristic dispersion under a fixed protocol.
• Figure 4: Relationship between within-scale dispersion (Avg. SD; Table \ref{['tab:descriptive-consistency']}) and human-alignment divergence (VAAR; Table \ref{['tab:kl_rate_main']}) across models. The dashed line is a faint descriptive trend fit on log(VAAR) (guide-to-the-eye only).
• Figure 5: Order robustness of VAAR under four questionnaire orderings. We report both overall VAAR shifts and path-level VAAR diagnostics to localize which value--action links drive deviations when AoDS is elicited first.