Growth First, Care Second? Tracing the Landscape of LLM Value Preferences in Everyday Dilemmas

People increasingly seek advice online from both human peers and large language model (LLM)-based chatbots. Such advice rarely involves identifying a single correct answer; instead, it typically requires navigating trade-offs among competing values. We aim to characterize how LLMs navigate value trade-offs across different advice-seeking contexts. First, we examine the value trade-off structure underlying advice seeking using a curated dataset from four advice-oriented subreddits. Using a bottom-up approach, we inductively construct a hierarchical value framework by aggregating fine-grained values extracted from individual advice options into higher-level value categories. We construct value co-occurrence networks to characterize how values co-occur within dilemmas and find substantial heterogeneity in value trade-off structures across advice-seeking contexts: a women-focused subreddit exhibits the highest network density, indicating more complex value conflicts; women’s, men’s, and friendship-related subreddits exhibit highly correlated value-conflict patterns centered on security-related tensions (security vs. respect/connection/commitment); by contrast, career advice forms a distinct structure where security frequently clashes with self-actualization and growth. We then evaluate LLM value preferences against these dilemmas and find that, across models and contexts, LLMs consistently prioritize values related to Exploration & Growth over Benevolence & Connection. This systemically skewed value orientation highlights a potential risk of value homogenization in AI-mediated advice, raising concerns about how such systems may shape decision-making and normative outcomes at scale.

💡 Research Summary

The paper investigates how large language models (LLMs) navigate value trade‑offs when giving advice in everyday dilemmas. Using a curated collection of 5,728 real‑world dilemmas from four Reddit advice subreddits—AskMenAdvice, AskWomenAdvice, CareerAdvice, and FriendshipAdvice—the authors first map the underlying value structure of human advice. They employ GPT‑4o to extract a concise “core value” for each of the two options in every dilemma, validating the extraction on a 400‑post sample with high inter‑annotator agreement (Cohen’s κ = 0.833, 92% accuracy). This yields 2,288 distinct fine‑grained values.

A bottom‑up hierarchical value framework is then built. All fine‑grained values are embedded with all‑mpnet‑base‑v2 and clustered via k‑means into 175 first‑level clusters, which are further grouped into 33 second‑level clusters and finally into four top‑level dimensions: Exploration & Growth, Security & Stability, Achievement & Impact, and Benevolence & Connection. Cluster names are generated by GPT‑4o and refined by human reviewers.

The authors construct value co‑occurrence networks for each subreddit, revealing substantial heterogeneity. The women‑focused subreddit shows the highest network density, indicating more complex value conflicts. Men’s, women’s, and friendship subreddits share a pattern centered on security versus respect/connection/commitment tensions, while career advice displays a distinct pattern where security frequently clashes with self‑actualization and growth.

To assess LLM value preferences, the same dilemmas are presented to three state‑of‑the‑art models—GPT‑4o, DeepSeek‑V3.2‑Exp, and Gemini‑2.5‑Flash—under zero‑temperature, order‑randomized conditions. Model choices are robust to option ordering (92.5% unchanged). Each chosen option is mapped to the top‑level value framework, allowing quantification of model preferences. Across all models and subreddits, the consistent finding is a strong bias toward Exploration & Growth over Benevolence & Connection. In career‑related dilemmas, the tension between Security & Stability and Exploration & Growth is especially pronounced.

The study makes four key contributions: (1) a scalable pipeline that leverages authentic online dilemmas to evaluate LLM value behavior, overcoming the limited ecological validity of synthetic benchmarks; (2) a data‑driven, four‑level hierarchical value taxonomy that offers an alternative to pre‑defined value taxonomies; (3) empirical evidence that current LLMs systematically prioritize growth‑oriented values, suggesting a potential for value homogenization in AI‑mediated advice; and (4) a publicly released dataset containing the original dilemmas, extracted values, hierarchical annotations, and model choices.

Limitations include reliance on GPT‑4o for value labeling (which may inject its own biases), focus on English‑language Reddit communities (limiting cultural diversity), and the binary‑choice paradigm that may not capture the full nuance of human value deliberation. Future work should expand to multilingual, multicultural sources, incorporate multi‑option and explanatory responses, and explore mitigation strategies to balance LLM value orientations. The findings raise important ethical considerations: if LLMs consistently amplify growth‑centric values, they could influence users toward individualistic decision‑making at the expense of relational and safety concerns, potentially reshaping normative outcomes at scale.