Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DailyDilemmas: Revealing Value Preferences of LLMs with Quandaries of Daily Life

Yu Ying Chiu, Liwei Jiang, Yejin Choi

TL;DR

DailyDilemmas provides 1,360 synthetic, non-clear-cut everyday moral dilemmas pairing two actions with stakeholders and associated values, enabling a cross-disciplinary assessment of LLM value prioritization. By mapping 301 values into five frameworks (WVSCulturalMap, Moral Foundations, Maslow, Aristotle’s Virtues, Plutchik’s Emotions) and analyzing action choices across GPT-4-turbo, Claude-3-haiku, and other models, the paper reveals consistent self-expression and care tendencies while exposing model-specific biases. It also evaluates alignment with OpenAI ModelSpec and Anthropic CAI principles and demonstrates limited steerability of value preferences via system prompts, underscoring practical challenges for end users in controlling model ethics in closed APIs. The work advances understanding of how real-world value trade-offs arise in AI decision-making and provides a dataset and analytic framework to guide safer, more aligned AI deployments in daily life.

Abstract

As users increasingly seek guidance from LLMs for decision-making in daily life, many of these decisions are not clear-cut and depend significantly on the personal values and ethical standards of people. We present DailyDilemmas, a dataset of 1,360 moral dilemmas encountered in everyday life. Each dilemma presents two possible actions, along with affected parties and relevant human values for each action. Based on these dilemmas, we gather a repository of human values covering diverse everyday topics, such as interpersonal relationships, workplace, and environmental issues. With DailyDilemmas, we evaluate LLMs on these dilemmas to determine what action they will choose and the values represented by these action choices. Then, we analyze values through the lens of five theoretical frameworks inspired by sociology, psychology, and philosophy, including the World Values Survey, Moral Foundations Theory, Maslow's Hierarchy of Needs, Aristotle's Virtues, and Plutchik's Wheel of Emotions. For instance, we find LLMs are most aligned with self-expression over survival in World Values Survey and care over loyalty in Moral Foundations Theory. Interestingly, we find substantial preference differences in models for some core values. For example, for truthfulness, Mixtral-8x7B neglects it by 9.7% while GPT-4-turbo selects it by 9.4%. We also study the recent guidance released by OpenAI (ModelSpec), and Anthropic (Constitutional AI) to understand how their designated principles reflect their models' actual value prioritization when facing nuanced moral reasoning in daily-life settings. Finally, we find that end users cannot effectively steer such prioritization using system prompts.

DailyDilemmas: Revealing Value Preferences of LLMs with Quandaries of Daily Life

TL;DR

DailyDilemmas provides 1,360 synthetic, non-clear-cut everyday moral dilemmas pairing two actions with stakeholders and associated values, enabling a cross-disciplinary assessment of LLM value prioritization. By mapping 301 values into five frameworks (WVSCulturalMap, Moral Foundations, Maslow, Aristotle’s Virtues, Plutchik’s Emotions) and analyzing action choices across GPT-4-turbo, Claude-3-haiku, and other models, the paper reveals consistent self-expression and care tendencies while exposing model-specific biases. It also evaluates alignment with OpenAI ModelSpec and Anthropic CAI principles and demonstrates limited steerability of value preferences via system prompts, underscoring practical challenges for end users in controlling model ethics in closed APIs. The work advances understanding of how real-world value trade-offs arise in AI decision-making and provides a dataset and analytic framework to guide safer, more aligned AI deployments in daily life.

Abstract

As users increasingly seek guidance from LLMs for decision-making in daily life, many of these decisions are not clear-cut and depend significantly on the personal values and ethical standards of people. We present DailyDilemmas, a dataset of 1,360 moral dilemmas encountered in everyday life. Each dilemma presents two possible actions, along with affected parties and relevant human values for each action. Based on these dilemmas, we gather a repository of human values covering diverse everyday topics, such as interpersonal relationships, workplace, and environmental issues. With DailyDilemmas, we evaluate LLMs on these dilemmas to determine what action they will choose and the values represented by these action choices. Then, we analyze values through the lens of five theoretical frameworks inspired by sociology, psychology, and philosophy, including the World Values Survey, Moral Foundations Theory, Maslow's Hierarchy of Needs, Aristotle's Virtues, and Plutchik's Wheel of Emotions. For instance, we find LLMs are most aligned with self-expression over survival in World Values Survey and care over loyalty in Moral Foundations Theory. Interestingly, we find substantial preference differences in models for some core values. For example, for truthfulness, Mixtral-8x7B neglects it by 9.7% while GPT-4-turbo selects it by 9.4%. We also study the recent guidance released by OpenAI (ModelSpec), and Anthropic (Constitutional AI) to understand how their designated principles reflect their models' actual value prioritization when facing nuanced moral reasoning in daily-life settings. Finally, we find that end users cannot effectively steer such prioritization using system prompts.
Paper Structure (52 sections, 6 figures, 15 tables)

This paper contains 52 sections, 6 figures, 15 tables.

Table of Contents

  1. Introduction
  2. DailyDilemmas: A Dataset of Everyday Dilemmas
  3. The Importance of Value-Based Theoretical Frameworks
  4. Dataset Construction
  5. An Analysis of Synthetically Generated Dilemma Vignettes and Human Values in DailyDilemmas
  6. Dataset Statistics and Topic modeling
  7. Mapping Human Values to Theoretical Moral Frameworks
  8. Verifying the Validity of DailyDilemmas with Human-Written Dilemmas
  9. Unveiling LLMs' Value Preferences Through Action Choices in Everyday Dilemmas
  10. Results
  11. Examining LLMs' Adherence to Design Principles and the Steerability of Value Preferences
  12. Case Study: Anthropic Constitutional AI
  13. Case Study: OpenAI ModelSpec
  14. Testing the Inference-Time Steerability of LLMs' Value Preferences
  15. Related Work
  16. ...and 37 more sections

Figures (6)

  • Figure 1: The Data schema and the synthetic data generation pipeline of DailyDilemmas. Each dilemma vignette presents two possible actions, along with their associated stakeholders and the human values implicated in each choice.
  • Figure 2: Representative examples and UMAP visualization of topic modeling for dilemmas in DailyDilemmas, spanning a diverse range of everyday topics.
  • Figure 3: Value distribution in DailyDilemmas based on five theories (Culture: World Values Survey, Foundation: Moral Foundations Theory, Needs: Maslow's Hierarchy of Needs, Virtues: Aristole's Virtue, Emotions: Plutchik's Wheel of Emotions) that also disclose GPT-4's bias during generation.
  • Figure 4: Normalized distribution of four representative models on their values preferences based on five theories with reduced dimensions. Since the model (GPT-4) used shows the unbalanced distribution of generated values in our five theories in Fig. \ref{['fig:percent_data']}, we decided to use the normalized percentages to adjust different dimensions in different theories into the same scale for meaningful visualization. Therefore, the normalized percentage is calculated by dividing the raw counts difference of value preferences. To interpret this graph, we should view each of the dimensions (e.g., Tradition on World Values Survey) to compare models.
  • Figure 5: Steerability of GPT-4 by system prompt. indicates effective modulation, where the upper triangles (after modulation towards supportive values) will have a higher score than the rounds (before modulation), and vice versa.
  • ...and 1 more figures