Auditing LLM-Governed Social Robots with Culture-Specific Moral Gradients

Technical Summary: Auditing LLM-Governed Social Robots with Culture-Specific Moral Gradients

Problem Statement
Large Language Models (LLMs) are increasingly integrated into social robots to enable quasi-autonomous decision-making in resource-constrained environments, such as prioritizing who receives scarce assistance first. While moral prioritization norms vary significantly across cultures based on age, status, and group size, current LLM moral audits remain predominantly English-centric and fail to test embodied contexts. This creates a diagnostic gap: without pluralistic calibration, LLM-governed robots risk imposing uniform defaults that scale into unequal access and harm user well-being in diverse cultural settings. Existing evaluation frameworks often treat cultural pluralism as a secondary concern or rely on Western-centric benchmarks, leaving the specific governance risks of LLM-robot allocation policies unexamined.

Methodology
The authors introduce a gradient-based audit framework designed to evaluate LLM moral trade-off behavior against culture-specific preference gradients. The methodology proceeds in two phases:

1. Scenario Construction and Grounding:

   • Evidence Base: The study derives scenarios from nine systematic literature reviews covering over 8,000 studies in social robotics, identifying three high-stakes domains: Care (healthcare/home), Education, and Public-Facing Services.
   • Dilemma Adaptation: The framework adapts the "Moral Machine Experiment" (MME), transforming its fatal "whom to spare" dilemmas into non-fatal "whom to assist first" allocation scenarios. Three core axes are preserved: Many vs. Few (MF), Young vs. Old (YO), and Higher vs. Lower Status (HL).
   • Cultural Benchmarks: The audit benchmarks LLM decisions against MME_Scores, which are min-max normalized Average Marginal Component Effects (AMCEs) derived from 117 countries. These scores serve as a descriptive baseline for cross-country preference strength, not as normative ground truth.
   • Experimental Setup: The study audits four LLMs (GPT-4o, DeepSeek R1, Mistral Large, Gemini 2.0 Flash Lite) across four country-language pairs (USA-English, China-Chinese, Japan-Japanese, Mexico-Spanish).
   • Prompting Regimes: Four prompting strategies are tested: PLAIN (no context), Cultural Calibration (CC, deployment context only), Zero-Shot Chain-of-Thought (ZSCoT), and Few-Shot Chain-of-Thought (FSCoT, adding contrastive cultural exemplars).
   • Scale: The experiment generates 57,600 decisions (100 samples per condition across 144 prompt conditions).

2. Evaluation Metrics (Two-Layer Analysis):

   • Layer 1 (Gradient Concordance): Uses Kendall's $\tau$ and Spearman's $\rho$ to test if LLMs differentiate between countries in an order consistent with documented MME preference gradients.
   • Layer 2 (Governance Typology): Analyzes individual cells to characterize failure modes across three dimensions:
     • Differentiation: Does the model track gradients, erase them, or invert them?
     • Direction: Does the output favor the majority-preferred option (A-territory) or the opposite (B-territory)?
     • Deliberation: Is the output variable or deterministic (collapsing to a fixed rule)?
   • These dimensions combine into a seven-bin governance typology ranging from "Calibrated" (best) to "Non-Tracking Rigidity" (worst).

Key Results

• General Failure to Differentiate: LLMs largely fail to differentiate between cultural contexts in moral trade-offs. The overall gradient concordance across 864 country-pair comparisons is near zero ($\tau = +0.086$).
• Cultural Asymmetry: Quality calibration is nearly twice as strong for Western-language decisions (English/Spanish) compared to Chinese and Japanese. Non-Western conditions face disproportionate miscalibration risks.
• Model Variability:
  • Mistral Large emerges as a significant outlier, showing meaningful concordance ($\tau = +0.315$) and the lowest rates of gradient erasure and deterministic rigidity.
  • GPT-4o exhibits the highest rates of "Gradient Erasure" (52.1%), producing identical outputs regardless of cultural context, particularly in "Many vs. Few" scenarios.
  • DeepSeek R1 shows the most instability, with negative concordance and contradictory allocation postures across languages.
• Prompting Effects: Prompting alone is insufficient to reliably correct cross-cultural gaps.
  • FSCoT (Few-Shot Chain-of-Thought with contrastive exemplars) produces the most consistent gains in gradient tracking, particularly for non-Western languages.
  • ZSCoT (Zero-Shot Chain-of-Thought) reduces determinism but often worsens gradient tracking, suggesting that reasoning capacity without cultural anchoring can lead to incorrect ordering.
• Failure Modes:
  • Gradient Erasure: Common in "Many vs. Few" dilemmas, where models apply a rigid utilitarian default regardless of cultural nuance.
  • Gradient Inversion: Models differentiate but in the wrong order, or apply opposing moral rules based on domain (e.g., favoring lower status in Care but higher status in Services).
  • Determinism: Over 50% of cells produce fully deterministic outputs, indicating a collapse of moral uncertainty and context sensitivity.

Significance and Claims
The paper claims to contribute a multilingual, domain-grounded audit instrument that makes LLM cultural posture legible and measurable prior to deployment. Its primary significance lies in:

1. Shifting the Audit Focus: Moving from simple bias detection to evaluating gradient tracking—whether models modulate the strength of their preferences according to cultural context, rather than just the direction.
2. Typologizing Governance Risks: Providing a framework to distinguish between different failure modes (e.g., cultural blindness vs. miscalibrated sensitivity vs. rigid defaults), which informs targeted mitigation strategies.
3. Model-Level vs. Prompt-Level Levers: The results suggest that model-level factors (e.g., training data diversity, architecture) are more robust levers for pluralistic calibration than inference-time prompting alone. While contrastive exemplars help, they cannot fully compensate for models that lack inherent cross-cultural sensitivity.
4. Pre-Deployment Gate: The authors position this audit as a necessary gate for LLM-robot governance, arguing that without such pluralistic evaluation, the deployment of these systems risks entrenching unequal access and undermining procedural legitimacy in global settings.

The study maintains a modest stance, clarifying that it uses MME data as a descriptive baseline of population tendencies rather than a normative prescription for robot behavior. It does not claim to solve moral alignment but rather to make the risks of misalignment visible and typologized for downstream adjudication.