A Statistical Case Against Empirical Human-AI Alignment

2025-02-20

AI Summary

This paper identifies an inherent statistical bias in empirical human-AI alignment—i.e., inferring AI objectives solely from observed human behavior—which systematically distorts alignment outcomes. Grounded in statistical principles, the authors formally characterize the root cause: non-random sampling of behavioral observations and absence of counterfactuals, leading to invalid causal inference. Through integrated statistical modeling and empirical case studies using human-centered decoding in language models, they demonstrate the unreliability of purely empirical alignment. In response, they propose a dual-track paradigm: (1) *normative alignment*, which encodes value-based priors as hard constraints on AI objective specification, and (2) *posterior empirical calibration*, which selectively incorporates behavioral data only within this normative framework. This approach reconciles theoretical rigor with practical deployability, offering a more robust methodological foundation and actionable pathway for reliable AI alignment.

Technology Category

• Humans and AI: Learning Human Values and Preferences
• Philosophy and Ethics of AI: Safety, Robustness & Trustworthiness
• Machine Learning: Calibration & Uncertainty Quantification

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Economics, Online Markets and Human Computation: Fairness and ethical considerations in crowd work and in human-in-the-loop AI systems
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Empirical human-AI alignment aims to make AI systems act in line with observed human behavior. While noble in its goals, we argue that empirical alignment can inadvertently introduce statistical biases that warrant caution. This position paper thus advocates against naive empirical alignment, offering prescriptive alignment and a posteriori empirical alignment as alternatives. We substantiate our principled argument by tangible examples like human-centric decoding of language models.

Problem

• Empirical human-AI alignment introduces biases
• Advocates against naive empirical alignment
• Proposes prescriptive and a posteriori alternatives

Innovation

• Advocates prescriptive AI alignment
• Proposes posteriori empirical alignment
• Critiques naive empirical alignment