A Voting-Based System for Ethical Decision Making (1709.06692v2)

Abstract: We present a general approach to automating ethical decisions, drawing on machine learning and computational social choice. In a nutshell, we propose to learn a model of societal preferences, and, when faced with a specific ethical dilemma at runtime, efficiently aggregate those preferences to identify a desirable choice. We provide a concrete algorithm that instantiates our approach; some of its crucial steps are informed by a new theory of swap-dominance efficient voting rules. Finally, we implement and evaluate a system for ethical decision making in the autonomous vehicle domain, using preference data collected from 1.3 million people through the Moral Machine website.

• The paper proposes a four-step framework that aggregates over 18M pairwise comparisons to derive collective ethical decisions.
• It models individual preferences using a Thurstone-Mosteller process, achieving up to 92.4% accuracy with 100 comparisons per voter.
• The approach efficiently summarizes models with voting rules like Borda count and Copeland to support scalable real-time decision-making.

Overview of the Voting-Based System for Ethical Decision Making

The paper "A Voting-Based System for Ethical Decision Making" presents a methodology to automate ethical decisions utilizing machine learning and computational social choice. The authors propose a four-step approach: (I) collecting preference data from individuals, (II) learning models of individual preferences, (III) summarizing these models into a collective model, and (IV) using this summary model for real-time decisions. The implementation focuses particularly on the autonomous vehicle domain, leveraging 1.3 million responses from the Moral Machine website.

Methodological Framework

In addressing the long-standing challenge of formalizing ethical decision-making principles, the authors draw on computational social choice to aggregate societal preferences into a collective decision framework. The approach foregoes the need for predefined ethical axioms by approximating ethical decision-making through large-scale data aggregation.

1. Data Collection: The authors utilized an extensive dataset from the Moral Machine project, consisting of 18,254,285 pairwise comparisons by 1,303,778 voters. This scale of data collection is unprecedented in studies on ethical decision-making in AI.
2. Preference Learning: Each voter's decisions are modeled using a Thurstone-Mosteller (TM) process with linear parameterization relative to alternatives' features. This relational model ensures consistency in capturing varied personal ethical views as random utility models.
3. Model Summarization: The summary model, crafted to be a single TM process with averaged parameters, is designed to scale efficiently for rapid decision-making. This permits application in real-time scenarios where ethical decisions are critical, such as autonomous vehicle operations.
4. Preference Aggregation: The developed theoretical framework enables the application of any swap-dominance efficient voting rule from computational social choice. Here, the Borda count and Copeland correspondences are used, which efficiently select from finite alternatives based on maximum mode utilities.

Empirical Evaluation

The authors validate the constituent steps of their approach using synthetic data and the Moral Machine dataset. Key findings include:

• Step II Accuracy: Learning voter-specific models reveals high accuracy, achieving up to 92.4% with 100 pairwise comparisons per voter.
• Summarization Robustness: Step III shows that summarization results in minimal accuracy loss, and the summarization maintains high fidelity in collective decision-making even with extensive voter diversity.

This instantiation process demonstrates that ethical decisions can be automated effectively and at scale. The method, thus, not only aligns computational feasibility with theoretical robustness but also accommodates the complexities inherent in real-world ethical dilemmas.

Implications and Future Work

This research underscores the potential for AI systems to incorporate collective ethical reasoning without explicit ethical rule definitions, relying instead on observed societal preferences. Practically, this represents a prototype framework for decision-making embedded in systems like autonomous vehicles, potentially influencing areas like public policy where collective ethics must be operationalized.

Theoretically, this work extends the paradigm of computational social choice into moral dimensions, suggesting opportunities for integrating more complex moral principles over time. Future developments would benefit from exploring richer utility representations and the feasibility of handling more nuanced ethical considerations as AI applications become increasingly pervasive.

Conclusion: This paper's proposed framework and algorithmic execution manifest a pragmatic step toward balancing ethical automation with real-time constraints—a critical progression for autonomous AI systems. While challenges in ethical data accuracy and representational equity remain, the groundwork laid here promises significant advancements in AI's capacity to emulate nuanced human ethical decision-making.