Privacy-preserving Crowd-guided AI Decision-making in Ethical Dilemmas

Moral preferences are sensitive personal data: The paper identifies that people's ethical views—such as how autonomous vehicles should handle life-or-death decisions—constitute private information that could be exploited if disclosed to malicious parties, highlighting why privacy protection matters in AI ethics research.

Crowd input is valuable but risky for ethical AI: While gathering public preferences through voting is a useful way to build AI systems that reflect societal values, the centralized collection of these moral judgments creates privacy vulnerabilities that current approaches don't adequately address.

Privacy protection comes in different strengths: The research proposes four different privacy-protection approaches (VLCP, RLCP, VLDP, and RLDP) that vary in scope—protecting individual voters versus entire records, and using centralized versus distributed methods—allowing organizations to choose appropriate privacy levels for their context.

Privacy and accuracy can coexist: A key finding is that the proposed methods can protect voter privacy while still maintaining accuracy in the AI system's ability to learn and apply moral preferences, suggesting that privacy protection doesn't require sacrificing the quality of ethical decision-making.

Differential privacy offers a quantifiable solution: By adopting differential privacy as a mathematical framework, the paper provides a concrete, measurable way to guarantee privacy protection in crowd-guided AI systems, moving beyond vague privacy promises to verifiable privacy guarantees.