How Do Fairness Definitions Fare? Examining Public Attitudes Towards Algorithmic Definitions of Fairness

Public preference for calibrated fairness: The study found that ordinary people tend to prefer "calibrated fairness" over other algorithmic fairness definitions when evaluating loan decisions, suggesting that technical experts and the public may align on at least one fairness approach.

Sensitive information influences fairness judgments: When people were shown the race of loan applicants (sensitive demographic information), their perceptions of what constitutes fair algorithmic decisions changed, indicating that context and transparency about protected characteristics matter in how people evaluate AI systems.

Gap between technical definitions and public values: The research highlights a disconnect between the multiple fairness definitions proposed by computer scientists and what ordinary people actually consider fair, suggesting that AI developers should consult public values when choosing which fairness approach to implement.

Support for affirmative action principles: The study's findings provided empirical support for affirmative action principles in algorithmic decision-making, indicating that public attitudes may favor definitions that account for historical inequities rather than purely "neutral" approaches.

Need for public input in AI governance: By demonstrating that non-experts have coherent and measurable fairness preferences, this work suggests that responsible AI development should include public consultation rather than leaving fairness definitions solely to technical experts.