Randomized Classifiers vs Human Decision-Makers: Trustworthy AI May Have to Act Randomly and Society Seems to Accept This

Randomization can be ethically necessary: The paper argues that truly fair and ethical AI decisions may require randomized (rather than deterministic) choices, because pure strategies in decision-making are vulnerable to exploitation and may not achieve mathematically optimal outcomes in game-theoretic terms.

Perfect information is impossible: AI systems making decisions with supervised learning inherently lack complete information about all relevant factors, making it impossible for any decision-maker—human or artificial—to make perfectly ethical choices based on complete knowledge.

Random classifiers perform comparably: The research demonstrates that randomized AI classifiers can perform at least as well as deterministic ones, suggesting that introducing randomness into AI decision-making need not compromise accuracy or performance.

Society accepts randomized decisions: Empirical findings indicate that people have a positive attitude toward randomized AI decision-making, suggesting public acceptance of this approach to achieving fairness—a crucial factor for trustworthy AI deployment.

Randomization prevents gaming the system: By using mixed strategies (randomized decisions), AI systems become harder to exploit or manipulate compared to predictable, deterministic decision rules, potentially making them more robust in adversarial settings.