From the Ground Truth Up: Doing AI Ethics from Practice to Principles

Explainability vs. Performance Trade-off: The paper argues that real-world AI practitioners face a fundamental tension between making AI systems explainable to users versus ensuring they perform accurately. Rather than assuming explainability is always necessary for trustworthiness, ground-level experience suggests practitioners must often choose between these competing goals based on specific contexts.

"Reliability" May Be More Useful Than "Trustworthiness": The authors challenge the widespread focus on AI "trustworthiness" in ethics discussions, proposing that "reliability"—a more concrete, measurable concept—better reflects what practitioners and users actually need from AI systems in practice.

Protection vs. Innovation Tension: Current AI ethics debates often frame the choice as either protecting users from harms or enabling innovation. The paper demonstrates that practitioners working on real problems experience these as interconnected rather than opposing goals, suggesting ethics frameworks need more nuance than this binary presents.

Theory-Practice Gap in AI Ethics: The paper's core finding is that AI ethics has become too focused on abstract principles developed by theorists, disconnected from the actual dilemmas and constraints faced by designers building AI systems. Grounding ethics in practitioners' lived experiences could make ethical frameworks more relevant and actionable.

Methodology Matters: By studying AI designers' real-world decision-making rather than applying pre-existing principles, the authors demonstrate that ethical insights generated from practice can reshape and improve theoretical debates in AI ethics.