Best Practices to Detect and Avoid Harmful Biases in Artificial Intelligence Systems

This document synthesizes the main findings derived from both on-desk research and a workshop with experts from various APEC economies. Significant findings from the study include:

1. APEC economies have developed different strategies to detect, avoid, and mitigate harmful biases in AI systems, with the most successful ones establishing robust institutional frameworks and actively implementing measures to address biases.
2. In terms of institutional strategies, common measures to reinforce trust in AI ecosystems include the creation of domestic policies, the adoption of ethical frameworks, and the establishment of guidelines and pilot projects.

Recommendations have been established in two fields. Institutional best practices include the adoption of domestic frameworks, participatory instances in early phases of AI system development, adopting a multistakeholder approach and integrating progressive regulatory mechanisms. Technical best practices encompass problem-oriented AI solutions, establishing multidisciplinary groups in earlier stages, and opening datasets for inspection and auditing errors during the pre-processing and labeling stage.