A Multistakeholder Approach Towards Evaluating AI Transparency Mechanisms

Different stakeholders need different types of transparency: The paper identifies that various groups affected by AI decisions—including technical experts, end users, regulators, and affected communities—have distinct transparency needs, yet current transparency tools are primarily designed for technical audiences.

Current transparency mechanisms fall short for non-technical stakeholders: The research found that the majority of deployed transparency mechanisms primarily serve technical stakeholders, suggesting that important groups like patients in healthcare or defendants in criminal justice may lack meaningful access to information about how AI systems affect them.

Industry context matters for transparency evaluation: By conducting their study across specific industries (healthcare, criminal justice, and content moderation), the researchers recognize that transparency needs and challenges vary significantly depending on the domain and its unique stakeholders.

Real-world testing is essential for responsible AI: Rather than evaluating transparency mechanisms in isolation, the paper emphasizes the importance of large-scale, mixed-methods user studies across actual organizations to understand how these tools function in practice with diverse stakeholders.

Multistakeholder involvement is critical for AI accountability: The research underscores that responsible AI development requires considering and including perspectives from all affected parties—not just those building the systems—when designing transparency and accountability mechanisms.