Certification Labels for Trustworthy AI: Insights From an Empirical Mixed-Method Study

Certification labels can meaningfully increase public trust in AI systems, particularly in high-stakes scenarios like healthcare or criminal justice decisions, according to the survey of 302 census-representative participants—suggesting that visible audit markers may help bridge the gap between expert-level AI documentation and public understanding.

Labels are more effective at addressing certain concerns than others: the study found that certification labels successfully mitigate data-related concerns such as privacy and data protection, but are less effective at addressing concerns about model performance and accuracy—indicating that labels alone cannot solve all trustworthiness issues.

High-stakes AI applications show stronger label effects than low-stakes ones, meaning people care more about certification when AI decisions directly impact their lives (medical diagnoses, loan approvals) versus routine recommendations, suggesting certification label design should be tailored to application context.

Current AI auditing practices are designed for regulators and experts, not the public, creating a communication gap that certification labels could help fill—but the study reveals this remains an "open challenge" requiring careful implementation beyond simply adding a label to an AI system.

Qualitative interviews identified both opportunities and limitations for certification labels, suggesting that while they show promise as a communication tool, they work best as part of a broader trustworthiness strategy rather than as a standalone solution for building public confidence in AI.