Certified AI System = Trustworthy? Exploring Expert and Lay User Perceptions and Needs Regarding AI Certification

Paper Title

Certified AI System = Trustworthy? Exploring Expert and Lay User Perceptions and Needs Regarding AI Certification

Publication Info

• Topic area: User perceptions and design of AI certification systems
• Keywords: AI certification, trust in AI, HCI, transparency, post-certification monitoring, certification fraud, user-centered design, AI governance, high-risk AI, expert vs. non-expert perceptions

Background and Problem

• Problem / challenge: Limited empirical understanding exists of how AI certifications influence user trust, especially across different user groups (experts vs. lay users). Static certification approaches may not address the evolving and context-dependent nature of AI systems.
• Significance: AI certification is critical for fostering trust, ensuring compliance with ethical standards, and mitigating risks in high-stakes applications. Understanding user expectations is essential for designing effective certification schemes.
• Motivation and related work: Previous studies show mixed results on the effectiveness of AI certifications in building trust. Research has focused on technical and regulatory aspects but has largely overlooked user-centered perspectives. This paper addresses this gap by exploring how experts and non-experts perceive AI certification and its associated processes.

Solution

• Proposed approach: A qualitative, user-centered study to investigate expert and lay user perceptions of AI certification, including trust impacts, post-certification monitoring, and fraud prevention.
• Novelty:
  1. Empirical comparison of expert and non-expert perceptions of AI certification.
  2. Identification of user needs for transparency, monitoring, and fraud prevention in certification schemes.
  3. Actionable recommendations for designing user-centered, adaptive AI certification systems.
• Procedure and key techniques:
  • Conducted semi-structured interviews with 30 participants (15 experts, 15 non-experts).
  • Included a drawing exercise to visualize participants’ mental models of the certification process.
  • Analyzed data using thematic analysis to identify differences in perceptions and expectations between the two groups.

Results

• Concrete findings:
  • Non-experts viewed certification as a trust cue, while experts were more skeptical and emphasized understanding the certification process.
  • Both groups preferred independent organizations or NGOs as certifiers, with experts more open to private companies.
  • Experts favored update-based post-certification monitoring, while non-experts preferred annual checks.
  • Fraud concerns included copying genuine labels, self-issued certificates, and unauthorized certifiers.
• Advantage over baselines:
  • Highlights the need for multi-level transparency to address varying user expertise.
  • Proposes context-specific and adaptive certification standards to handle AI system diversity.
  • Suggests robust fraud prevention mechanisms combining technical, legal, and awareness-based measures.
• Experiments / evaluation:
  • Participants included 15 AI experts (average age 33.67 years) and 15 non-experts (average age 39.53 years) from diverse educational and professional backgrounds.
  • Data collection involved interviews and drawing exercises, analyzed using thematic coding.
• Limitations and future work:
  • Sample limited to WEIRD (Western, Educated, Industrialized, Rich, Democratic) countries and academia-heavy expert group.
  • Results are qualitative and exploratory, limiting generalizability.
  • Future work should include industry experts, non-WEIRD participants, and explore perceptions of human-AI collaborative systems.

Summary

This study investigates how AI certification influences trust among experts and non-experts, revealing significant differences in perceptions and expectations. Non-experts rely on certification as a trust signal, while experts demand transparency and robust processes. Both groups emphasize the importance of independent certifiers, ongoing monitoring, and fraud prevention. The findings highlight the need for user-centered, context-specific certification schemes with multi-level transparency and adaptive standards. Recommendations include hybrid certification models, automated monitoring pipelines, and secure, verifiable certification labels. These insights contribute to designing AI certification systems that are both technically robust and socially relevant.