Factors Evaluated by Trust Engines

Are you ignoring a crucial factor for trust during AI implementation? Risking low adoption and ineffective use of AI tools you purchase. Ethics in AI implementation makes sure that technologies align with human rights, avoiding harm, and promoting equitable care for everyone. Essential for building trust among patients and healthcare professionals. Here are 12 important factors to consider to increase trust in AI: Governance:  Establish ethics committees to oversee AI deployment and auditing. Regulatory: Stay compliant with relevant laws and regulations for AI in healthcare. Bias Auditing: Collaborate with vendors to audit tools for bias. Monitoring: Implement ongoing monitoring of AI performance, safety, and ethical compliance. Privacy & Data Security: Enforce strict data protection measures, and limit data retention by vendors. Explainability & Transparency: Require vendors to provide transparent AI models or explanations of decision-making processes. Risk Management: Use risk management frameworks to map,prioritize and mitigate AI-related risks. Education & Literacy: Provide ongoing education for healthcare professionals about ethical use of AI. Informed Consent: Inform patients about the use of AI in their care, and rights to opt out. Clinical Oversight: Review and approve all AI-generated outputs before action. AI Policies: Develop and maintain clear, robust policies regarding AI use. Due Diligence: Evaluate vendors’ ethical practices, data security measures, transparency, and regulatory compliance. Trust is essential if we want AI to be adopted in healthcare. What are you doing to increase trust in AI tools in your organization - for patients and healthcare professionals alike?

I was reflecting on the variety of risk calculations and security scores we all rely on. Having worked cyber risk calculations through financial services companies’ model risk management (MRM) programs, I’ve seen firsthand the level of scrutiny applied. But many other models in use today don’t receive that same level of evaluation—though they probably should. If you’re outside of financial services, how do you replicate that level of investigative rigor? Is there a single “right” cyber model, or does it align more with the axiom that “all models are wrong, but some are useful”? Far too often, trust in cyber risk models is assumed rather than assessed. My latest article in ISACA Journal (Volume 2, 2025) introduces a structured framework for evaluating trust in cyber risk models. Drawing from Aristotle’s rhetorical principles—logos, ethos, and pathos—the framework decomposes trust into three tiers: attributes, artifacts, and evidence. This approach ensures that models are not just mathematically sound, but also transparent, validated, and empirically supported. For organizations relying on cyber risk models, understanding these trust factors is essential to making informed, defensible decisions. Read more in ISACA Journal: https://lnkd.in/ewgfeQCR

This paper from Oct. 10, 2024, "A Comprehensive Survey and Classification of Evaluation Criteria for Trustworthy Artificial Intelligence" by Louise McCormack and Malika Bendechache reviews literature on how to evaluate Trustworthy Artificial Intelligence (TAI). The study focuses on the 7 principles established by the EU's High-Level Expert Group on AI (EU HLEG-AI), outlined in their "Ethics Guidelines for Trustworthy AI" from 2019 (https://lnkd.in/ghha89W9), and further developed in the "The Assessment List for Trustworthy AI" in July 2020 by the AI HLEG (https://lnkd.in/gYWtZ6mk). The paper identifies significant barriers to creating uniform criteria for evaluating trustworthiness in AI systems. To help moving this area forward, the authors analyze existing evaluation criteria, maps them to the 7 principles, and proposes a new classification system to help standardize TAI assessments. Link to paper: https://lnkd.in/gzVDYdaR * * * Overview of the Evaluation criteria for the 7principles of Trustworthy AI: 1) Fairness (Diversity, Non-discrimination): Evaluated using group fairness (metrics based on parity, confusion matrices, etc.) and individual fairness metrics (e.g., counterfactual fairness). Complex fairness metrics are used for specific sensitive data scenarios. 2) Transparency: Assessed through data transparency (data collection, processing, and assumptions), model transparency (how models are developed and explained), and outcome transparency (how AI decisions are understood and challenged). 3) Human Agency and Oversight: Includes evaluating human control (ability to stop AI when needed) and the human-AI relationship (user trust, satisfaction, and understandability). 4) Privacy and Data Governance: Measured using differential privacy (introducing randomness for data protection) and assessing data leakage. Compliance with data governance is evaluated through processes for data collection, processing, and consistency. 5) Robustness and Safety: Robustness is measured by how well AI performs under variable conditions (e.g., unseen data). Safety is evaluated through resilience to attacks, general accuracy, and fallback plans for system failures. 6) Accountability: Assessed through auditability (traceability and documentation) and risk management (documenting how risks are managed across AI development and deployment stages). 7) Societal and Environmental Well-being: Includes evaluating the societal impact (workforce, culture, and harm potential) and sustainability (environmental and economic impact, including energy use and resource consumption). The authors conclude that more research is needed to develop standardized, quantifiable evaluation metrics, specific to different AI applications and industries with sector-appropriate benchmarks.