Establishing and Evaluating Trustworthy AI: Overview and Research Challenges

AI systems need meaningful human control: The paper identifies "human agency and oversight" as a foundational requirement for trustworthy AI, emphasizing that humans must remain meaningfully involved in decisions, particularly in high-stakes domains where AI systems have previously produced "unexpected or undesirable outcomes."

Fairness and non-discrimination require ongoing vigilance: The paper highlights fairness as a core trustworthiness requirement, recognizing that AI systems can perpetuate or amplify discrimination. However, it notes this is complicated by context-dependency—what counts as fair varies significantly depending on the specific application and stakeholders involved.

Transparency alone isn't enough; explainability matters: While the paper identifies both transparency and explainability as essential requirements, it distinguishes between them, suggesting that simply making AI systems more transparent doesn't automatically make them understandable to users and affected communities.

Real-world testing is critical but underutilized: The paper identifies "investigations in real-world contexts" as a major overarching research challenge, indicating that many trustworthiness evaluations happen in controlled settings rather than where AI actually impacts people's lives.

Trustworthy AI requires teamwork across disciplines: The paper emphasizes that establishing trustworthy AI demands interdisciplinary collaboration beyond computer science alone, requiring input from ethicists, social scientists, domain experts, and other fields to adequately address the six trustworthiness requirements.