Trustworthy AI

AI systems today are powerful but unreliable: The paper notes that while AI has achieved "good enough performance" to operate in public spaces and homes, these systems "can be brittle and unfair," highlighting a critical gap between capability and trustworthiness that must be addressed before widespread deployment.

Trust is essential for AI adoption: The authors argue that for society to benefit from AI, we must first establish mechanisms to trust these systems—suggesting that without trustworthiness, the promise of AI cannot be realized regardless of its technical capabilities.

Formal verification offers a path forward: The paper proposes formal verification—a mathematical approach borrowed from trustworthy computing—as one method to ensure AI systems are reliable and safe, representing a shift toward more rigorous validation methods beyond traditional testing.

Trustworthy AI requires new research: The authors frame trustworthy AI as raising "the ante" on both trustworthy computing and formal methods, indicating that existing approaches are insufficient and new research questions must be explored to adequately address AI's unique challenges.

Multiple trustworthy properties are needed: By drawing on decades of trustworthy computing research, the paper suggests that trustworthy AI systems should possess multiple desired properties (such as fairness and reliability), rather than optimizing for a single metric.