Towards Modelling and Verification of Social Explainable AI

Social Explainable AI needs security testing: The paper identifies that current SAI systems focus on delivering basic functionality but largely ignore potential vulnerabilities to malicious attacks or errors—a gap that formal verification methods can help address.

Multi-agent system tools can protect AI transparency: The researchers demonstrate that verification techniques from multi-agent systems research can be adapted to test whether decentralized, transparent AI environments can resist attacks from compromised AI modules, suggesting a practical pathway for security assessment.

Formal verification is feasible for AI accountability: Using the STV model checker, the authors show that it's possible to formally model and verify SAI environments, indicating that rigorous mathematical approaches can help ensure these systems behave as intended rather than relying solely on testing and monitoring.

Compromised AI modules pose real risks: The paper highlights that individual AI components within a system can be malicious or faulty, and that social AI systems—which emphasize decentralization and human interaction—need explicit defenses against these internal threats.

Early-stage AI governance requires proactive security design: By addressing security vulnerabilities at the modeling stage rather than after deployment, this work suggests that responsible AI development should incorporate formal verification methods from the outset, particularly for systems designed to be transparent and user-focused.