A Human-Centric Assessment Framework for AI

There is currently no standard way to evaluate whether AI explanations are actually understandable to humans. The paper identifies this gap and proposes that domain experts should be the judges of whether AI system explanations make sense, rather than relying on technical metrics alone.

A "human-centric" evaluation approach can measure AI performance by comparing it directly to expert judgment. By having domain experts evaluate both an AI system's solutions and a human expert's solutions, researchers can determine whether the AI performs comparably to human-level work—similar to how the Turing test measures machine intelligence.

The framework can incorporate real-world uncertainty in expert opinions. Rather than assuming experts always agree perfectly, the assessment method accounts for cases where domain experts themselves are uncertain about correct answers, making evaluations more realistic and fair.

This approach is flexible and can be adapted for different types of AI assessments. The paper demonstrates two different applications: measuring classification accuracy and evaluating whether explanations are actually useful to humans, suggesting the framework could work across various AI systems and domains.

Trustworthy AI requires moving beyond technical metrics to human judgment. The paper emphasizes that responsible AI development needs assessment methods grounded in how real experts and users actually interact with and understand AI systems, not just algorithmic performance measures.