An Open API Architecture to Discover the Trustworthy Explanation of Cloud AI Services

Explainability can be added to existing AI services without redesigning them: The paper demonstrates that feature contribution explanations (showing which parts of input data influenced a prediction) can be provided for cloud AI services through an open API architecture, meaning organizations don't need to rebuild their AI systems from scratch to make them more transparent.

Trustworthiness of AI services can be measured and compared: The researchers developed metrics to evaluate both model performance and "XAI consistency" across leading cloud vision AI services, providing a practical framework for organizations to assess and compare the trustworthiness of different AI providers' offerings.

Data quality directly impacts explanation reliability: The study found that data augmentations (variations of training data) resulted in measurable improvements in XAI consistency metrics, suggesting that how AI systems are trained affects not just accuracy but also the quality and reliability of explanations they provide.

Open standards enable better AI governance: By designing the explainability service as cloud-agnostic (working across different providers), the architecture supports independent auditing and assessment of AI services, reducing vendor lock-in and enabling organizations to hold multiple AI providers accountable to the same transparency standards.

Reproducibility strengthens AI accountability: The system collects provenance data from operational pipelines to enable reproducibility, meaning organizations can verify and audit how specific predictions were made—a critical requirement for responsible AI deployment in high-stakes applications.