Proposal Description

Statistical science has a foundational role to play in the responsible development and deployment of artificial intelligence. As AI systems move into consequential decision environments spanning healthcare, digital advertising, environmental monitoring, and public governance, questions of transparency, accountability, and robustness have become urgent scientific priorities. This session brings together four speakers and a session chair from Nigeria/Germany, Turkey, USA, and Italy whose work addresses these challenges from complementary statistical perspectives. Explainable machine learning for cross-domain decision support, adversarial risk analysis for Bayesian forecasting systems under strategic manipulation, interpretable ML-based classification for sustainable waste management, and generative AI methods for scalable Bayesian computation will be discussed. Together, the talks position statistical rigor as the cornerstone of trustworthy AI — not merely as a technical requirement, but as a prerequisite for ethical and socially responsible deployment of AI across science, industry, and public policy.

The four talks in this session address distinct but interconnected dimensions of trustworthy and robust AI, spanning interpretability frameworks, adversarial robustness, applied classification, and generative computation.

Awe introduces a unified framework for interpretable and accountable machine learning, operationalized through civic.icarm, a new R package that integrates data preprocessing, predictive modeling, fairness evaluation, and explainability diagnostics. The framework employs a range of machine learning models and applies feature importance analysis, partial dependence profiling, and local explanation methods to make model outputs actionable and auditable. The work positions interpretability not merely as a technical property but as an instrument of accountability and participatory governance across healthcare, education, finance, and public policy.

Ekin examines a different dimension of AI trustworthiness: robustness under adversarial attack. Focusing on sequential Bayesian count data forecasting in the context of click fraud, the paper develops an adversarial risk analysis framework that models data poisoning attacks from both attacker and defender perspectives. By considering how perturbations propagate through filtering and forecasting uncertainty mechanisms across multiple attacker archetypes, the framework provides a statistically grounded approach to quantifying and mitigating the impact of deliberate manipulation on AI-driven decision systems.

Kocadagli demonstrates the practical power of interpretable ML in an industrial sustainability context. Using FTIR-ATR spectroscopic data from 200 wood-waste samples, a decision support system is developed and benchmarked across eight ML classifiers — including ANN, random forests, SVM, and Naive Bayes — achieving classification accuracy approaching 100%. The interpretable design of the system, which identifies 52 key wavelengths as discriminating features, enables practitioners to understand and trust the basis of automated classifications, providing an efficient and ergonomic alternative to classical chemometric methods in waste wood management.

Polson addresses the computational frontier of Bayesian AI, proposing generative AI methods to replace traditional Bayesian computation. By training deep neural networks on large simulated datasets, the approach learns the inverse Bayes map — mapping observations to posterior distributions — without relying on closed-form densities. Deep Quantile Neural Networks provide a flexible framework for inference and prediction, demonstrated on traffic flow and epidemiological data. This model-free approach to Bayesian computation offers significant scalability advantages and opens new directions for integrating generative AI into mainstream statistical practice.