CAUSAL HORIZONS IN AI: STATISTICAL INNOVATIONS FOR ETHICAL AND ROBUST INTELLIGENCE

The aim of this article is to explore the evolving intersection between Statistics and Artificial Intelligence (AI), particularly in the context of Machine Learning (ML) advancements, to identify critical gaps and propose pathways for future research. The scope encompasses a systematic review of recent literature on statistical methodologies applied to AI systems, focusing on explainability, testing, evaluation, and integration in data platforms. It is examined how traditional statistical tools are being adapted or revolutionized by AI techniques, such as in parameter calibration, uncertainty quantification, and performance metrics. The scientific methods employed include a comprehensive literature search on the Web of Science and Clarivate databases, resulting in the selection of 30 relevant articles based on relevance to "statistics in AI" themes like explainable AI (XAI), differential testing, agentic recovery and probabilistic modeling. Each article was analyzed using content summarization tools to extract key statistical contributions, methods, and limitations. Thematic analysis was conducted to synthesize findings, identifying patterns in statistical applications (e.g., bootstrapping, F1-scores, constraint analysis) and gaps via qualitative coding. Results reveal robust advancements in statistical integration for AI reliability, e.g., multimodal evolutionary algorithms for model calibration achieving 1 - 4% improvements in loss functions and process reward models outperforming outcome-based ones by 19% in multi-tool reasoning accuracy. However, a prominent missing domain emerges: the application of causal statistical inference to mitigate biases in AI decision-making systems, where current literature emphasizes descriptive and predictive stats but under-explores counterfactual reasoning for ethical AI deployment. This gap forms the "missing link", enabling novel connections between causal stats and AI fairness. Three research questions (RQs) are developed to bridge this and are integrated throughout, guiding the discussion. Future research directions include empirical validation of causal stats in XAI tools, development of open-source benchmarks for bias-causal links, and interdisciplinary collaborations to embed causal literacy in AI curricula. This work underscores statistics' pivotal role in making AI more interpretable, robust, and equitable, paving the way for trustworthy AI ecosystems.