Date of Award
3-19-2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Information Science
First Advisor
Ningning Wu
Abstract
Large Language Models (LLMs) now underpin critical systems in education, healthcare, and policy advisory contexts, yet remain susceptible to the recursive amplification of bias, misinformation, and error (BME). These distortions, driven by feedback loops between user interactions, model retraining, and data reuse, accelerate epistemic drift and erode information quality over time. This dissertation advances a comprehensive governance architecture to detect, quantify, and mitigate these risks across the full AI lifecycle. The research introduces the BME Metric Suite, a standards-aligned diagnostic framework comprising five original metrics: Bias Amplification Rate (BAR), Echo Chamber Propagation Index (ECPI), Information Quality Decay (IQD), Pretraining Diversity Index (PTDI), and Architectural Hallucination Risk Score (AHRS). These metrics provide quantifiable indicators of drift, redundancy, and hallucination, forming the analytic foundation for continuous lifecycle assurance. Empirical validation using Six Sigma statistical process control, simulation, and structured prompt testing demonstrates the reliability, interpretability, and cross-domain generalizability of these methods. Building on these diagnostics, the dissertation operationalizes the AI Lifecycle Audit and Governance Framework (ALAGF), an integrated model that unites MIDCOT (Multi-Dataset IQ Drift & Cost Optimization Training) for data governance with SymPrompt+ for oversight of structured human–AI interaction. Together, these components establish a unified governance ecosystem aligned with ISO/IEC 42001, ISO/IEC 23053, ISO 8000, and the NIST AI Risk Management Framework. The outcome is a scalable, auditable pathway for integrating technical AI assurance with ethical, regulatory, and organizational governance. The dissertation concludes by introducing Continuous Audit Maturity (CAM) and proposing the “Measure–Monitor–Govern” model as a repeatable standard for lifecycle-aligned AI stewardship.
Recommended Citation
Rutherford, Dale A., "Governing BME Amplification in Large Language Models: A Standards-Aligned Framework for LLM Data Lifecycle Risk Detection and Ethical AI Governance" (2026). Theses and Dissertations. 1329.
https://research.ualr.edu/etd/1329
