Trustworthy AI: Ensuring Reliability and Accountability from Models to Agents

Bias affects overlooked populations: The research demonstrates that traditional demographic categories often miss important subgroups vulnerable to biased ML decisions. The thesis introduces kernel-based methods to achieve "multiaccuracy" across complex subpopulations, suggesting that fairness efforts must look beyond standard demographic categories to protect all affected groups.

Multiple accurate models can give conflicting results: A critical finding is that different ML models can be equally accurate yet make contradictory predictions for the same individual. This "predictive multiplicity" problem means that model accuracy alone doesn't guarantee fair or consistent outcomes, highlighting the need for additional safeguards when deploying AI systems.

Watermarking can verify AI-generated content without major quality loss: The research characterizes the trade-off between detecting watermarks in AI-generated text and maintaining text quality, showing that watermarks can be embedded in large language models to prove content provenance. This addresses accountability concerns by enabling detection of AI-generated content while preserving usability.

LLM agents show promise but need careful evaluation: The thesis develops the first simulator for fully AI-driven supply chains, finding that LLM-based agents outperform human teams and reduce costs. However, this emerging capability underscores the importance of rigorous evaluation frameworks for autonomous AI systems before real-world deployment.

Trustworthy AI requires multiple complementary approaches: Rather than a single solution, the research demonstrates that reliability and accountability require combining techniques from information theory, optimization, and statistical learning—suggesting that responsible AI deployment demands layered safeguards addressing bias, arbitrariness, provenance, and autonomous behavior simultaneously.