| dc.description.abstract | In an era where data increasingly drives our world, the proliferation of Artificial Intelligence (AI) is undeniable. AI has become integral, not just in our digital lives but in mission-critical domains from medical diagnostics to nuclear systems. This surge in AI utilization brings enhanced efficiency and economic gains but also introduces a disproportionate reliance on autonomous systems without guaranteed formal verification.
The fundamental assumption of any real-world AI application is that training data is representative of real-world encounters. However, the dynamic and inconsistent nature of the real world poses a challenge to this assumption. Integrating complex AI systems into our daily life means imposing static assumptions on a chaotic environment. The risks of such oversights are negligible in consumer electronics but potentially catastrophic in sectors like finance or military technology. Addressing the unpredictability inherent in the real world is not new in AI development, yet unknown variables persist—global pandemics, economic crises, and geopolitical conflicts. The most perilous assumption in AI is presuming all contingencies are accounted for.
This dissertation explores the implications of AI agents operating in dynamic real-world scenarios. It delves into the assumptions AI systems make, their limitations, and how we can enhance methodologies to ensure AI is utilized safely, effectively, and responsibly. Through a blend of practical applications and theoretical foundations of AI, this work proposes methodological innovations aimed at cross-disciplinary impact and real-world problem-solving. By scrutinizing AI behavior and limitations, this research seeks to refine existing practices and foster the beneficial application of AI across various domains. | |
