Effective human oversight in Artificial Intelligence

Artificial Intelligence is one of the most transformative forces of the twenty-first century. How its diverse benefits can be harnessed while mitigating its various risks, is becoming increasingly important. In response, in July of 2024, the European Commission published the EU Artificial Intelligence Act, with the intention of regulating the development and deployment of AI within the European Union. The overarching goal of the Act is to promote human-centric and trustworthy AI by safeguarding the health, safety and fundamental rights of individuals through a risk-based framework. The Act introduces human oversight as one of the core measures of protection in regulating high-risk AI systems. This thesis explores the role of deployers in implementing oversight and provides insights into how holding overseers accountable can enable effective human oversight. The thesis begins by providing an overview of the EU AI Act, including its definition of oversight and the requirements set for it. Oversight is also examined through relevant ethical guidelines. After providing a definition for accountability and presenting how AI enables its unfair allocation, the thesis determines what constitutes effective oversight, demonstrates how allocating accountability to overseers can enable it in practice and defines the role of deployers in its implementation. The findings indicate that deployers play a critical role in oversight implementation in practice and that effective oversight requires adherence with legal requirements, ethical principles and fulfilment of oversight goals. Moreover, the findings reveal that accountability enhances human task performance, cognitive complexity, sense of morality and information accumulation and processing, consequently, enabling effective human oversight over AI systems. This thesis contributes to the ongoing global discourse on Artificial Intelligence regulation, focusing on the EU AI Act and its imminent implementation. It links accountability and human oversight, aims to bridge the gap between regulation and practical implementation and generally contributes to the discussion around human-centric AI. Ultimately, the thesis provides insights for deployers of high-risk AI systems on the importance of effective human oversight in promoting responsible Artificial Intelligence and safeguarding human autonomy and fundamental values.