Authority and Human Oversight in Agentic AI

Topic Description

AI agents increasingly act on people's behalf. They send messages, change code, spend money and pass tasks to other agents. Each action depends on authority that someone delegated. In practice, this authority is often unclear, and it can be hard to know who allowed what, and who is accountable when things go wrong. Regulation such as the EU AI Act calls for meaningful human oversight, but offers little guidance on how to achieve it for agents.

This PhD will investigate how authority is delegated to, and exercised by, AI agents, whether in software development or in the wider settings where agents act on people’s behalf, such as workplaces, public services and everyday life.

Possible directions include:

• How authority is granted to, and exercised by, AI agents.
• Human oversight of agents in practice, and what makes it meaningful.
• Accountability and traceability when agents act for others or hand tasks to other agents.
• How regulation such as the EU AI Act shapes agentic software practice.
• How authority is expressed, negotiated and understood in natural-language interaction with agents.

Skills Required:

A good degree (2:1 or above) in Computer Science, Software Engineering, HCI or a closely related subject, or a postgraduate conversion degree in Computing. Familiarity with current software development practice is essential, and some understanding of AI agents or LLM-based tools is desirable. You should have experience of, or a strong interest in, empirical research methods (e.g. interviews, surveys, case studies, qualitative and quantitative analysis). Curiosity about the human, organisational and regulatory sides of computing, and the ability to communicate ideas clearly to different audiences, are important.

Background Reading: