This Master's thesis investigates the design and implementation requirements for Large Language Model (LLM)-based chatbots in Swiss correctional facilities, addressing the critical challenges of digital inequity and social isolation among inmates. While AI-driven technologies show promise for enhancing rehabilitation and digital literacy, their deployment in high-security environments presents unique technical, functional, and ethical challenges that remain underexplored.

Building on Siegmann & Bendel's (2024) work on social robots in prisons, this study extends the discourse to more accessible chatbot technologies. The research employs a pragmatic, qualitative case study approach, combining systematic literature review, document analysis of Swiss and international digitalization initiatives, semi-structured interviews with seven correctional system experts, and technical prototyping to validate offline LLM feasibility.

The findings reveal stringent requirements across multiple dimensions: technically, systems must operate fully offline with robust security controls and tamper-proof hardware; functionally, they require multilingual support, simple interfaces, and clear institutional oversight mechanisms; content must be strictly vetted with comprehensive prohibitions on security-compromising information; and ethically, implementations must preserve human interaction while ensuring data privacy and preventing algorithmic bias. The thesis contributes a comprehensive requirements framework and a practical implementation guide for Swiss correctional facilities, balancing the potential benefits of digital tools with the paramount security and human-centric principles of correctional work. The research demonstrates that while LLM-based chatbots can support inmate education and reintegration, their successful deployment demands careful calibration between technological innovation and institutional constraints.