Julian Benz 

Motivation 

Industrial paper-printing machines are highly instrumented and generate dense process and control time series. A large share of their lost availability originates not from steady-state production but from changeovers, the transitions between two production runs, during which many interacting parameters must be brought back to a stable operating point. How to perform a changeover quickly and reliably is knowledge that experienced operators hold largely tacitly, rather than something captured in any system.

Data-driven methods can help characterize changeover phases, uncover recurring machine states and identify influential parameters. But in a setting like this, purely automated optimization falls short: the relevant expertise is experiential and tacit, decisions carry operational and safety consequences, and operators must understand, trust and ultimately own the resulting changes. This is why human–AI collaboration matters here: the goal is not to replace operators, but to design AI systems that work with them, keep them in the loop, and strengthen rather than sideline their competence.

In cooperation with BATO, an early-stage startup building intelligence for operations, we therefore offer master theses at this intersection, with the chance to work on real industrial data. The two topics below each take up a different facet of human–AI collaboration in this context; one will be selected and refined together with you, according to your background and interest. A thesis can also be a starting point for a deeper involvement with BATO, for example through a working student position.

Topic 1: Explainable AI as a Reflective Partner 

Explainable AI is usually framed as a way to justify a system's recommendation. Yet where operators already hold rich process knowledge, the more valuable role of explanations may be to trigger reflection, confronting practitioners with patterns in their own changeover behaviour, showing where their routines diverge from data-driven expectations, and prompting them to form and test hypotheses rather than simply comply.

At the same time, such assistance carries a critical trade-off. The same support that improves short-term performance may, if consumed passively, erode operators' own problem-solving and process expertise over time, an industrial counterpart to the intellectual deskilling documented in knowledge work. The thesis therefore investigates how explanations and feedback should be designed so that operators engage critically with their own practice rather than following the system passively, and how this design shapes their competence over time: whether it maintains and enhances skills (upskilling) or erodes them (deskilling), and how the two differ between experienced and novice users.

Depending on your background, the project can be design-oriented (a prototype plus user evaluation) or more empirical, with a focused research question developed together.

Topic 2: Integrating Operators' Tacit Knowledge

Much of what makes a changeover fast and stable lives as tacit, experiential knowledge in the heads and hands of operators, and is never represented in the data a model learns from. Purely data-driven optimization therefore risks ignoring exactly the expertise that distinguishes good practice, while operators, in turn, have no structured way to feed their knowledge back into the system.

The thesis explores how operators' tacit knowledge can be elicited and integrated into a data-driven optimization approach, for instance through human-in-the-loop feedback, knowledge-elicitation interfaces, or hybrid mechanisms that combine machine-discovered patterns with human process expertise. The aim is a concept, and where feasible a prototype, in which system and operator jointly produce better outcomes than either alone.

Profile

• Interest in interdisciplinary research on human interaction with AI systems
• Python fundamentals; basic understanding of ML / time-series (depth depending on the chosen approach)
• Self-driven, structured working style and curiosity for real production data
• English skills

Contact

We offer an exciting research topic with strong relevance to both academia and practice, close supervision, and the opportunity to develop theoretical, methodological, and practical skills. If you are interested, please send a current transcript of records, a short CV, and a brief motivation (2–3 sentences) to Julian Benz (julian-david.benz@kit.edu)

Literature 

• Kosmyna, Nataliya, Eugene Hauptmann, Ye Tong Yuan, u. a. „Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task“. arXiv:2506.08872. Preprint, arXiv, 31. Dezember 2025.
• Förster, M., Broder, H. R., Fahr, M. C., Klier, M., & Fink, L. (2025). Tell me more, tell me more: the impact of explanations on learning from feedback provided by Artificial Intelligence. European Journal of Information Systems, 34(2), 323-345.
• Förster, M., Schröppel, P., Schwenke, C., Fink, L., & Klier, M. (2024). Choose Wisely: Leveraging Explainable AI to Support Reflective Decision-Making. International Conference on Information Systems.