Type: Master

Date: Immediately

Supervisor: Nils Ness, Florian Leiser 

Artificial intelligence systems are increasingly integrated into tasks that traditionally required human intelligence—from visual perception to natural language processing and complex decision-making. While these systems often match or even exceed human-level performance in isolated benchmarks, they do so via fundamentally different mechanisms. Understanding how reasoning unfolds in both humans and AI agents is critical to improving interpretability, aligning system behavior with human expectations, and ultimately designing more robust and transparent AI systems.

In this thesis, we invite a master’s student to investigate how reasoning processes in humans and artificial agents compare across different task types. Rather than focusing on performance outcomes alone, this project emphasizes the nature of reasoning and task understanding—asking: how do intelligent agents (biological or artificial) approach a task, represent it internally, and solve it?

To structure this analysis, the thesis will center around 4–6 well-defined tasks from different cognitive domains:

• Handwritten Digit Recognition (e.g., MNIST)
• CAPTCHA Solving 
• Text Summarization 
• Clinical Decision Support 
• Standardized Test Reasoning (e.g., MCAT/Assessment Centers)

Each task will serve as a lens through which to compare human and machine reasoning approaches using a task-specific but theory-informed framework. The theoretical backbone includes concepts from Distributed Cognition [1] and Cognitive Offloading [2], which help explain how humans interact with tools and environments to externalize cognitive processes. We aim to assess whether similar patterns or strategies appear in AI systems—e.g., through attention mechanisms, memory modules, or multi-agent systems.

Objectives and Methodology

• Taxonomy Development: Based on literature, develop a comparative taxonomy of reasoning across selected tasks. This taxonomy will capture dimensions such as modality of input, degree of abstraction, reliance on external aids, learning mechanisms, and interpretability.
• Morphological Box Construction: Map the taxonomy into a morphological box—a structured overview of how reasoning varies across tasks and agent types (human vs. AI). This enables systematic comparison and identification of reasoning "gaps" or mismatches.
• Literature Review Approach: The thesis is non-empirical and will rely on structured literature analysis using snowballing techniques (backward/forward citations) to ensure a comprehensive yet focused review. Both AI research and cognitive science literature will be used.

Selected References

[1] Hollan, J., Hutchins, E., & Kirsh, D. (2000). Distributed cognition: Toward a new foundation for human-computer interaction research. ACM Transactions on Computer-Human Interaction, 7(2), 174–196.

[2] Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688.

[3] Nickerson, R. C., Varshney, U., & Muntermann, J. (2013). A method for taxonomy development and its application in information systems. European Journal of Information Systems, 22(3), 336–359.