2025
Reinhard, P., Li, M.M., Fina, M. & Leimeister, J.M.. (2025). Fact or Fiction? Exploring Explanations to Identify Factual Confabulations in RAG-Based LLM Systems. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (CHI EA ’25), April 26–May 1, 2025, Yokohama, Japan. ACM.
The adoption of generative artificial intelligence (GenAI) and large language models (LLMs) in society and business is growing rapidly. While these systems often generate convincing and coherent responses, they risk producing incorrect or non-factual information, known as confabulations or hallucinations. Consequently, users must critically assess the reliability of these outputs when interacting with LLM-based agents. Although advancements such as retrieval-augmented generation (RAG) have improved the technical performance of these systems, there is a lack of empirical models that explain how humans detect confabulations. Building on the explainable AI (XAI) literature, we examine the role of reasoning-based explanations in helping users identify confabulations in LLM systems. An online experiment (n = 97) reveals that analogical and factual explanations improve detection accuracy but require more time and cognitive effort than the no explanation baseline.
Reinhard, P., Moritz, J. M., Wagner, S., & Li, M. M.. (2025). From human-human to human-ai delegation: a leadership theory driven investigation of delegation. Paper presented at the Annual meeting of the academy of management (aom), Copenhagen, Denmark.
The emergence of generative AI (GenAI) has transformed work by enabling humans to delegate tasks like writing and coding to GenAI agents such as ChatGPT. While existing studies highlight AI capability awareness and perceived competence as drivers of delegation, they overlook parallels between human-AI and human-human delegation. Our ongoing research proposes that human-AI delegation can be understood through a leadership lens, with leadership experience and traits as key predictors. Hence, we investigate whether individuals with leadership experience demonstrate higher delegation levels than those without such experience. In an initial online experiment (n=48), participants were grouped by leadership experience and AI transparency to decide whether to delegate or personally perform image classification tasks. Preliminary findings indicate that under a low-transparency condition, leadership experience results in higher delegation rates. However, leadership alone does not significantly predict delegation. Transparency in GenAI consistently leads to higher delegation, while greater domain knowledge corresponds to lower delegation rates. Our ongoing research seeks to deepen understanding of delegation behavior and its predictors in the age of GenAI.
Reinhard, P., Liessmann, A., Weinzierl, S., Zilker, S., Li, M. M., Matzner, M., & Leimeister, J. M.. (2025). Event log construction from multimodal data – a reference architecture for exploiting process mining in it service management. Paper presented at the European conference on information systems (ecis), Amman, Jordan.
Process mining holds substantial potential to discover and optimize processes utilizing event log data. However, current applications primarily rely on (semi-)structured data from process-aware information systems, limiting their capacity to incorporate multimodal data from diverse sources, particularly in domains like IT service management (ITSM). While existing stand-alone approaches can extract event log data from unstructured sources such as videos, documents, or bot logs, they fall short of leveraging the full range of real-world data available in ITSM. To address this gap, our research focuses on developing a reference architecture for constructing event logs from multimodal data. This architecture integrates diverse data types, construction functions, and process mining use cases. Following a design science research methodology, we aim to evaluate the architecture through a software artifact leveraging real-world ITSM data and incorporating state-of-the-art generative AI. In this study, we present the preliminary reference architecture and share early insights from expert evaluations.
Reinhard, P., Li, M. M., Oeste-Reiß, S., & Bretschneider, U.. (2025). Genai-cop: a reusable co-creation process for identifying generative ai agents. Paper presented at the International conference on design science research in information systems and technology (desrist), Montego Bay, Jamaica
Generative AI (GenAI) can enhance organizational processes and productivity. To realize these benefits, organizations must design GenAI agents that augment human work. A key challenge lies in making sense of the diverse forms of GenAI agents and aligning them with existing work processes. To address this, we propose a reusable co-creation process for identifying GenAI agents (GenAI-CoP) that enables organizations to involve domain experts in leveraging GenAI’s potential for their products and workflows. Grounded in action design research (ADR), our approach draws on collaboration engineering to develop GenAI-CoP. We iteratively refined and tested it through simulations, expert interviews, and pilot tests. Our research contributes to GenAI and collaboration engineering literature by introducing a reusable, bottom-up identification procedure for GenAI agents. GenAI-CoP packages facilitation expertise, allowing practitioners to execute it without prior training in tools or techniques. They gain actionable guidelines to identify augmentation potential and summarize it as GenAI agents.