AI Safety, Cybersecurity, and Inclusion through Text Analytics
Permanent URI for this collectionhttps://hdl.handle.net/10125/112418
Browse
Recent Submissions
Item type: Item , Identifying and Mitigating Risks Emerging from Uncertainty in AI Agents: A Stakeholder-Centered, Scenario-Based Approach(2026-01-06) Sonnabend, David; Reinhard, Philipp; Li, MaheiToday, AI agents are already capable of performing impressive tasks. However, their non-deterministic nature introduces novel risks that must be understood and mitigated carefully. This study adopts a scenario- based approach to identify and mitigate risks stemming from AI agents’ inherent uncertainties. In a workshop with 30 stakeholders from a medium-sized German software company, participants developed structured risk scenarios based on projections of system behavior, tasks, users, and environment. Qualitative analysis of 35 scenarios and 12 interviews enabled the identification of design principles that enhance stakeholders’ perception of risk manageability, while also assessing the approach’s strengths and limitations. Findings highlight financial risks as predominant and reveal stakeholder preferences for confidence estimators and human-in-the-loop strategies, alongside novel, context- specific mitigation approaches that remain underexplored in current academic literature. These results demonstrate the effectiveness of the scenario- based method in uncovering context-specific risks and fostering the development of mitigation strategies across diverse practitioner groups.Item type: Item , Robust Sentiment Analysis in Service Systems: Enhancing XLNet with Adversarial Training(2026-01-06) Shaw, Laxmi; Ekin, TahirIn service-intensive industries, sentiment analysis plays a pivotal role in understanding customer experiences and driving data-informed decisions. However, these systems are vulnerable to adversarial text manipulations that undermine their reliability. This paper investigates the use of XLNet, a permutation-based transformer model, for sentiment analysis in service contexts and evaluates its robustness under adversarial attacks. We propose a defense mechanism based on adversarial training with embedding-level augmentation. Experiments on diverse benchmark datasets show that while XLNet performs well on clean data, adversarial training significantly enhances robustness without sacrificing accuracy. Our findings offer actionable insights for building trustworthy sentiment analysis pipelines in service systems.Item type: Item , Introduction to the Minitrack on AI Safety, Cybersecurity, and Inclusion through Text Analytics(2026-01-06) Cogburn, Derrick; Wong, Haiman; Ekin, Tahir