IRIS Publications

Abstract

How much can Large Language Models (LLMs) influence the attitudes and opinions of their users? Answering this question requires controlled pre/post-treatment experiments, where participants interact with LLMs that consistently adopt a predefined political stance. Such experiments, however, are only possible if LLMs can be reliably steered to hold these stances throughout the interactions. In this work, we evaluate whether state-of-the-art LLMs can be effectively stance-controlled in German, thereby enabling experiments on human–LLM interactions. First, using a corpus of realistic user prompts, we find that LLMs are predominantly neutral, making them infeasible for said experiments. We then show that a prompt-based stance control method can reliably guide models to argue for or against a particular topic. Finally, we analyze confounding factors like topic and stance of the initial user prompts. We find that control is easiest when the target stance aligns with topical priors of the model or a user’s prompt. Further, the models maintain a comparable style across target stances — a key prerequisite for pre/post-treatment experiments. Taken together, our results demonstrate that stance-controlled LLMs are feasible and practically useful for experiments on user attitude change.

Abstract

Jailbreaking -- bypassing built-in safety mechanisms in AI models -- has traditionally required complex technical procedures or specialized human expertise. In this study, we show that the persuasive capabilities of large reasoning models (LRMs) simplify and scale jailbreaking, converting it into an inexpensive activity accessible to non-experts. We evaluated the capabilities of four LRMs (DeepSeek-R1, Gemini 2.5 Flash, Grok 3 Mini, Qwen3 235B) to act as autonomous adversaries conducting multi-turn conversations with nine widely used target models. LRMs received instructions via a system prompt, before proceeding to planning and executing jailbreaks with no further supervision. We performed extensive experiments with a benchmark of harmful prompts covering several sensitive domains. This setup yielded an overall jailbreak success rate across all model combinations of 97.14\%. Our study reveals an alignment regression, in which LRMs can systematically erode the safety guardrails of other models, highlighting the urgent need to further align frontier models not only to resist jailbreak attempts, but also to prevent them from being co-opted into acting as jailbreak agents.

Abstract

This paper presents a human-centered cyber-physical system integrating gesture recognition, depth sensing, and motion sensing to support human–robot collaboration (HRC) in timber assembly. The system enables users to control cobots on demand, adapts cobot trajectories to workspace changes, and ensures safety through motion-triggered stops, while a visual interface displays real-time sensor data and planned actions. We evaluated the system in a small-scale timber assembly study with 21 participants—Novice Academics, Experienced Academics, and Novice Professionals—performing assembly tasks under three modes: Human Agency (gesture control), Robot Agency (autonomous sensing), and Combined Agency (integrated). Results show that gesture control achieved the highest perceived usability and the lowest mental and physical demand, whereas autonomous sensing produced the fastest completion times (median 6.02 min). The combined mode, intended to merge both paradigms, paradoxically led to slower performance (median 10.15 min) and higher cognitive load, as users had to coordinate overlapping control schemes. Although some task-time data for professionals were unavailable, qualitative feedback revealed consistent trends: academics valued autonomy and transparency, while professionals prioritized safety and reliability. These findings highlight that user background profoundly shapes collaboration preferences and that flexible interfaces—allowing adjustable safety thresholds, feedback modes, and levels of autonomy—are essential for effective and inclusive HRC design in construction contexts.

Abstract

This research addresses the gap between technical advancement and human factors in human--robot collaboration (HRC) education for digital fabrication. Current pedagogical approaches predominantly focus on technological capabilities while neglecting critical aspects of user diversity, accessibility, and contextual adaptability essential to Industry 5.0 implementation. We present a human-centered HRC design methodology integrating feminist technoscience (FTS) dimensions with behavioural fabrication techniques. The framework consists of three components: (1) an FTS framework and framework to guide and evaluate HRC design; (2) a cyber-physical system enabling real-time sensor-driven robot behavior; and (3) a modified Double-Diamond Approach (DDA) that systematically guides students through creative processes, for iterative design. The methodology was tested through a four-day workshop with graduate architecture and engineering students, who developed four distinct HRC workflows incorporating voice control, gesture recognition, computer vision feedback, and multilingual interfaces. Using the FTS rubric, we observed that students were able to translate abstract concepts---such as user diversity, agency, and bias---into concrete design decisions, despite limited prior experience in robotics or feminist critique. These outcomes suggest that the integrated approach effectively supports students in addressing practical HRC challenges while remaining attentive to the social and ethical dimensions of automation. This work contributes a reproducible methodology for preparing future practitioners to design HRC systems that are technically robust while remaining responsive to user needs, thereby advancing the transition from Industry 4.0 to 5.0 priorities.

Abstract

Cognitive diversity has driven societal progress, but it can provide challenges in daily workplace interactions. Hence, we developed a modular training paradigm using virtual reality (VR) and structured cognitive exercise to build transferable skills and enhance awareness of cognitive differences. The training strengthens core executive functions, which are subsequently applied to simulated and real-life scenarios. To evaluate performance impacts and technology acceptance, 17 neurodivergent adults explored our VR-based spatial working memory training module. Descriptive results for task performance and the UTAUT dimensions showed performance improvements, low anxiety levels, and high perceived self-efficacy. Participants further expected high effort and low performance gains in work situations, leading to low engagement intentions. These findings suggest that the abstract nature of the training module may have hindered its perceived applicability to practical tasks. While the training demonstrates potential for skill enhancement, refining its relevance to workplace scenarios is crucial for sustained use.

Abstract

This paper presents a novel adaptive training system that integrates a Virtual Reality (VR) environement, real-time neurofeedback (NF) via functional near-infrared spectroscopy (fNIRS), and tactile stimulation, specifically designed to enhance cognitive skills in autistic adults. Central to this system is an integration platform that enables real-time, synchronised interplay between user behaviours and neurophysiological data. The presented neuroadaptivity loop employs a feedback mechanism that continuously assesses users’ cognitive states through fNIRS, allowing the system to dynamically adjust environmental complexity and send tactile cues based on real-time mental state indicators. This multimodal approach fosters cognitive executive functions relevant for managing everyday situations. By prioritising user-centred design through iterative feedback from autistic individuals, the platform ensures usability and safety while promoting skill transfer from structured training to real-world applications. The implications of this system extend to personalised, adaptive technologies in workplace training, highlighting important considerations regarding user empowerment.

Abstract

Text-to-image models are increasingly popular and impactful, yet concerns regarding their safety and fairness remain. This study investigates the ability of ten popular Stable Diffusion models to generate harmful images, including sexual, violent, and personally sensitive material. We demonstrate that these models respond to harmful prompts by generating inappropriate content, which frequently displays troubling biases, such as the disproportionate portrayal of Black individuals in violent contexts. Our findings demonstrate a complete lack of any refusal behavior or safety measures in the models observed. We emphasize the importance of addressing this issue as image generation technologies continue to become more accessible and incorporated into everyday applications.

Abstract

Given the ever-evolving demands of the professional world, higher education plays a vital role in equipping students with strategies for self-organized and sustainable skill development. This enables students to quickly and independently adapt to new knowledge and skills throughout their careers. Therefore, it is of the essence to integrate methods that enhance self-regulation skills into our study programs, alongside the instruction of specific subject-matter expertise. Addressing these demands, we introduce a focused training that embeds the assistive software tool focUS into a structured seminar concept, leveraging eduScrum elements and accompanying learning communities. After introducing the results of a pilot evaluation with a small group of interdisciplinary doctoral students, we discuss possibilities of technical and conceptual integration in course curricula and learning counseling in higher education. Taken together, our approach indicates value for empowering future professionals across various domains to unleash their full potential.

Abstract

Introduction The persistent prevalence of distractions challenges people's capacities to study and work productively. Digital tools can support focusing on meaningful tasks with features like time tracking, feedback, or rewards. Evidence exists for benefits regarding behavioral focus, distraction management, and motivation, but also for hesitation to use digital support at all. Building on the Unified Theory of Acceptance and Use of Technology (UTAUT), this research explores which factors can foster or hinder the intention to use an exemplary software to support self-regulation. Methods A sample of 96 adult volunteers watched a short introductory video explaining how the software focUS fosters self-regulated studying and working. Subsequently, participants completed an online survey to capture their willingness to use focUS, expected gains and challenges, and individual characteristics. Results Participants expressed stronger intentions to use focUS when they expected increased benefits in performance. By trend, lower levels of expected effort also hinted on stronger intentions to use focUS. Contrary to expectations, participants lacking previous experience with software in the scope described, who anticipated higher effort when using focUS, tended to express stronger intentions to use it. Discussion Taken together, our evidence suggests that highlighting specific performance improvements may encourage the use of digital support for tasks that require self-regulation. Ensuring the use of digital support to be as effortless as possible could provide yet another compelling argument to use it. Particularly for inexperienced users, sparking curiosity for the challenges of the unknown might be a worthwhile strategy to reduce hesitation towards emerging technologies.

Abstract

In today's high-choice media environment, the relations between news consumption and (extreme) political views receive extensive attention. While social media and news aggregators are often blamed for political polarization, theories from media psychology highlight the importance of additional aspects of news consumption, aside from the media types used, in impacting the formation of political views and thereby, for instance, polarization. However, these aspects are typically studied in isolation, neglecting the heterogeneity in where and how news can be consumed. Based on this observation, the present work investigated the conjoint relations of various aspects of news consumption across media types with political views and their extremity. Study 1 included two samples of Nþinspace=þinspace714 (sample 1) and Nþinspace=þinspace721 (sample 2) participants completing an online survey assessing political news consumption habits and general political views, namely, ideology and affective polarization. Study 2 included two samples of Nþinspace=þinspace725 (sample 3) and Nþinspace=þinspace713 (sample 4) participants completing scales on issue-specific political views, namely, attitudes on war and the war against Ukraine. Latent Profile Analyses revealed four mostly media type-specific news consumer profiles replicated across samples, indicating that individuals choose one media type, like social media, aggregators, or print media, and devote considerable time to consuming political news from various sources, including attitude-incongruent news and diverse perspectives, from the same selected media type. While these profiles differed on sociodemographic and personality variables, few significant relations were found between these profiles and political views. These findings highlight the complexity of these relations and challenge the assumption that only certain media types, and news consumption via them, drive political extremity.

Abstract

The present study explores individual differences related to the perception of the media coverage of immigrants as biased. Building on previous research, relations with the extremity in preexisting attitudes toward immigrants and in affective ratings of non-immigrants versus immigrants are examined. Additionally, the present study extends previous work by investigating dogmatism and intellectual humility, including their potential moderating roles on the relations of extremity in attitudes and affective ratings with perceived media bias. A sample of Nþinspace=þinspace212 (59\% men) individuals from the general German population completed self-reports on their preexisting attitudes and affective ratings, as well as dogmatism and intellectual humility online. Moreover, participants rated their perception of news media coverage of immigrants as biased against their views. Results indicate that particularly more extreme negative attitudes toward immigrants and affective ratings favoring non-immigrants relative to immigrants are positively associated with perceiving the media coverage of immigrants as biased. No robust relations of dogmatism, intellectual humility, or their interactions with the extremity scores with perceived media bias were found. These findings underline the importance of negative extremity in (out) group-related attitudes and affect in perceiving the media coverage as biased. New approaches investigating media perceptions from a group-related perspective are discussed.

Abstract

Workers are increasingly embracing Artificial Intelligence (AI) to optimise various aspects of their operations in the workplace. While AI offers new opportunities, it also presents unintended challenges that they must carefully navigate. This paper aims to develop a deeper understanding of workers’ experiences with interactions with automated agents (AA) in the workplace and provide actionable recommendations for organisational leaders to achieve positive outcomes. We propose and test a simulation model that quantifies and predicts workers’ experiences with AA, shedding light on the interplay of diverse variables, such as workload, effort and trust. Our findings suggest that lower-efficiency AA might outperform higher-efficiency ones due to the constraining influence of trust on adoption rates. Additionally, we find that lower initial trust in AA could lead to increased usage in certain scenarios and that stronger emotional and social responses to the use of AA may foster greater trust but result in decreased AA utilisation. This interdisciplinary research blends a systems dynamics approach with management theories and psychological concepts, aiming to bridge existing gaps and foster the sustainable and effective implementation of AA in the workplace. Ultimately, our research endeavour contributes to advancing the field of human-AI interaction in the workplace.

Abstract

Pretrained language models (PLMs) have been shown to encode binary gender information of text authors, raising the risk of skewed representations and downstream harms. This effect is yet to be examined for transgender and non-binary identities, whose frequent marginalization may exacerbate harmful system behaviors. Addressing this gap, we first create TRANsCRIPT, a corpus of YouTube transcripts from transgender, cisgender, and non-binary speakers. Using this dataset, we probe various PLMs to assess if they encode the gender identity information, examining both frozen and fine-tuned representations as well as representations for inputs with author-specific words removed. Our findings reveal that PLM representations encode information for all gender identities but to different extents. The divergence is most pronounced for cis women and non-binary individuals, underscoring the critical need for gender-inclusive approaches to NLP systems.

Abstract

Offensive speech is highly prevalent on online platforms. Being trained on online data, Large Language Models (LLMs) display undesirable behaviors, such as generating harmful text or failing to recognize it. Despite these shortcomings, the models are becoming a part of our everyday lives by being used as tools for information search, content creation, writing assistance, and many more. Furthermore, the research explores using LLMs in applications with immense social risk, such as late-life companions and online content moderators. Despite the potential harms from LLMs in such applications, whether LLMs can reliably identify offensive speech and how they behave when they fail are open questions. This work addresses these questions by probing sixteen widely used LLMs and showing that most fail to identify (non-)offensive online language. Our experiments reveal undesirable behavior patterns in the context of offensive speech detection, such as erroneous response generation, over-reliance on profanity, and failure to recognize stereotypes. Our work highlights the need for extensive documentation of model reliability, particularly in terms of the ability to detect offensive language.

Abstract

Social media revolutionized political participation possibilities, and movements like Fridays for Future (FFF) rely on them. Understanding the factors driving the novel online modes of political participation is crucial. Consequently, this project explores the relations between diverse group identification components and social media political participation types in the context of FFF, drawing on the Social Identity Model of Pro-Environmental Action and extending previous work by treating group identification as multidimensional construct. Two German quota samples (N = 619, 45 \% men; N = 616, 44 \% men) completed online surveys assessing group identification components and social media political participation types applied to the FFF context. Generalized linear models were used to investigate their relations. Centrality – the extent to which individuals perceive their membership in FFF as an important aspect of their self-concept – was the strongest and most consistent predictor of Counter, Follower, and Expressive Engagement via social media in both samples. Additionally, each social media political participation type exhibited distinct connections with different group identification components. These results underscore the relevance of group identification in emerging forms of social media political participation. The diverse connections between distinct group identification components and participation types underscore the need to examine both considering their diverse forms.

Abstract

We investigated salivary biomarkers of stress, more specifically, cortisol and alpha-amylase, to evaluate effects of individualized music listening (IML) in people with dementia.

Abstract

Authorship Profiling (AP) aims to predict the demographic attributes (such as gender and age) of authors based on their writing styles. Ever-improving models mean that this task is gaining interest and application possibilities. However, with greater use also comes the risk that authors are misclassified more frequently, and it remains unclear to what extent the better models can capture the bias and who is affected by the models' mistakes. In this paper, we investigate three established datasets for AP as well as classical and neural classifiers for this task. Our analyses show that it is often possible to predict the demographic information of the authors based on textual features. However, some features learned by the models are specific to datasets. Moreover, models are prone to errors based on stereotypes associated with topical bias.

Abstract

The paper explores integrating cyber-physical systems in construction, targeting workforce shortages, productivity issues, and environmental concerns. It highlights the shift towards human-robot collaboration (HRC) to enhance productivity and safety. However, HRC implementation in prefabrication and construction encounters challenges like low productivity and worker lack of trust. To address these, the paper advocates for reflexive research methods to understand workers’ needs and promote inclusivity in HRC design. It examines the relevance of feminist technoscience perspectives (FTS) in addressing industry needs and fostering inclusivity in HRC design. Through an analysis of three case studies in HRC, the paper identifies research gaps and proposes strategies to enhance HRC design for prefabrication and construction. It emphasizes the significance of the FTS Framework in guiding these efforts. The FTS Framework assesses and informs HRC design for construction. It raises designers’ awareness of power dynamics, biases, and inclusivity issues. The chosen case studies illustrate collaborations among diverse agents, human augmentation in design and fabrication processes, and usability of HRC designs. The paper evaluates the case studies using the four perspectives of the FTS Framework: Embodiment, Reconfiguration, Interrelations, and Participation. These perspectives guide efforts to adapt HRC designs to workers’ needs, address on-site challenges, redistribute agency for inclusivity, and involve users in the design process. Leveraging the FTS Framework, the paper aims to address automation challenges in construction and facilitate the transition ofHRC from laboratory settings to industry applications.

Abstract

We describe the details of the Shared Task of the 5th ACL Workshop on Gender Bias in Natural Language Processing (GeBNLP 2024). The task uses dataset to investigate the quality of Machine Translation systems on a particular case of gender robustness. We report baseline results as well as the results of the first participants. The shared task will be permanently available in the Dynabench platform.

Abstract

In this paper, we revisit the seminal work of Garimella et al. 2019, who reported that dependency parsers learn demographically-related signals from their training data and perform differently on sentences authored by people of different genders. We re-run all the parsing experiments from Garimella et al. 2019 and find that their results are not reproducible. Additionally, the original patterns suggesting the presence of gender biases fail to generalize to other treebank and parsing architecture. Instead, our data analysis uncovers methodological shortcomings in the initial study that artificially introduced differences into female and male datasets during preprocessing. These disparities potentially compromised the validity of the original conclusions.

Abstract

Automating robotic assembly in architectural construction is challenging due to material uncertainties and the buildup of tolerance experienced in assembling many parts. Implementing AI technologies, including various machine learning algorithms in robotic assembly, has demonstrated significant potential for robots to respond to this uncertainty. This research builds upon previous implementations of singular algorithms by combining haptic teaching with deep reinforcement learning (DRL) in a single workflow to improve robot autonomy in responding to uncertainties in timber assembly. Haptic teaching bridges the gap between simulation and reality inherent to DRL, while DRL improves the generalizability of haptic teaching and speeds up the agents’ learning process. The developed workflow is tested through various lap joint assembly experiments. The effectiveness of this combined approach is assessed through various experiments that refine and evaluate the methodology, providing valuable insights into enhancing robot capabilities to manage material uncertainties and deviations. Additionally, this research considers the evolving role of human workers in collaborative construction environments with robots.