How do product teams evaluate LLM-powered products? As organizations integrate large language models (LLMs) into digital products, their unpredictable nature makes traditional evaluation approaches inadequate, yet little is known about how practitioners navigate this challenge. Through interviews with nineteen practitioners across diverse sectors, we identify ten evaluation practices spanning informal ‘vibe checks’ to organizational meta-work. Beyond confirming four documented challenges, we introduce a novel fifth we call the results-actionability gap, in which practitioners gather evaluation data but cannot translate findings into concrete improvements. Drawing on patterns from successful teams, we contribute strategies to bridge this gap, supporting practitioners’ formalization journey from ad-hoc interpretive practices (e.g., vibe checks) toward systematic evaluation. Our analysis suggests these interpretive practices are necessary adaptations to LLM characteristics rather than methodological failures. For HCI researchers, this presents a research opportunity to support practitioners in systematizing emerging practices rather than developing new evaluation frameworks.

Interdisciplinary teams developing complex technologies such as healthtech struggle to align disciplinary perspectives, stakeholder priorities, and evolving problem framings, particularly during rapid iteration, when existing collaboration tools offer limited support for in-session negotiation. We present KNIT, an AI-mediated framework that conceptualises AI-generated artefacts as computational boundary objects. KNIT supports convergence by externalising anonymised individual inputs into shared artefacts, including semantic clusters and stakeholder-centred problem reframings, that surface differences in interpretation and make them available for negotiation. We evaluated KNIT in workshops with seven early-stage healthtech teams (28 participants), analysing 190 interaction episodes using Carlile’s 3T framework. KNIT supported knowledge boundary crossing across syntactic (95.0%), semantic (86.3%), and pragmatic (84.8%) levels. We contribute empirical evidence and design principles showing how computational boundary objects mediate distinct boundary-crossing mechanisms, demonstrating that representational transformation rather than automation is the primary mechanism through which AI enables convergence across disciplinary boundaries.

Adoption of Fully Autonomous Vehicles (FAVs) depends on trust, which is defined as confidence in a vehicle's dependability, safety, and predictability.In cooperative driving scenarios, trust must exceed ego vehicles to include other autonomous vehicles and their coordination.This is challenged by unexpected multi-agent interactions, diminishing human control, and limited system transparency.We hypothesize that enhancing transparency by providing information about ego vehicle, other cooperative vehicles, and road conditions can foster trust.This is achieved by visualizing vehicleto-everything (V2X) information via augmented reality (AR) interfaces.To test this in a safe environment, we conducted a withinsubjects experiment in a Virtual Reality (VR) driving simulator with AR overlays.Participants experienced three interface concepts: (A) no transparency, (B) system-level transparency (ego vehicle intentions only), and (C) environment-level transparency (cooperation intentions, planned paths, and infrastructure).Results show that environment-level transparency, despite the higher cognitive workload, enhanced trust in both ego and cooperating FAVs.

Conversational agents (CAs) are rapidly advancing across industry and academia and it is crucial to consider the values embedded within these systems. Value-sensitive design practices have benefited AI-based systems, but have not yet been widely applied to CAs. This paper introduces the Value-Sensitive Conversational Agent (VSCA) Framework. The framework uses collaborative design (co-design) activities to guide CA creators and CA users to collaboratively create three key artefacts that elicit CA users’ values and are technically useful for CA creators to drive implementation forward, resulting in value embodied CA prototypes. The paper presents the practical framework and toolkit, followed by a mixed-method evaluation through design workshops, semi-structured interviews, and a comparative survey. Results show that the framework and toolkit increase CA creators’ value-sensitivity, empower CA users, enhance collaboration, and produce value-embodied prototypes. Based on this work, we offer 14 guidelines to practically support value sensitivity in CAs.

As AI systems become increasingly prevalent in critical domains, ensuring their alignment with stakeholders’ values is essential. However, recent studies have revealed deficiencies in socio-technical design processes and design activities for AI, particularly in eliciting diverse stakeholder values and integrating them into system design and development. To investigate these challenges empirically, we conducted 30 semi-structured interviews with AI practitioners. Our findings reveal several key challenges faced during AI design and development. Firstly, practitioners struggle with identifying and involving stakeholders due to uncertainties regarding user demo-graphics, and a lack of interdisciplinary expertise. Secondly, they encounter obstacles in integrating values into technologies, citing practical complexities, unclear responsibilities, and limited support. This paper presents a concept map detailing these four primary barriers and discusses potential strategies and recommendations for overcoming them. These strategies revolve around improving value elicitation practices, facilitating more meaningful stakeholder engagement, and understanding the impact of stakeholder values By qualitatively describing the barriers to integrating stakeholder values into AI systems, this study contributes to the emerging field of value-sensitive AI.

Abstract This study addresses the challenge of limited AI literacy among the general public hindering effective participation in AI codesign. We present a card-based AI literacy toolkit designed to inform nontechnical audiences about AI and stimulate idea generation. The toolkit incorporates 16 competencies from the AI Literacy conceptual framework and employs ‘What if?’ prompts to encourage questioning, mirroring designers’ approaches. Using a mixed methods approach, we assessed the impact of the toolkit. In a design task with nontechnical participants (N = 50), we observed a statistically significant improvement in critical feedback and breadth of AI-related questions after toolkit use. Further, a codesign workshop involving six participants, half without an AI background, revealed positive effects on collaboration between practitioners and end-users, fostering a shared vision and common ground. This research emphasizes the potential of AI literacy tools to enhance the involvement of nontechnical audiences in codesigning AI systems, contributing to more inclusive and informed participatory processes.

Value Sensitive Design (VSD) is a framework for integrating human values throughout the technology design process. In parallel, Responsible AI (RAI) advocates for the development of systems aligning with ethical values, such as fairness and transparency. In this study, we posit that a VSD approach is not only compatible, but also advantageous to the development of RAI toolkits. To empirically assess this hypothesis, we conducted four workshops involving 17 early-career AI researchers. Our aim was to establish links between VSD and RAI values while examining how existing toolkits incorporate VSD principles in their design. Our findings show that collaborative and educational design features within these toolkits, including illustrative examples and open-ended cues, facilitate an understanding of human and ethical values, and empower researchers to incorporate values into AI systems. Drawing on these insights, we formulated six design guidelines for integrating VSD values into the development of RAI toolkits.

This article introduces a method for conducting a fully online collaborative design workshop requiring no facilitation which we refer to as a Self-guided Collaborative Online Workshop (SCOW).The article provides three main contributions.Firstly, we present a process for the conversion of a faceto-face facilitated design workshop into a SCOW using a method we call the "playboard" which draws on concepts from CSCL literature.Secondly, we evaluate the efficacy of SCOWs using an iterative evaluation with 75 participants, including measures for participant satisfaction, subjective and objective learning outcomes, quality of the online and self-guided experience, and comparison with face-to-face workshops.Results across all measures indicate that the self-guided workshop was as successful as the in-person facilitated original.Moreover, participants reported advantages of the more scalable format including improved access to those with non-visible disabilities and in the Global South.Finally, based on our findings, we present a set of recommendations for others interested in using SCOWs as an inclusive and scalable way to support collaborative experiences.