Eighteen Recommendations Advance GenAI Lab Studies, Addressing Five Key Challenges

Researchers are grappling with how to rigorously evaluate generative AI systems, given their inherent unpredictability. Hyerim Park (BMW Group and University of Stuttgart), Khanh Huynh (BMW Group and LMU Munich), and Malin Eiband (BMW Group), alongside et al., detail significant methodological hurdles in a new study examining user interaction with these technologies. Their work presents a reflective analysis of four lab-based user studies , encompassing in-car assistants and design tools , to identify five key challenges arising from generative AI’s stochastic nature, such as managing user trust and interpreting ambiguous outputs. Crucially, the team proposes eighteen practical recommendations, distilled into five guidelines, to help researchers design more robust and transparent evaluations of generative AI, ultimately enabling more comparable and trustworthy research in this rapidly evolving field.

Crucially, the team proposes eighteen practical recommendations, distilled into five guidelines, to help researchers design more robust and transparent evaluations of generative AI, ultimately enabling more comparable and trustworthy research in this rapidly evolving field.

GenAI Evaluation Challenges in HCI Labs require novel

Scientists have demonstrated a crucial step forward in evaluating generative artificial intelligence (GenAI) systems within controlled laboratory settings, addressing a significant challenge in human-computer interaction (HCI) research. The research establishes a framework of five guidelines and eighteen practice-oriented recommendations, offering researchers actionable strategies for conducting more transparent, robust, and comparable studies of GenAI systems. Unlike traditional rule-based systems, GenAI models generate varied outputs even with identical inputs, disrupting core assumptions of controlled lab studies such as consistency and comparability. This unpredictability affects all stages of user research, from task definition and prototype development to data collection and analysis, yet systematic exploration of how to navigate these complexities has been lacking.

The study unveils that simply measuring user-facing outcomes like task performance or satisfaction is insufficient; a deeper understanding of the evaluation process itself is critical for drawing reliable conclusions about GenAI usability and user experience. This detailed logging allows researchers to differentiate between issues stemming from the interface design and those inherent to the GenAI system itself, improving the accuracy of evaluation. Experiments show that adapting evaluation methods is not merely about acknowledging limitations, but about proactively designing studies that account for GenAI’s inherent variability. The guidelines offer practical recommendations, such as carefully considering the level of fidelity needed in prototypes, balancing control with the need for realistic GenAI behaviour, and developing methods for capturing and interpreting user feedback in the context of stochastic outputs. The research opens new avenues for evaluating adaptive and intelligent systems, ensuring that future HCI research can accurately assess the potential of these powerful technologies.

GenAI Evaluation Challenges and Practice Guidelines are crucial

The study employed a cross-case reflection and thematic analysis, scrutinising all phases of each study to pinpoint specific issues and formulate actionable solutions. This innovative approach enabled the team to move beyond simply identifying problems to developing concrete strategies for improved evaluation. Experiments began with participants interacting with the GenAI prototypes in controlled laboratory settings, performing pre-defined tasks designed to elicit specific behaviours and responses. Researchers carefully logged all user interactions, including inputs, system outputs, and observed behaviours, creating a detailed record of each session.

Crucially, the team also implemented system-level logging, capturing data on internal events such as hallucinations and latency, providing a granular view of the GenAI’s performance beyond the user interface. This dual-level logging, user actions and system behaviour, proved essential for disentangling interface issues from inherent GenAI variability. Participants completed questionnaires and engaged in think-aloud protocols, articulating their perceptions of the system’s trustworthiness and how well it understood their goals. To address the challenge of interpretive ambiguity, researchers developed a coding scheme to differentiate between problems stemming from the interface design and those originating from the GenAI’s stochastic outputs.

This involved multiple coders independently analysing session recordings and resolving discrepancies through discussion, ensuring inter-rater reliability. Furthermore, the study pioneered a reframing of onboarding procedures to proactively manage participant expectations regarding system unpredictability. Instead of presenting the GenAI as a consistently reliable tool, onboarding materials explicitly acknowledged its potential for variability, preparing users for unexpected outputs and encouraging them to view the system as a collaborative partner. This approach aimed to reduce frustration and improve the validity of the evaluation by minimising the impact of negative surprise.

GenAI Reliance Hinders Novel Interaction Exploration, stifling. The authors note a limitation in that the study focused on lab-based evaluations, and further research is needed to explore how these challenges manifest in field studies or with larger-scale deployments. Future work could investigate the application of these guidelines across diverse GenAI applications and user groups, ultimately advancing the field’s ability to rigorously assess and improve these increasingly prevalent technologies.