Attunement Costs: How Interaction Locks Users into Specific LLMs

Lieu : Salle 3.216 - Campus Grand Paris & en ligne

Horaire : 12h00–13h30

Intervenants : Joachim Henkel

Discutante : Soukaina Karama - Cloud Services Consultant, Fifty-Five

Bio : 

Prof. Henkel (*1965) is a professor of technology and innovation management. His research focuses on patent management, standards, technology acquisitions, digitalization, and open and user innovation. His research has been published in leading journals, such as Academy of Management Discoveries, the Harvard Business Review, the Rand Journal of Economics, Research Policy, and the Strategic Management Journal. He is an Associate Editor with Academy of Management Discoveries and serves on the editorial review boards of Industrial and Corporate Change, Innovation: Organization & Management, and Research Policy. He regularly consults firms in the ICT industries, with a focus on patent litigation in the context of standards. Prof. Henkel received a degree in physics from the University of Bonn, a Ph.D. in economics from the University of Mannheim, and was an assistant professor at Ludwig Maximilians University Munich. After his Ph.D., he worked at the consulting firm, Bain & Company for two years. He was a visiting scholar at University College London, MIT Sloan School of Management, Harvard Business School, and Singapore Management University. Prof. Henkel was appointed full professor at TUM School of Management in 2004

Abstract:

Large language models (LLMs) present near-identical conversational interfaces that impose minimal upfront learning costs. Traditional accounts of switching costs would therefore predict that users switch freely toward better-performing or cheaper alternatives. However, we find that while users experiment with alternative models, they remain attached to specific models, even when objectively superior or cheaper options are available. This pattern is not readily explained by status quo bias, habit, or inertia, which predict that users would not switch in the first place. We argue that this reflects a novel form of switching cost—attunement costs—rooted in the interaction paradigm of LLM-enabled software, which offers three affordances: natural language intent articulation, probabilistic high-dimensional generativity, and discursive output engagement. These affordances shift the responsibility for effective interaction from developers, who predefine interfaces to guide interaction in traditional software, to users, who must acquire interaction-based knowledge. This knowledge follows a distinctive learning trajectory and transfers imperfectly across LLMs. Because it is tacit and concealed behind near-identical user interfaces, its limited transferability cannot be anticipated ex ante. Attunement costs therefore manifest ex post, as degraded performance following a switch and can lead users to revert to the incumbent model. We empirically test a necessary condition of this theory by submitting 50 HumanEval coding tasks, each in four prompt variants reflecting relevant prompting techniques, to 46 LLMs (9,200 prompt–model combinations). We find substantial model-level heterogeneity: prompting techniques that improve one model's performance degrade another's, with little cross-model agreement on the optimal technique. These findings are consistent with the proposed model specificity of interaction-based knowledge. The theory extends affordance-based accounts of human-computer interaction, introduces attunement costs as a novel form of procedural switching cost, and offers a demand-side boundary condition for debates on LLM market structure and competition—a question of growing consequence as LLMs become embedded across an expanding range of private and professional technologies.

Pour toute information complémentaire, n’hésitez pas à contacter Margherita Pagani : margherita.pagani@skema.edu