“How May I Help You?” A Multidimensional Analysis of GPT-Based Customer-Service Chatbots

This study examines how GPT-based customer-service chatbot language is evaluated when chatbot style is systematically constructed through Biber’s Multidimensional Analysis (MDA). Ten chatbot conditions were developed, representing opposing poles of five MDA continua: involved versus informational discourse, narrative versus non-narrative discourse, explicit versus situation-dependent discourse, persuasive versus non-persuasive discourse, and abstract/impersonal versus concrete/non-abstract discourse. The chatbot outputs were iteratively validated through MDA-based linguistic analysis before being used in simulated customer-service helping situations. A total of 120 Saudi university students interacted with all ten conditions and completed a 19-item post-interaction evaluation questionnaire after each interaction. Participant-level composite scores were analyzed using a one-way repeated-measures ANOVA and five planned paired-samples comparisons with Holm adjustment. Chatbot condition had a significant overall effect on evaluation, F(9, 1071) = 1420.82, p < 0.001, partial η2 = 0.923. All five planned contrasts were significant after Holm correction (all ps < 0.001), with 95% confidence intervals excluding zero and large standardized paired effects (dz = 3.75–5.65). Informational, non-narrative, explicit, non-persuasive, and concrete/non-abstract discourse received higher evaluations than their contrasting conditions. Dimension 3 showed the largest observed contrast, with explicit discourse receiving the highest evaluation. The findings demonstrate how MDA can function not only as a descriptive framework for register analysis but also as a prospective framework for constructing, validating, and evaluating measurable linguistic variation in generative-AI customer-service discourse.