Programmable Expressiveness in Non-Social Tasks: A Mixed-Methods Study of Middle School AI Learning

Abstract

Background. Programmable expressive features—such as speech, facial expressions, and chatbot-style dialogue—are often promoted as tools to enhance engagement in educational robotics. While prior research shows benefits in socially-oriented tasks like storytelling or group collaboration, it remains unclear how student-controlled expressive blocks affect learning when the task itself is non-social. This study isolates the impact of such features in a context where expressiveness is not instructionally required. Method. We conducted a controlled, two-cohort study with 41 middle school students (ages 10–12) during a one-day AI-and-robotics workshop using the Doodlebot platform. Students in the experimental group had access to optional blocks enabling the robot to speak, emote, and use GPT-based responses. These features were hidden from the control group. All participants completed identical programming tasks (e.g., maze navigation, visual classification) that did not require social interaction. Data sources included pre/post surveys, facilitator notes, and student code. We applied the Mann–Whitney U test [1, 2] and reflexive thematic analysis [3, 4] to examine outcomes. Results. The expressive condition showed no significant gains in programming confidence or peer trust, but performed significantly worse on the post-workshop concept quiz (p = .007, r = .41). Qualitative data revealed that students in this group often used expressive blocks for entertainment rather than learning, leading to distraction, off-task behavior, and increased reliance on adult facilitation. Contributions. This study contributes (i) empirical evidence on the limitations of robot expressiveness in non-social learning contexts, (ii) a mixed-methods protocol for analyzing classroom robot deployments, and (iii) design guidance for aligning robot behavior with pedagogical intent. Implications. Expressiveness in educational robots should be contextually deployed—not assumed beneficial by default. In technical, goal-driven tasks that do not involve social reasoning, unscaffolded expressiveness may introduce cognitive overhead or divert attention. We propose a “dial-a-sociality” model, where robot behavior can be flexibly tuned to match the demands of the learning environment.