Persuasion, pride and prejudice: factors associated with students' resistance to a social robot's incorrect mathematical solution.

This exploratory study investigates how secondary school students resist or align with a social robot that challenges their reasoning in a curriculum-based mathematics task. Twenty-two students collaborated with a Furhat robot on a geometry problem whose three parts were familiar from class. In the critical part, all students produced task-relevant evidence for the correct solution, while the robot proposed and defended an incorrect solution through argument-based persuasion, without deliberately invoking peripheral persuasive cues. We examined resistance to the robot's informational social influence in two ways. First, as an endpoint outcome, 14/22 students aligned with the robot: three immediately after dissent and eleven after a persuasion debate, while eight resisted. Second, as a process index, we introduce an approach to quantify early expressed resistance: students' responses within the first two persuasion debate exchanges were descriptively coherent with endpoint outcomes. Rather than testing causal or predictive relations among endpoint outcomes, early resistance, and questionnaire or interaction measures, we interpret observed patterns, prior theory, and interactional design to formulate seven hypotheses for future research: Students' resistance may be stronger with higher self-perceived competence (1), with behavioural patterns consistent with self-efficacy when cognitive demands are higher (2), and with stronger task-directed attention before conflict, suggesting that early engagement may provide descriptive markers of later resistance (3). Resistance may be weaker among students reporting more frequent conversational AI/LLM use and stronger preferences for human-like robot attributes (4). For the robot, incorrect arguments may appear more objective and credible when they are delivered through neutral wording, tone, and facial expressions without explicit uncertainty cues (5), and when the arguments, while incorrect, are heuristically plausible (6). Finally, students who aligned may be more vulnerable to later robot influence (7), since uncertainty and trust in the robot's information appeared to rise together during debate until alignment. These hypotheses suggest that calibrated robot behaviours (e.g., signalling uncertainty when the information is less reliable) may help sustain students' resilience and self-efficacy when they engage with persuasive yet fallible educational robots.