Human Tutoring, Generative AI Tutoring, and Independent Internet Search: A Randomised Between-Subjects Experiment on Students’ Responses to Academic Challenge

Generative artificial intelligence (AI) is increasingly available as a source of interactive academic support, raising questions about how it shapes learning. Research has focused on performance outcomes, with less attention to learners’ behavioural and affective responses to academic challenge, even though productive engagement with challenge supports deeper learning. In this randomised between-subjects experiment, students aged 16–18 at an English sixth-form college worked on two sections of challenging mathematics questions with one of three forms of support: one-to-one human tutoring, a GenAI tutor or unguided internet search. Measures included self-reported effort and perceived difficulty, questions attempted, post-task stress and subsequent challenge-seeking. Support condition was associated with different response patterns. Students using the GenAI tutor reported higher effort and perceived difficulty than those receiving human tutoring by the second section. Descriptively, post-task responses showed a lower proportion of stressed responses under GenAI tutoring than internet search. Human tutoring showed the lowest perceived difficulty and highest number of attempts in both sections, although significant pairwise differences varied by section. It also showed descriptively the lowest proportion of stressed responses. Challenge-seeking did not differ by condition and was instead strongly associated with trait challenge-seeking. The findings suggest that the GenAI tutor used in this study offered a form of support distinct from both human tutoring and unguided internet search, with potential value during independent practice rather than as a replacement for human teaching. They demonstrate the importance of evaluating AI-supported learning beyond performance outcomes and through behavioural and affective responses to difficulty.