Guided AI dialogue was followed by more developed reading questions, with task order limiting causal interpretation

An educational assistant can be designed to help a reader ask a better question instead of simply supplying an answer. Xiaoyu Zhao and colleagues studied this possibility through a guided dialogue mode built around ERNIE Bot. After reading, students proposed questions for classmates, and researchers examined how numerous, clear and cognitively demanding those questions were. This places a learner-generated artifact at the center of evaluation, while raising an important design question: does improved questioning during support carry over when the support is absent?

The instructional quasi-experiment involved 112 sophomore students at a university in central China. Each student first read one article in a general digital-reading mode and posted questions. After a break, the student read a second article and engaged in AI-supported dialogue before completing the same kind of question-generation task. Experts selected texts considered comparable in length, relevance and conceptual demand. Nevertheless, the ordinary reading always preceded the AI-supported reading, so mode, text and order were not independently varied.

The agent was designed to encourage inquiry through examples, follow-up prompts, counter-questions and evaluative feedback, rather than simply return direct answers. Learners retained some control over how they responded or redirected the conversation. Researchers collected 1,234 student-generated questions and 579 dialogue segments, then coded the questions for cognitive level and features such as clarity, relevance and logical structure. They also used sequence analysis to examine how learners combined dialogue strategies. These measures describe questioning performance and observed interaction, not the full internal process of comprehension.

Students produced an average of 5.90 questions after the AI-supported phase, compared with 5.13 after general reading. Rated overall question quality was also higher, with reported means of 5.86 and 5.18 on the study's eight-point scale. Higher-performing questioners showed different patterns of reflection, evaluation and adjustment in their dialogue. The findings suggest that a scaffolded conversation can be a useful setting for practicing inquiry. However, associations between dialogue patterns and question quality do not establish that a particular strategy independently caused the difference.

The design warrants a restrained conclusion. Repeating the task may help students understand what is expected, and the second text may elicit different questions even if experts regard the texts as comparable. There was no counterbalanced sequence to separate these influences. The short activity also did not test autonomous transfer, stable metacognitive regulation or later questioning without AI. Coding reliability supports consistency of the ratings, but cannot resolve those design limitations. The result is evidence of stronger observed performance in the supported phase, not proof of a durable increase in a general questioning ability.

For AIEDHK, a useful next study would reverse the order for some learners and include a third, unfamiliar text read without assistance. Teaching could ask students to identify what a question assumes, explain why it matters and connect it to a specific passage. Teachers could assess the quality of those justifications rather than reward the number of questions alone. These are proposed extensions beyond the paper. The practical opportunity is to make inquiry visible and teachable, while checking whether the learner can continue to formulate worthwhile questions after the conversational support has been removed.