June 29, 2026

Should AI Ask First? How Proactive Mentoring Shapes Learner–AI Interaction in Self-Directed Learning 

A major promise of AI in education is that every learner could get something closer to one-to-one tutoring: explanations, hints, prompts, and feedback on demand. But the paper “Should AI Ask First? Investigating the Effects of Proactive vs Reactive AI Mentoring in Self-Directed Learning” points to a persistent problem: simply making an AI tutor available does not mean learners will use it. Many learners underuse help systems and do not know when help would be useful. This motivates the paper’s central question: should an AI mentor wait for learners to ask for help, or should it proactively step in? 

The Study

The researchers conducted a between-subjects experiment with 81 participants who completed a self-directed learning task based on a 14-minute instructional video about Bayes’ theorem. Participants were randomly assigned to one of two AI mentoring conditions. 

In the reactive condition, the AI mentor provided support only when learners explicitly requested help. 

In the proactive condition, the AI mentor initiated short pedagogical prompts during the video. These prompts were designed to be context-aware and minimally disruptive. They appeared at natural task boundaries, such as the beginning of new video chapters, and were adapted using lightweight personalization based on pre-test performance and learner behaviour. 

The proactive mentor also used a cooldown mechanism to avoid overwhelming learners. If a learner had recently interacted with the AI, the system waited before initiating another prompt. Learner actions such as pausing or rewinding also triggers context-specific support at a later point in the video. 

Participants completed a pre-test, watched the video with access to the AI mentor, completed a post-test, and then answered a final questionnaire about perceived control, choice, usability, workload, and learning experience. 

Key Findings 

Interactions
Proactive mentoring strongly increased interaction. Learners in the proactive condition exchanged an average of 10.51 messages with the AI mentor, compared with 2.89 messages in the reactive condition. 
It also reduced non-interaction. In the reactive condition, 31.8% of learners never interacted with the AI mentor, compared with only 8.1% in the proactive condition. This means learners were almost four times more likely to interact when the AI initiated support. 

Proactive mentoring also changed the structure of learner–AI conversations. Learners in the proactive condition engaged in deeper conversations, discussed more distinct topics, accepted more help, attempted more answers, and contributed more useful input. Their communication was also less likely to be off-topic. 

In contrast, learners in the reactive condition more often used the AI for isolated questions, summaries, reformulations, or direct requests. Their exchanges tended to be shorter and less sustained. 

A clustering analysis identified several learner interaction profiles. One particularly important profile, the help-accepting learner pattern, appeared only in the proactive condition. This suggests that proactive AI support can scaffold learners into forms of engagement they may not initiate on their own.
 

Learning Outcomes and Learner Agency 
Despite these strong interaction effects, proactive mentoring did not significantly improve short-term learning gains. Both groups improved from pre-test to post-test, but the difference between conditions was not statistically significant. The paper therefore shows that initiative policy can strongly shape how learners interact with AI without necessarily producing immediate measurable learning gains in a single session. 

The study also highlights an important trade-off. Learners in the reactive condition reported greater perceived choice and control over their interactions. Proactive mentoring increased engagement, but it also reduced learners’ sense that they controlled when the AI entered the learning process. 

This finding points to a core design challenge for educational AI: systems should support learners who need help, but they must also preserve learner autonomy. 

Takeaway

The study shows that AI mentoring is not only about what an AI system can explain, but also about who takes the initiative, when support is offered, and how learners experience control. 

Proactive AI mentoring can reduce passive behaviour, increase learner engagement, and encourage more sustained help-accepting dialogue. However, more interaction does not automatically lead to higher short-term learning gains, and proactive support may come at the cost of perceived autonomy. 

The researchers argue that initiative should be treated as a first-class design dimension in AI-supported learning. Future AI mentors should move beyond fixed proactive or reactive strategies toward adaptive, mixed-initiative approaches that offer support when learners need it while still allowing them to preserve control over the learning process. 

References

• Otmani, K., Bodonhelyi, A., Bühler, B., Kasneci, E. (2026). Should AI Ask First? Investigating the Effects of Proactive vs Reactive AI Mentoring in Self-directed Learning. In Artificial Intelligence in Education. AIED 2026. Lecture Notes in Computer Science( ), https://doi.org/10.1007/978-3-032-29760-0_53