Improving Hybrid Human-AI Tutoring by Differentiating Human Tutor Roles Based on Student Needs

Synthesis: In a large-scale quasi-experiment with 635 students (grades 5-8), hybrid human-AI tutoring produced substantial gains over AI-only tutoring: +25% time on task, +36% skill proficiency, and +61% standardized academic growth. The study's core innovation was a differentiated tutoring policy: students below the grade median received proactive human-initiated support, while those above received reactive, on-demand support. Proactive tutoring showed marginally higher growth (+75%, p = .065) and was particularly beneficial for students farthest below the cutoff, helping narrow achievement gaps.

Connection to Broader AIED Research

This study extends the taxonomy in The Evidence Base on AI in K-12: A 2026 Review by showing that the human component remains differentiable and valuable even alongside AI. It also connects to Personalized Learning frameworks and speaks to Equity concerns about achievement gaps.

What this means for practice

• Administrators. Allocate human tutor time by need instead of evenly. The 635 consented students in grades 5-8 below their within-grade median received proactive, tutor-initiated support and showed 75% greater MAP growth than the reactively supported group (p = .065) — the same tutor hours, aimed where the marginal benefit was largest.
• Administrators. Fund the human layer rather than AI alone: students gained an additional 61% in standardized MAP growth (p = .003) in the human-AI period compared with the AI-only baseline, about 2× the expected national NWEA growth norms.
• Instructors. Reserve proactive outreach for students below the cutoff and leave higher performers on reactive, on-demand support: time on task (+25%) and skill proficiency (+36%) improved significantly with human-AI tutoring overall but did not differ significantly between the proactive and reactive groups.
• Administrators. Re-examine the cutoff each term rather than fixing it. Dividing students into proactive and reactive groups at the median may oversimplify the relationship between need and optimal tutoring intensity.
• Researchers. Compare the differentiated policy as a whole against a business-as-usual control condition; the authors identify that randomized design as the stronger causal test.

Limitations

• Quasi-experimental evidence from one school. The findings come from a difference-in-discontinuities design in a single middle school in a Mid-Atlantic U.S. state during the 2024-2025 school year, with 635 of approximately 1,000 eligible students opting in through informed consent.
• Part of the sample lacks prior scores. Prior-year state test scores were available for 557 of the 635 students; the remaining 78 were newly enrolled, so the cutoff placement rested on less information for them.
• The headline differentiation effect is marginal, and null at the cutoff. Proactive tutoring's advantage over reactive tutoring was 75% on average but at p = .065, and the full-sample trend of a 26% improvement was likewise not significant (p = .086).
• A dichotomous median split. Assigning support at the median may oversimplify the relationship between student need and optimal tutoring intensity, as the authors state — the students farthest below the cutoff are precisely the ones the design cannot distinguish.

Citation

Gurung, A., Gao, G., Gutterman, J., Thomas, D. R., Gupta, S., Branstetter, L., Brunskill, E., Aleven, V., & Koedinger, K. R. (2026). Improving hybrid human-AI tutoring by differentiating human tutor roles based on student needs. Proceedings of the 19th International Conference on Educational Data Mining (EDM'26).