Human tutoring, generative AI tutoring and independent internet search: A randomised between-­subjects experiment on students' responses to academic challenge

Synthesis: Ennion and McLellan randomized students at an English sixth-form college to a human tutor, Khan Academy's GenAI tutor (Khanmigo Lite), or unguided internet search on mathematics problems set above the A-level curriculum. Human tutoring produced the calmest experience: the lowest perceived difficulty in both sections, the most attempts, and the fewest stressed students. The GenAI tutor did not make the work feel easier, but its students reported more effort than human-tutored students by the second section and, descriptively, less stress than internet searchers. The authors read this as a distinct, partial form of co-regulation useful in independent practice, not a replacement for teaching; no learning measure was taken.

Key Findings

1. Human tutoring made hard problems feel manageable. Students with the tutor reported the lowest perceived difficulty in both sections, and the gap widened in the second.
2. Human tutoring kept students starting problems. They attempted significantly more questions than the GenAI group in the first section and the internet group in the second.
3. The GenAI tutor sustained effort without easing difficulty. By the second section its students reported significantly higher effort than human-tutored students, alongside higher perceived difficulty.
4. Open web search was the most stressful route. Stress ran at 32.0% in the internet group, against 19.2% under the GenAI tutor and 5.8% under human tutoring.
5. Support type did not change appetite for harder problems. Behavioral challenge-seeking did not differ by condition and tracked pre-existing dispositions instead.
6. Effort is not evidence of learning. The study measured engagement, not understanding; higher effort can read as persistence or as unresolved difficulty.

How the experiment was run

160 volunteers aged 16–18 took part at a large sixth-form college in England; 11 were excluded for technical disruptions, leaving 149 (52 human, 50 GenAI, 47 internet). A random-number generator assigned each student one support. All followed a 90-min session with two 20-min sections of four problems, validated as exceeding the A-level curriculum. Effort and difficulty were single 5-point items after each section, attempts were counted out of four, stress was asked once at the end, and a final task offered six problems as a challenge-seeking measure.

Three kinds of help, three patterns

The human condition had one mathematics tutor in the room, guiding on request but withholding answers. The GenAI condition used Khanmigo Lite, a GPT-4-based tutor reached through a paid ChatGPT account and built to guide reasoning with hints. The internet condition used DuckDuckGo without AI features.

Difficulty separated the groups twice. Internet students rated the first section significantly harder than human-tutored students (p = 0.002); in the second, both the GenAI tutor (p = 0.007) and internet search (p < 0.001) were rated harder than human tutoring.

Effort separated them only later: no difference in the first section (p = 0.139), but significantly higher effort under the GenAI tutor than human tutoring in the second (p = 0.025). Attempts ran the other way, with the human group highest in both sections, more than the GenAI group in the first (p = 0.024) and the internet group in the second (p = 0.048).

Stress followed the same order: calm most under human tutoring, least under internet search.

A distinct and partial form of support

The authors place the conditions on a continuum of scaffolding and co-regulation. Human tutoring looked like both together: the best difficulty profile, the most attempts, the calmest responses, and lower effort that may mean less exertion was needed, not less engagement. The GenAI tutor sat between that and internet search, where students managed difficulty and affect alone.

Higher effort alongside higher difficulty is ambiguous: it could be productive struggle sustained by responsive hints, or students working harder because difficulty never got resolved. With no learning measure or interaction logs, scaffolding and co-regulation remain interpretive lenses, and the preregistered prediction of greater effort, attempts and challenge-seeking under GenAI support was not supported.

Why challenge-seeking barely moved

Behavioral challenge-seeking did not differ by condition (p = 0.225) but tracked trait challenge-seeking strongly (mean rank 89.90 against 49.07, r = 0.50). The form of help shaped how students experienced difficulty, not how much they sought next; one session will not shift a disposition built over years.

What this means for practice

• Instructors: treat support as task design, not convenience. Human tutoring gave the calmest, steadiest experience, so keep a person available when students must get through difficulty.
• Instructors: coach help-seeking explicitly. Students were told to ask for guidance, not answers, but not trained to do it, so specify what to request and what to refuse.
• Instructional designers: expect an AI tutor to raise effort and difficulty together, and add a resolution step, such as a worked solution or teacher check, so effort is not stranded.
• Administrators: position GenAI tutors as a replacement for unguided search, not for staffing. Stress ran at 32.0% in the internet group against 19.2% under the GenAI tutor.

Limitations

• One 90-min session in a single high-attaining, digitally confident sixth-form college, with self-selecting volunteers and 65.3% studying Mathematics, so findings may not generalize.
• Effort, difficulty and stress were single-item self-reports, stress was asked once, and problems sat above curriculum level.
• Attempts counted questions started, not accuracy, so a higher count could mean persistence or floundering; one tutor's style, the novelty of the GenAI tutor and variable web quality also differ across conditions.
• Khanmigo Lite's model snapshot was unavailable and may have changed between November 2024 and July 2025, so results may not generalize to other versions or to GenAI tutors as a category.
• The study departed from its preregistration: internet search replaced the planned no-support condition, and the exact analyses were not prespecified, so section-specific findings are exploratory.

Connected Concepts

• Scaffolding: contingent guidance that keeps a task within reach
• Productive Failure: difficulty that stays workable long enough to teach
• Help-Seeking: asking for a hint rather than an answer
• Self-Regulated Learning: managing effort and affect during independent work
• Intelligent Tutoring: what a GenAI tutor shares with human tutoring, and lacks
• Anxiety and Stress: affect while encountering difficulty

Connected Articles

• One click away: AI tutoring with Khanmigo in a two-year school experiment: a two-year Khanmigo trial where engagement bounded the effect
• StudentBench: AI and human tutoring yield equivalent GRE learning gains: AI and human tutoring compared on learning gains rather than engagement
• AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting: a purpose-built AI tutor that beat in-class active learning
• ChatGPT-generated help produces learning gains equivalent to human tutor-authored help on mathematics skills: ChatGPT hints matching human-authored hints on mathematics learning
• The effortless trap: Productive struggle, AI, and the illusion of learning: where AI sits relative to the cognitive work of learning

Citation

Ennion, M., & McLellan, R. (2026). Human tutoring, generative AI tutoring and independent internet search: A randomised between-­subjects experiment on students' responses to academic challenge. British Journal of Educational Technology. https://doi.org/10.1111/bjet.70100