A Rigorous Study of 2,379 Students Found AI Tutoring Made Grades Worse

By Hector Herrera | October 28, 2026 | Education

A randomized controlled trial gave college students access to a GPT-4o–based AI study assistant and then measured what happened to their final grades. The result: students who used the AI tool earned lower grades and participated less in instructor-led learning activities than peers who studied without it. The study — one of the most methodologically rigorous evaluations of AI tutoring tools in formal education yet published — lands as a direct challenge to the assumption driving billions of dollars of EdTech investment: that always-available AI assistance improves learning.

What the Study Found

The trial enrolled 2,379 college students, randomly assigning them to either a GPT-4o–powered AI study assistant or a control group with no AI tool access. Both groups received the same coursework, the same instructors, and the same assessments.

Key findings:

• Students in the AI group earned lower final grades than their control-group peers.
• AI-group students showed significantly lower participation in instructor-designed learning activities — office hours, study groups, discussion sections.
• The performance gap was not explained by prior academic ability, subject area, or level of AI tool usage. Even students who used the assistant heavily underperformed.

The researchers' interpretation centers on a concept called productive struggle — the process of working through difficult material independently before getting help. That friction is not a bug in learning; decades of cognitive science research suggest it is central to how humans consolidate knowledge into long-term memory. When an AI tool removes that friction instantly, learners may get the answer without building the mental architecture that lets them recall or apply it later.

Why This Is Significant

This is not a think piece or a survey of student preferences. It is a randomized controlled trial — the gold standard of evidence in research — with nearly 2,400 participants. The study design rules out the most obvious confounding explanations (students who chose to use AI were already struggling, or the AI was bad at explaining things).

The result is significant because it runs directly against the commercial narrative. EdTech platforms selling AI tutoring tools to schools, universities, and parents are doing so largely on the logic of access and personalization: every student can have a patient, always-available tutor. That logic is intuitive. This study says the evidence does not support it — at least not in its current form.

The broader pattern emerging from 2026 AI education research:

• A Digital Education Council survey released in September found 88% of students use AI for coursework, but faculty report comprehension gaps widening — not closing.
• Multiple universities running AI pilot programs have reported the same phenomenon: usage is high, measured learning outcomes are flat or declining.
• The gap between what students believe AI tutoring does for them and what assessment data shows is consistent and statistically significant across multiple studies now.

None of this means AI has no place in education. It means the current dominant deployment model — AI as on-demand answer machine — may be optimizing for the wrong outcome.

What This Means for Schools and Administrators

For school districts and universities evaluating or expanding AI tutoring deployments, this study should trigger a formal evidence review. The questions to ask vendors are no longer just about privacy and accuracy — they are about learning outcomes.

Practical implications:

• Outcome accountability clauses. Contracts for AI tutoring tools should include measurable learning outcome benchmarks, not just engagement metrics. Usage data tells you students are using the tool. Grade data tells you if it's working.
• Friction-preserving design. Some EdTech companies are already experimenting with AI tutors that withhold direct answers and instead prompt students through Socratic questioning. That design philosophy aligns better with what cognitive science says about productive struggle. It also tends to produce lower engagement scores — which is why it has not dominated the market.
• Differentiated access. The blunt policy choices — ban AI entirely, or allow it unrestricted — both have costs. The research suggests a middle path: AI assistance that scaffolds without replacing the cognitive work may preserve the learning benefits while reducing the harm.

What to Watch

The researchers behind this trial are expected to publish follow-up data on whether the grade gap persists into subsequent semesters, and whether students who used AI heavily show differential retention on cumulative assessments. That longitudinal data will be more consequential than any single endpoint grade. If AI tutoring use correlates with worse performance over a full academic year — or across students' transcripts — it will force a genuine reckoning with whether the EdTech industry's AI bet is producing what it promises.