Kelly Wen
Co-Founder, EdPilot
Access to one-on-one tutoring has always been one of the most powerful predictors of learning outcomes. It's also always been unevenly distributed. AI changes both of those facts simultaneously.
For most of the history of higher education, access to one-on-one academic support has been unevenly distributed in ways that mostly track with resources. Students at well-funded institutions have tutoring centers, abundant TA support, small discussion sections, and faculty who have time for them. Students at under-resourced institutions, or students who can't afford private tutoring, or students with work and family obligations that constrain when they can seek help — these students have always operated with a meaningful disadvantage.
AI-supported learning changes this. A student who can access AI that knows their course materials and can explain concepts in multiple ways, at any hour, doesn't have an advantage equivalent to private tutoring — but they have something meaningfully better than nothing, at a cost approaching nothing.
This is one of the genuinely significant things about the current moment. The access question is changing.
The educational research on one-on-one tutoring is fairly consistent: individual tutoring, done well, substantially improves learning outcomes compared to lecture-based instruction. The effect size in well-designed studies is large enough that researchers have spent decades trying to understand why and how to replicate it at scale.
The reasons tend to cluster around a few mechanisms. Tutors catch misunderstandings immediately rather than letting them compound. Tutors adjust explanation depth and approach to the individual student's responses. Tutors maintain productive engagement rather than letting attention drift. Tutors provide immediate feedback on practice.
AI that's well-designed for educational contexts can replicate some of these mechanisms. Not all — the relationship component of effective tutoring isn't easily replicated — but the feedback and adaptation dimensions are addressable.
The positive case for AI tutoring is real. So are the risks of implementing it poorly.
AI that's optimized for student satisfaction rather than learning tends to produce students who feel good about their understanding without developing it. AI that smooths all productive difficulty out of the learning process removes the struggle that produces durable knowledge. AI that's unconnected to course materials builds confidence grounded in potentially wrong answers.
The design choices matter enormously. AI that guides students toward answers through questions produces different outcomes than AI that provides answers. AI that requires students to demonstrate understanding before moving forward produces different outcomes than AI that accepts any response as engagement.
Getting these design choices right requires people who understand how learning works — which is primarily faculty.
If AI significantly improves the availability of foundational support — explanations, practice, conceptual clarification — the nature of what the classroom needs to do changes.
Class time that was previously devoted to explaining concepts students could get elsewhere becomes available for something else: application, discussion, analysis, synthesis, the things that require a room full of people engaging with ideas together.
This is already happening in fields that have experimented with flipped classroom models. The potential with AI-supported preparation is that students arrive with more robust foundational understanding, making the in-class time more productive.
The professors who adapt to this most effectively will be the ones who think clearly about what their class time is for and design accordingly — not assuming it needs to do the same things it did before, but asking what it can accomplish that AI-supported study can't.