Learning on Credit: How AI Postpones the Help Students Need

Executive summary

Personal instruction has always been a vital part of education. Office hours, tutoring, after-school help, and one-on-one instruction have unlocked potential in countless students, enabling them to achieve things they never could have achieved on their own. Seeking out that help, and the effort it demands, is part of how students learn.

Generative AI has now changed how students pursue success. It has not reduced how much help they need; it has postponed that need until the damage is harder to undo.

I have tutored college chemistry in one-on-one and group settings since 2021, and I built a tutoring company around understanding how students learn. Before generative AI, struggling students reached us in the first one to three weeks of a semester. They carried shaky, pattern-based understanding that was already starting to fail, because it skipped the point of learning: understanding why, not just what. Fixing that was straightforward. Now AI keeps those shortcuts working much deeper into the semester, until the material finally outpaces them and their understanding collapses due to the missing foundation. The damage follows students into later courses and into any situation that asks them to apply concepts they never learned or reason through problems they were never taught to think through.

In effect, students are learning on credit. They borrow understanding they have not earned, and AI has raised their credit limit. The bill still comes due, only later, larger, and with interest.

This is not a story about AI alone. The cracks were always there: courses that reward memorized patterns over conceptual understanding, support that answers questions with those same patterns, and instruction that many capable students cannot follow as delivered. AI exposes and widens those cracks by making the shortcuts cheaper, faster, and more convincing.

The core problem is incentives. The incentive itself has not changed: grades reward correct answers. What has changed is the path to them. What once took time, effort, and real thinking is now nearly free and instantly available. A right answer reached the wrong way still earns the grade, even if it causes problems later. Institutions that redesign assessments, connect proper one-on-one instruction to their own curriculum, and train the people who teach students one-on-one can make real understanding the easier path. Those that do will produce stronger graduates and gain a genuine advantage in attracting students. And because the damage stays hidden in homework grades until exams and later courses, institutions that wait will discover it in their outcomes rather than in time to prevent it.

Key takeaways

AI has not reduced how much help students need. It has postponed that need until the damage is harder to undo.
Students are learning on credit: they borrow understanding they have not earned, and the bill comes due later, larger, and with interest.
The cracks predate AI. Courses reward memorized patterns, support answers with the same patterns, and many capable students cannot follow instruction as delivered.
The core problem is incentives. A right answer reached the wrong way still earns the grade.
Institutions that redesign assessments, connect one-on-one instruction to their curriculum, and train the people who teach one-on-one can make real understanding the easier path.

1. What more than 7,000 hours of tutoring revealed

Four lessons from tutoring shaped everything that follows: confusion is predictable, students take the fastest path available, teaching why matters more than teaching what, and live interaction is what makes it all work.

Confusion is predictable

When I started tutoring, I answered whatever questions students brought, as a traditional tutor would. Within a few semesters, it was clear that nearly everyone was confused about the same concepts, in the same places, for the same reasons. The same confusion appeared across different institutions and different levels of the subject, from AP Chemistry to college organic chemistry.

Increasingly, students at both levels also share the same habits. Many now arrive at college having already spent years using AI to get through high school coursework.

Upon recognizing this pattern, I changed my approach. I began opening sessions with a short 15 to 20 minute explanation of the concepts behind each topic. These explanations differed from what students heard in lecture, because they focused on simple explanations, analogies, and understanding why things happen rather than what to do to get the right answer. Students’ questions nearly disappeared, because the explanations had gotten ahead of the confusion. Over the following years, those explanations grew into a complete curriculum, which became the foundation of how Orango teaches, and most sessions became a full hour of targeted instruction on whatever unit a student was in.

The lesson was not that these students lacked ability. They could understand the material. They could not understand it the way it had been taught to them, and they needed a different kind of instruction. This matches decades of chemistry education research, which has documented the same predictable misconceptions across topics from bonding to equilibrium, often persisting into upper-level college courses (Kind; Jenkins and Shoopman). Chemistry is our example, but the steady demand for tutoring across every difficult subject suggests the problem is far wider.

Students take the fastest path

Since 2021, I have worked with more than 500 students, from different backgrounds and on different career paths. Nearly all of them shared one tendency: they looked for the quickest, easiest way to solve each type of problem.

You can hear it in one of the most common complaints in higher education: “That exam was so unfair. It was nothing like the practice exam!” Anyone who has taught has heard it. But should the real exam look like the practice exam? If the goal is to test understanding, it cannot. A good exam should look unfamiliar while testing the same concepts. That is the only way to tell whether students understand the material or have simply learned patterns that fail the moment something new appears. Imagine a surgeon who finds an abnormality and suddenly has no idea what to do.

The complaint is universal because the tendency is universal. When the reward goes to patterns and ease, students learn patterns, and applying knowledge to anything new becomes nearly impossible. As AI makes patterns available on demand, institutions that want to keep producing elite graduates will have to test application in new and deliberate ways, or the ability to think critically will erode.

Teach the why, not just the what

The first part of the solution is what instruction emphasizes. Much has been made of teaching to learning styles, such as visual or auditory learners. Research has not supported the idea that matching instruction to a preferred style improves learning (Pashler et al.). Teaching emphasis gets far less attention, and in our experience it matters much more.

Students do differ in how they approach problems. Within 15 to 30 minutes of working with someone, a trained tutor can usually tell whether they are a natural conceptual thinker, who needs to know why everything happens, or a strong pattern recognizer, who can solve difficult problems but cannot explain how. Knowing that helps the tutor decide where to start and how to explain.

But every one of those students benefited from the same emphasis: learning why things happen instead of simply what to do. Once a conceptual foundation is built on why, application becomes much easier, and “that was nothing like the practice exam” stops being a problem. The student has learned to take an abstract idea and apply it to a situation they have never seen. That is what education has always been about.

Interaction is the active ingredient

The second part of the solution is live, personal instruction. As my tutoring company grew, I recorded video lessons so students could learn on their own. Many watched a video, did not understand it, and then came to a session where I taught the same lesson almost word for word. In person, they understood it.

The content was identical. The difference was the interaction. In a live session, a tutor can see a student’s face and notice confusion they have not voiced. They can ask a question, hear how the student is thinking, and adjust before the misunderstanding sets in. In effect, a good tutor can solve problems before the student knows they have them.

The research on tutoring points the same way. A meta-analysis of randomized tutoring experiments found consistently large positive effects on learning, and larger effects when tutors were teachers or trained paraprofessionals rather than volunteers or parents (Nickow, Oreopoulos, and Quan). Those studies cover K-12 students, not college, but the mechanism is the same: a skilled person responding to how a particular student thinks.

2. What AI changed: learning on credit

AI has not removed students’ need for instruction. It has moved the moment they discover that need from the first weeks of a semester to the middle of it, or later.

Before AI: a low credit limit

For years, struggling students followed a predictable path. Within the first one to three weeks, they came to us saying they did not really know how to do anything. They had picked up a few problem-solving methods, often from online videos or poor tutoring, that worked on early problems. As concepts began to stack, that shaky foundation could not hold the weight.

Every one of those methods was pattern recognition. Students knew which steps to follow for a familiar-looking problem, but not why the steps worked or how they connected to the underlying concept. They were borrowing understanding they did not have, and the debt came due quickly. Once we rebuilt the conceptual foundation, everything that had seemed arbitrary made sense.

After AI: a much higher limit

Generative AI supplies an endless stream of increasingly advanced patterns. Ask a chatbot for a way to solve a problem and it will produce one. When that method stops working, it will produce another.

In our experience, the result is that students can keep borrowing far longer. Where they once reached us in weeks one to three, many now reach us around weeks six to eight, and some not until just before the final exam. That is when the material becomes advanced enough that AI-supplied patterns no longer carry them, and they try to apply a conceptual foundation that does not exist to difficult conceptual problems.

High school students show the same pattern, often earlier, because they have had these tools for a larger share of their education. In a Pew Research Center survey conducted in fall 2025, 54 percent of U.S. teens ages 13 to 17 said they had used AI chatbots for help with schoolwork, and one in ten said they do all or most of their schoolwork with a chatbot’s help (Pew Research Center, 2026). A year earlier, 26 percent had used ChatGPT for schoolwork, double the share in 2023 (Pew Research Center, 2025).

These timeframes are what our tutors have observed across our own students, not a controlled measurement. However, the pattern is consistent with controlled research. In a randomized study of high school math students, those given unrestricted GPT-4 access improved their practice performance. When access was removed, they performed worse than students who never had it. A version designed to guide students rather than hand them answers largely avoided that harm (Bastani et al., PNAS, 2025).

The bill comes due, with interest

By week six or eight, fixing the problem is no longer a matter of filling a gap. A student has spent weeks building methods on the wrong foundation, and those habits have to be untangled before real understanding can be built.

Think of someone who never learned the streets of their own city. They can drive to the store by memorizing the same turns they always take. Start them from the other side of town and they are lost. That is the position almost every one of these students ends up in.

The cost also does not end with the course. Later courses build on the same concepts, and admissions exams such as the MCAT and DAT test exactly the conceptual reasoning the shortcut skipped.

In the short run, it looks as if students need less help. They do not. They need the same amount, but they start later, when it is harder to give.

3. Two ways courses fail

AI creates a different problem depending on what a course rewards, and most courses fall into one of two groups.

Courses that reward patterns

Many courses still let students succeed by memorizing procedures and recognizing problem types. Before AI, that was already a weakness, but it had a built-in check. Problems quickly became difficult enough that students who could not work out the patterns on their own had to read the textbook, go to office hours, or find a tutor. The effort of putting the pieces together is itself what produced understanding.

AI removes that check. A student can now get a working pattern for nearly any problem type on demand. It is now possible to get through most or all of such a course without understanding how or why anything works.

Courses that demand understanding

Other instructors design courses that require students to explain mechanisms and reason about why things happen. These courses resist shortcuts better, but AI creates a different risk here.

Chatbots produce fluent, confident explanations. Some are wrong, and many are right in general but do not match how a particular instructor teaches the concept. A student can reach the correct answer through an explanation that conflicts with the course, or through reasoning that is wrong even though the pattern happens to work. AI will also give wrong answers in a convincing way, and give students an incorrect thinking pattern that turns into long term issues as content evolves and gets more difficult. In the Bastani study, researchers documented the standard chatbot giving incorrect answers with logical errors, even in high school math (Bastani et al.).

The result is the same debt in a different form: a right answer built on the wrong method, which eventually fails the student.

4. It’s an incentive problem

Students are not misusing AI out of laziness. They are responding rationally to the incentives institutions have set.

Think of it as a market. The goal students are paid for, in grades, has not changed. What has changed is the price of reaching it. A correct answer that once required hours of struggle, a trip to office hours, or after-school help, now costs a few seconds with a chatbot. When a cheaper, faster route to the same reward appears, people take it.

AI did not create these weaknesses; it removed what used to contain them. Assessments that could be passed with patterns, instruction many students could not follow, and support that answered questions without building understanding have existed for years. What limited the damage was that getting around them took effort, and that effort often produced learning on its own. AI has made getting around them nearly effortless, so every weakness shows at once.

Institutions exist to teach students information and, just as importantly, to teach them how to think. If the most efficient way to succeed in a course is to let a chatbot do the thinking, students will do exactly that.

Many people worry that AI will take jobs because it will think for us. The greater risk is that AI gives us a reason not to think, and we gradually lose the ability.

The outcome may look the same, but the diagnosis is different, and so is the solution. Changing the outcome means changing the incentive.

5. Why this can’t wait

Three features of this problem make waiting expensive.

Your data won’t show it until it’s too late

The usual early warning signs stay quiet. Homework and practice scores look healthy because AI is helping produce them. In a national survey of more than 3,800 U.S. college students conducted in late 2025, 60 percent said they use AI daily or weekly to check answers on homework or class assignments, and 64 percent use it weekly for coursework they do not understand (Gallup, 2026).

The gap shows up later: on exams, in the next course in the sequence, or in a student who quietly switches out of a major. By the time it reaches an institution’s metrics, a full cohort has already passed through.

It shows up in the numbers that matter

Every student who fails a gateway course, changes majors, or leaves takes something with them: tuition revenue, a graduate the program counted on, and an outcome that shapes rankings and reputation. Students who move on to medical, dental, or graduate school carry the gap into admissions exams and beyond, where it reflects on the institution that prepared them.

For schools and districts, the stakes are just as concrete

Schools face the same hidden debt, with different consequences. In K-12, a gap in understanding is most likely to surface in state assessments, course failure rates, graduation and readiness measures, and the number of students who arrive at college unprepared. No study has yet measured AI’s effect on these outcomes directly, but they are where gaps in understanding have always shown up.

The behavior is already widespread in schools. In the same Pew survey, 59 percent of teens said students at their school use AI to cheat at least somewhat often (Pew Research Center, 2026).

The problem also deepens inequity. Families who can pay for skilled tutoring can buy their children real instruction; other students are left with whatever a chatbot supplies. A recent study of private tutoring centers in the United States found that nearly half of tutoring users come from the top tenth of the income distribution, and that this tutoring helps its users while widening achievement gaps and shaping the test scores by which states and families judge schools (Goodman and Nitkin, 2026). The study is a working paper that has not yet been peer reviewed, but its pattern matches what tutors see every day: real instruction goes to the students whose families can pay for it.

Every semester of delay adds to the debt

Each term without a response sends another cohort through carrying the same debt. AI keeps improving, so the credit limit keeps rising, and the patterns it supplies carry students further before they fail. Students now arriving at college have had these tools for much of high school, so the habits start earlier.

Institutions are also behind their own students. In the Gallup survey of college students, a majority of students used AI for coursework at least weekly, yet more than half said their school discourages or prohibits it (Gallup, 2026). Policy alone is not changing how students learn. Redesigning the system around them will.

6. What institutions should do

The goal is to make real understanding pay more than pattern-learning, so that seeking real help becomes the smart choice rather than the slow one. That does not mean banning AI. It means redesigning the system around it.

Assess understanding, not just answers. Ask students to explain why a method works, predict what changes if a condition changes, or diagnose an error in a worked solution. Write exams that look unfamiliar but test familiar concepts. When exams reward reasoning, students have a reason to understand. Some courses will need more conceptual work, others more application. The point is to measure what the course is actually meant to teach.
Connect tutoring to the curriculum. Generic help competes poorly with a chatbot. Institutional tutoring built around each course’s sequence, notation, and methods gives students something AI cannot: help that matches exactly what they will be graded on. That makes seeking real help worth it for students, and it gives the institution something concrete to show prospective students and, over time, the employers who hire its graduates.
Train the people who teach. TAs, tutors, and many teachers are hired or trained for what they know, not for how well they teach it. Training them to diagnose confusion, teach the why, ask guiding questions, and avoid simply giving answers makes human help highly effective and worth seeking. For colleges, that means TA and tutor training; for schools and districts, professional development focused on teaching emphasis rather than only on new tools. In a large randomized trial, tutors given real-time guidance toward these practices produced measurably better student mastery, with the biggest gains among the weakest tutors (Wang et al., 2024).
Build tutoring that works, in school and after school. High-dosage tutoring, meaning frequent sessions in small groups with trained tutors, tied to what students are learning in class, is among the best-supported interventions in education research. In two large randomized trials in Chicago high schools, daily small-group math tutoring raised math test scores by 0.18 to 0.40 standard deviations, improved grades in both math and other courses, and produced gains that persisted into later years. The researchers attributed the effect to more personalized instruction, the same active ingredient described in section 1 (Guryan et al., 2023). For districts, well-designed in-school and after-school tutoring is one of the most direct ways to give every student the kind of help that today only goes to families who can pay for it.
Give students clear guidance on AI. Students need to know when AI supports learning and when it replaces it. The Bastani study suggests that AI designed to guide rather than answer avoids much of the harm, so the question is how AI is used, not whether (Bastani et al.). Guidance that comes from instructors themselves carries weight and creates a more effective learning environment.

Where to start

None of this requires redesigning an institution all at once. The most effective first step is usually small and contained: a diagnostic of one high-enrollment gateway course, or training for the TAs who support it. A single course shows where students’ understanding breaks down, produces results that can be measured within a semester, and builds the case for going further.

An opportunity, not just a defense

This is not only about protecting standards at elite institutions. For colleges, universities, and schools working to attract students, a support system that visibly works is a real advantage. A prospective student and their family can see the difference between “we have a tutoring center” and “our tutors are trained, and they know your professor’s course.”

The institutions that act first will have the strongest story to tell: come here, and you will actually learn, and graduate ready for what comes next.

Conclusion

A degree has always meant two things: that a graduate knows the material, and that they learned how to think while earning it. The second came from the struggle of achieving the first, often with the help of a good teacher.

AI has made that struggle optional. Left alone, the incentives will push students toward the shortcut, and institutions will graduate people with credentials but without the reasoning those credentials once guaranteed.

None of this is inevitable. Institutions that redesign assessments, connect support to their curriculum, and train the people who teach will produce stronger students than they did before AI, not weaker ones.

How Orango can help

Orango Consulting helps colleges, universities, schools and districts, and tutoring companies put these changes into practice. Our work comes directly from the experience described in this paper: years of one-on-one instruction, and building and training a tutoring team of our own.

TA Training & Development. A one or two-day program, scaled to your TA team and courses, that trains TAs to know the professor’s course, diagnose confusion, teach the why, run effective office hours and one-on-one sessions, and guide students’ use of AI.
Tutoring Program Design & Implementation. Building a new tutoring program or revamping an existing one: tutor selection and training, alignment with each course, quality standards, and student uptake. Design takes four to six weeks, and we stay on through launch and implementation for as long as you want support.
Curriculum Audit & Course Redesign. A course-by-course review of what is taught, emphasized, and assessed, then a redesign so grades reflect real understanding.
Faculty, Teacher & Staff Workshops. Full-day workshops and keynote talks on teaching the why, assessment design, and AI, for college faculty and K-12 teachers alike.
Consulting for Tutoring Companies. Tutor hiring and training, program design and packaging, positioning for the AI era, and operations, from a founder who has built and run a tutoring company. Available as a project or by the hour.
Strategic Advising. An experienced advisor on any of these areas, by the hour or on retainer, whether or not you engage us for anything else.

Every engagement starts the same way: we diagnose before we prescribe. We learn how students actually move through your courses and support systems, find where understanding breaks down, and fix the causes rather than the symptoms.

About the author

Matt Breidigan is the founder and CEO of Orango. He began tutoring college chemistry in 2021 and founded Orango in 2023 after seeing the same problem again and again: disorganized instruction that left capable students frustrated and underprepared. Since 2021, he has personally taught over 7,000 hours of one-on-one instruction to more than 500 students.

A graduate of the University of Wisconsin–Madison’s Chemical Engineering program, Matt led process development and technical projects in industry before building Orango. He brought that analytical, systems-minded approach to teaching and developed Orango’s signature “why”-based method, which teaches students why concepts work rather than just what steps to follow, so they can apply their knowledge to problems they have never seen.

That method became the foundation of the company. Matt built and trained a team of tutors, handpicked from former students and held to a standardized instructional approach, so every student receives the same quality of instruction. Orango delivers its tutoring remotely to high school and college students across many institutions.

Orango Consulting brings the same approach to institutions and to other tutoring companies, helping colleges, universities, schools and districts, and tutoring businesses build tutoring programs, train TAs and tutors, and redesign curriculum and academic support for a world where every student has AI.

References

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakçı, Ö., and Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26).
Goodman, J., and Nitkin, S. (2026). Kumon up: Private tutoring centers and student achievement. NBER Working Paper 35671.
Guryan, J., Ludwig, J., Bhatt, M. P., Cook, P. J., Davis, J. M. V., Dodge, K., et al. (2023). Not too late: Improving academic outcomes among adolescents. American Economic Review, 113(3), 738–765.
Jenkins, J. L., and Shoopman, B. T. Identifying misconceptions that limit student understanding of molecular orbital diagrams. Science Education International, 30(3), 152–157.
Kind, V. Beyond Appearances: Students’ misconceptions about basic chemical ideas. Royal Society of Chemistry.
Marken, S. (2026). AI is routine for college students, despite campus limits. Gallup (Lumina Foundation–Gallup 2026 State of Higher Education study).
Nickow, A., Oreopoulos, P., and Quan, V. (2020). The impressive effects of tutoring on PreK-12 learning: A systematic review and meta-analysis of the experimental evidence. NBER Working Paper 27476.
Pashler, H., McDaniel, M., Rohrer, D., and Bjork, R. (2008). Learning styles: Concepts and evidence. Psychological Science in the Public Interest, 9(3), 105–119.
Pew Research Center (2025). About a quarter of U.S. teens have used ChatGPT for schoolwork, double the share in 2023.
Pew Research Center (2026). How teens use and view AI. Survey of U.S. teens ages 13 to 17, conducted September 25 to October 9, 2025.
Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., and Demszky, D. (2024). Tutor CoPilot: A human-AI approach for scaling real-time expertise. arXiv:2410.03017.

Learning on Credit