Eight Rules for Teaching in AI World

Kevin A. Bryan - August 2026

How should university teaching change due to AI? A professor's job is to decide what to teach and how to present this material. The job of the course structure is to ensure students learn that content. AI presents three issues: the link between performance and student knowledge has been broken, students need to be able to use AI effectively in their future life, and AI should let us improve how much students learn. We need to modify our courses, our expectations, and our evaluations because of AI, but we can do so in a way that makes education more effective than ever.

Here are eight rules for how to modify our teaching in AI World.

1. Decide What Should Be Learned Without AI

What should students know how to do? In some sense, we don’t care how someone writes a good essay or solves an engineering computation. These can be evaluated on quality alone. The problem is that much of what we can measure in the classroom is only a proxy for what we want to evaluate. No one is going to invert the matrix by hand when they do a regression, but knowing how to do so is a good proxy for your ability to extend that knowledge. We currently give many assignments that are good proxies when AI isn’t used, but are terrible when AI is used: a French term paper written with AI is useless when perfect AI translation exists. To design assignments, we need to be explicit about when we are trying to figure out whether the student can do something independent of AI.

2. Teach Better Than the Pre-AI Status Quo

Don’t be deluded: even prior to AI, some students slacked off. Some cheated. Some studied the absolute minimum required to pass a class. Indeed, US college students’ study time fell from 40 hours per week to 27 hours between 1961 and 2003. In a study of Chegg, a site where assignments and answers could be posted, a quarter of finance students copied wrong answers placed there, and those students learned less afterward. In a perfect world, we would know, for each student and in real time, how much they understand relative to the learning goals we have set. We would give them tools to be able to correct those errors. And we would understand where our own teaching has been misleading or could be improved. Just going back to the teaching and evaluation methods of a generation ago is insufficient.

3. Give Students an Incentive to Learn

Students exert more effort when they are tested. Study effort is usually hard to observe, but looking at All Day TA data, we can see how many interactions students have with a study tool week-by-week and across classes. There is a large – in many cases, 5-10x – increase in use on exam weeks, and also a large increase in study effort in courses that have exams versus those that don’t. This is consistent with other research. Oreopoulos et al found that detailed study plans, study reminders, and coaching on grades barely affected university study performance. On the other hand, students who took an intermediate exam had much higher pass rates and grades on the final. Likewise, setting a goal for completing online practice exams led to higher performance while a goal for the final course grade was much less effective. All of this matters more in AI World: since it is harder to force effort through take-home assignments, courses need to be structured with regular, graded assignments that check learning.

4. Don't Give Assignments That Are AI-Cheatable

Go through your syllabus and find every take-home assignment or exam worth nontrivial credit. I assure you: AI can get a top mark in your class, today, when used by students who understand how to use it. And many students will cheat in this way. Indeed, I wrote on November 30, 2022, the day ChatGPT was released, that take-home exams and problem sets should no longer be used. A recent large-scale study found a huge drop in the time spent on homework questions which were AI-susceptible, with a 25% fall in performance on those topics at exam time. A study from China found that post-AI, homework scores went up 18% and completion time fell 30%, but exam performance fell 20%. So how can we get students to study in a way where they actually learn? Three options. First, only give in-class assignments like “viva voce” oral exams, exams written on blue books, and so on. The problem is that these are very time-intensive to evaluate. Second, use AI to fight AI. On term papers, I ask for a précis early, show students what a frontier model can already produce from it, and tell them their final term paper grade has to be better. How? Get into the library stacks, get on the phone, visit sites, create data. The third option is to design at-home assignments that are not cheatable, but which still let students learn. For instance, I use All Day TA’s “Intelligent Quiz”, graded 0 or 1. Students pass if they get 10 or more questions right, no matter how many they get wrong. Th AI requires students to try to explain their logic on questions they get wrong, meaning that students learn anyway. Since the grade doesn’t depend on how many are done right, there is less incentive to cheat. And because it is fully automated, methods like this can be scaled to even large classes.

5. Get Students to Learn More Efficiently with AI

For students to learn effectively, we should use the precise language and content of your specific class, apply pedagogical techniques like asking students to explain or teach a concept rather than just giving them the answer, and draw on ideas like spaced repetition, where we return to topics students struggled with, and mastery learning, where we ensure students have reached a certain bar on each topic before moving on. The problem is this was, prior to AI, incredibly expensive, even though Bloom famously argued for the large benefit of this style of tutoring (his result was later revised down to smaller but still qualitatively huge effects). AI in a harness drawing on those ideas, however, makes it cost-feasible. A six-week study of Nigerian students found a 0.31 standard deviation gain from an AI tutor, and a World Bank meta-analysis found 0.12 standard deviations improvement even from much older AI and learning technology with very short timelines on average, in line with gains from Khanmigo’s AI despite very limited student usage, and a similar study of a tutor-type AI system with Harvard physics students. Students "studying with AI" by just getting homework answers clearly doesn't work. But AI also makes possible class-specific, individualized, cost-effective mastery learning at scale.

6. Use AI to Improve Our Teaching

AI can not only help students learn better, but also help professors teach better. If we teach something in a confusing way, how do we know? Perhaps mistakes appear at exam time, or we overhear a misunderstanding in the hallway. But generally, it’s pure guesswork. An old literature in education argued that eliciting evidence of student misunderstanding has large learning benefits. AI can help automate this. For example, AI study tools might capture a thousand student interactions per week in a mid-sized class. These can be used to (and in All Day TA's case, are used to) summarize the cause of common misunderstandings and report that back to the professor before the next class. Letting professors learn quickly where they have been imprecise can help: in a study of over 1,000 online computer science teachers, randomized feedback availability increased their use of student ideas by about 13%. Because AI digitizes student studying, it becomes possible for us to better understand where students are struggling and hence to improve our courses.

7. Personalize Assignments

Traditionally, we give all students the same assignment. This is not because we think that’s a more effective way for the students to learn. Rather, it’s just because the cost of individualizing and then grading heterogeneous assignments is too high. That said, having students work on problems exactly at the level they are struggling is long-known to improve learning, just as asking weightlifters at the gym to all lift identical barbells would be much less effective than targeting their current individual levels. A randomized evaluation of Mindspark, an adaptive learning system from before modern LLMs, generated gains over four months of 0.37 standard deviations in mathematics and 0.23 in Hindi, in line with “Teaching at the Right Level” interventions. Using generative AI, a randomized experiment on learning Python in Taiwan found that students whose problem sets were adaptively selected by an LLM scored 0.15 standard deviations higher on an in-person, no-AI final than students receiving a fixed sequence. All of this is consistent with a large body of pedagogical research on the benefits of distributed practice and evaluation. Letting AI handle the task of choosing problems and then grading them lowers the cost of student-specific assignments.

8. Raise Standards

If AI lowers the cost of learning content, doing research for a term paper, finding and analyzing data, and so on, then we should raise expectations! The sort of draft that earned an A ten years ago should no longer earn an A. In randomized workplace studies, writing quality improves and time to completion falls with AI, while consultants working on tasks within AI's competence domain produce better work. There is no excuse for sloppy citation, basic investigation of data, error-ridden prose, and so on: students should be expected to learn more in your class, and they should be expected to produce better work whenever they are allowed to use AI directly in any part of the project. This only works if you have a very clear AI use standard. In my courses, I use the following AI policy:

You may use AI in this course. If you use AI to write, you must specify in an attached note precisely what AI assisted with. Understand that with AI, my expectation for the quality of your ideas and writing is higher than it used to be - you need to show that you are able to provide value above and beyond AI. For readings, you must read them first yourself! Do not summarize them with AI. It will be very obvious in class if you do. I want your ideas; I can ask GPT for its thoughts without coming to class, as can your classmates.

The clearer you are about your expectations about where students should use AI to improve their work, and where they must avoid AI to permit other learning, the more students are able to get that balance right. Rather than throw our hands up and complain about what AI has done to universities when we see students GPT an assignment, we can instead make minor adjustments to our courses that draw on AI to improve student learning, deter cheating, make evaluation more effective, and raise the expected standard students will hit.

Kevin A. Bryan | University of Toronto