6 minute read
I Got AI Out of My Classroom. Here’s How You Can Get AI Out of Yours.
Less than a year ago, I was ready to quit teaching. AI had made my job and my passion utterly miserable and impossible. Cheating was brazen and rampant. Students were utterly checked out. The education business I had spent my adult life building, that had brought me so much joy and helped so many students build their intelligence, had turned into a factory of misery and idiocy.
I should add that my classes are online and based on challenge problems. Individual problems can take 30-60 minutes to solve, and students can easily access AI. It’s the worst possible situation.
I won’t go into every single experiment that I tried. There were just too many. Instead, I’ll share the methods and principles that worked.
Education isn’t about what a student does; it’s about what a student can do. Education is measured by skill and intelligence, not by the number of pages written or number of problems solved. To make my programs resistant to AI cheating, I made one key decision: if a student’s only skill was using AI, that student would not pass the class.
I knew I couldn’t detect every single type of AI use. I also knew that I couldn’t make every single task AI proof. But I didn’t need to. I just needed to have enough AI proof tasks that students who couldn’t solve problems without AI also couldn’t pass the class.
My classroom is unusual, so the AI proof tasks I use will likely be quite different from what you use. But the principle will be the same.
Take the most vulnerable type of in-person course, for example: writing-heavy humanities courses. Today, students who can’t write at all can pass those courses because they use AI on all take-home essays. But that is something that, with a little creativity, can be addressed. Just as most educational institutions set aside entire days for exams, they can also set aside entire “Writing Days” for in-class essays. A full 8-10 hours is more than what most students spend on a 5 page paper anyway. As long as the paper topic is specific enough that students can’t memorize generic AI-generated papers in advance of the Writing Day, and students have no AI access during the writing day, a full-day, in-class writing day can accurately measure a student’s writing skill.
From there, it’s a relatively simple question of adjusting the grading system to ensure that students who can’t write also can’t pass the class. For example, in-class papers might be 85% of the grade, with no “50% minimum” grading schemes. Students who can’t write coherently (a much larger percentage than you might expect) should get single digit grades. That single digit, by the way, should be a 0.
Exacting grading systems should function as incentives, not punishments. Institutions should have writing tutors or other support systems that help students develop their reading and writing skills. There must always be a mechanism for students to bring their skills to the required level.
The principle of measuring skill, and using that to determine who passes, is even easier to apply to STEM classes than in humanities classes. Making the entire grade based on live, in-class assessments is easy. Tests and quizzes are a natural way to test proficiency in STEM subjects. As long as the students who know nothing fail the course (as in, get an F, not a C), the course cannot be destroyed by AI.
Using skill assessment to determine who passes and who fails will require a somewhat seismic shift in how Americans look at grading. At the moment, most institutions allow students to substitute diligence for skill. Homework grades, participation grades, group projects, and the like allow students with no proficiency whatsoever to still pass courses. This approach is supposedly humane, as it spares students the trauma of failing a class and the social embarrassment of repeating a course. But that approach is untenable in the AI era. Diligence can no longer be treated as a substitute for skill. AI has made diligence impossible to measure, and grading diligence has always been a dubious approach anyway.
There is one other, major change I made to my classes, and every humanities teacher should make to their classes. Teachers have discovered that AI detectors don’t work. Students put AI-generated content into those same AI detectors and then incrementally change words until the AI detectors views the content as human produced. Thus, there is no realtime AI detector that works. I doubt that any will ever be created.
The Big Realization that I had is that there was no requirement for realtime detection. AI improves every year. So do AI detectors. A student might trick today’s AI detector, but it’s unlikely that his trickery will fool an AI detector from a few years (or a few decades) in the future. Each year, institutions can audit student papers from prior years with the newest AI detectors, retroactively failing students or stripping degrees as appropriate.
Retroactive AI audits make it possible for teachers to assign term papers, hard math proofs, and other tasks that take more than a full day. The nice thing about AI detectors is that they are more accurate with longer blocks of text. Thus, a 20 page term paper that fooled today’s AI detectors won’t fool AI detectors from several years in the future.
Many teachers worry about student backlash to harsher AI policies. But what I saw in my classroom wasn’t rage. It was relief. The fact is that many students use AI because other students use it, and they want to be able to compete. People cheat because others cheat, and they don’t want to be the only sucker who didn’t cheat. But if students have an excuse to avoid cheating, they’re happy to take it. If students know that avoiding AI doens’t make you a sucker, but rather makes you someone who will still have a degree in 10 years while cheaters lose degrees and jobs, they are relieved to avoid AI.
Education is about developing skill, not obedience. Intelligence, not volume of work, is the only measure of education. By making skill development a requirement for passing a course, almost any teacher can stop AI cheating in almost any course.