The dean of Harvard College spent the first weeks of the school year defending one sentence. David J. Deming told students the college should accept, and maybe even encourage, AI in classes built around writing. Chasing cheaters, he said, wrecks the trust between teachers and students.
He is right, and I said the same thing in April 2023, in a piece called The Socratic AI, when most colleges were still writing rules against the tool. No school can control what a student does on a laptop at midnight.
His distinction is a barbell. Encourage AI where it can deepen learning. Protect the assessments where a student has to show what he can do himself. That is a defensible line, and it leaves one question open. In the courses on the encouraging end, the long papers and the projects, what exactly is the deepening supposed to look like?
Deciding not to police something is not a plan. It only clears the space. What you put in that space is the whole job, and almost nobody is talking about that part.
Here is what I put in it.
The tool is not the lesson
Arguing with the tool is the lesson.
I use ChatGPT in my classroom as another voice in the room. A student says something. I type it into the machine and read the answer back out loud. Then I ask the student to respond to it.
Last term I asked a group for a cutting-edge idea that would shape the future. They came back with GLP-1 drugs and healthcare. I knew what was wrong with it. It is a real story, but it is a story that has already happened. It has been on magazine covers for two years.
I could have said that. If I had, I would have been the professor who embarrassed a team in front of the room, and the rest of the class would have learned to offer safer answers next time.
So I typed their idea into ChatGPT instead and asked whether it was cutting edge. The machine answered the way I expected. Widely covered, well understood, already priced in.
Now they were not answering me. They were answering the machine first, and then me. They went back to work, and what they brought back was an idea nobody in the room had heard before. That was the whole point.
That is the trick. I already know what is wrong with the first answer. I let the machine be the one who pushed back.
This reverses the usual relationship with the machine. The student is not asking it to think for him. He is thinking against it. The machine proposes, challenges, and sometimes gets it wrong. His job is to question it, defend his position, and decide. The learning is not in the answer the machine gives. It is in the friction between the student and the machine.
It goes both ways
This is the part people miss. The same machine can do a student's thinking for him or force him to think harder. It depends on who is holding it and why. In his dorm room at midnight, it finishes the work. In my classroom, it starts an argument.
A professor now has an assistant. So does the student. The tool did not choose which of those it would be. We choose, every time we write an assignment.
One catch. This only works if the student knows the material. If he never did the reading, he cannot tell that the machine is wrong. He just nods at it.
Berkeley Law understands this. Its default rule keeps AI out of the work students submit for credit, down to outlining and editing, and instructors can lift that rule where it serves the course. The school is explicit about why. Graduates need to be fluent with these tools and able to think, write, and judge without them. That is not a rule against my classroom. It is the condition for it. Learn to think, then argue with the machine about it.
What we are actually deciding
Harvard will publish its guidelines this academic year, and other universities will read them closely. The question is not whether they allow AI. That fight is over. The question is what they ask a teacher to do with it once it is in the room. Permission takes an email. Teaching takes the hour, the preparation, and the people, and no budget has paid for that yet.
We can count how many students finish and how fast. We have no easy way to count whether they can think, and when institutions have to choose, the thing with a number has a habit of winning.
Socrates did not give his students better answers. He gave them better questions. Perhaps that is the real opportunity of AI in education. Not teaching students how to get an answer from the machine, but teaching them to question it. Never letting it have the last word.