Adrienne Kennedy didn't teach me how to write. Adrienne Kennedy. The woman who wrote Funnyhouse of a Negro, who wrote A Movie Star Has to Star in Black and White, the plays that read like a mind splitting open in real time, like consciousness itself coming apart in your hands while you're trying to hold it still long enough to read the next line. She didn't teach me how to write. She taught me how to sit in a room with someone whose interior life I had no access to and not pretend I did. Not pretend I knew. Not pretend I could guess. Not pretend, period. She'd had a whole career writing across the inside of people, and she ran her classroom like she ran the work. Nobody got to coast on what they assumed about anybody else. Including themselves. Especially themselves.
That's the foundation. That's the whole thing. Twenty years in a classroom and everything I've done comes back to that one rule. Don't walk in assuming you know who you're talking to. Don't.
Which brings me, the long way around, to AI.
Because the conversation about AI in the classroom right now is being conducted, mostly, by two kinds of professor. Two. And both of them are getting it wrong. Wrong in different ways but wrong, and wrong in ways that are going to look, in about three years, like the educational equivalent of the school boards that banned rock and roll because the kids might dance.
The first kind bans it. Bans it. Treats the thing like contraband, like a cheat code, like a moral failure, like something they can hold the line against if they just frown hard enough at the syllabus. Drafts honor-code language. Threatens consequences. Issues memos. Pretends the machine isn't sitting in every student's pocket, didn't help write the last three emails the professor herself sent this morning, isn't being used right now, right now, by the university administration to write the very policies banning its use. (Look it up. Run any institutional policy document through any detector. The numbers will make you laugh.) That, friends, is professorial cowardice. That's pretending you can teach in 2025 the way you taught in 2015 if you just hold the line hard enough, if you just insist on the old contract loud enough, if you just glare at the students hard enough at the start of the quarter. You can't. The line is gone. The line was gone the minute the thing went public, the minute it could write a passable essay in thirty seconds, the minute your students figured out, before you did, that the rules had changed and nobody was going to tell you.
The second kind embraces it uncritically. Hands the students a list of prompts on day one like it's a goddamn party favor. Tells them AI is a tool. A tool, see, like a calculator. Like a word processor. Like a search engine. Just a tool, they say, with the same conviction your dentist uses to tell you the next part won't hurt. (It hurts.) Pretends the comparison is clean. It isn't. It's never been. A calculator doesn't generate the question. A word processor doesn't write the sentence. A search engine doesn't have an opinion about the result. The comparison is doing rhetorical work the comparison cannot bear, and the professors using it know that, somewhere, and use it anyway because the alternative is harder. They let students offload the part of writing that used to be the part that taught them how to think, and then they grade the output as if the thinking happened. That's professorial laziness. Pure, unleavened, unadorned professorial laziness. It's outsourcing the hard part of pedagogy to a vendor based in Northern California and calling it innovation.
The actual job is harder than either. Way harder. Orders of magnitude harder. And here's the part nobody in the current conversation, nobody, seems to know: the actual job isn't new.
I've been teaching it for years. Years. I just didn't know that's what I was teaching. Nobody did. There wasn't a word for it yet.
I ran a research class at San Quentin. Yes, that San Quentin, the one with the bay view and the death row and the men inside doing time you can't imagine doing. The students had to produce a research paper to complete their degree. Topic of their choice. Their choice. Which sounds simple until you remember where you are. The prison library was, let's say, limited. Limited like a kid's bookshelf. Limited like a waiting room. The research material the students actually needed, the real material, the stuff that would let them do real work, had to come from outside. From the world. From the libraries and journals and databases that exist outside the wire.
So here's how it worked. The student writes a research question. The student writes it carefully, because they get exactly one shot per cycle. Outside researchers, mostly volunteers, take the question and go looking for relevant texts. They find what they find. Whatever they think is relevant, based on what they understood the question to be asking. They print it. They send it to the prison. It goes through the warden's office, gets reviewed, eventually, eventually, gets delivered into the student's hands. The student reads it. Decides what's useful and what isn't. And then, here's the part, here is the part, the student writes a new set of research prompts. More specific. Better targeted. Sharper. Trying to compensate for whatever the outside researcher misunderstood, missed, missed on purpose, missed because of their own politics, their own assumptions, their own academic training, their own blind spots. Outside researcher gets the new prompts. Hunts again. Prints. Ships. Warden. Student. Iterate. The cycle ran as long as the quarter ran. Weeks of latency. Months of work. One paper.
You see what this is, right? You see it?
It's prompt engineering. It's prompt engineering. With a turnaround time of weeks instead of seconds and a human intermediary instead of a model.
What I had to teach those students, years before anyone outside a research lab was using the word AI in a sentence that didn't involve science fiction, was the entire skill set the AI moment now demands of every undergraduate in every writing class in this country. How to write a prompt clear enough that someone with no context can act on it. How to evaluate what comes back without being seduced by it just because it took three weeks to arrive and you're tired. How to read the bias of whatever or whoever did the retrieval, what they considered worth finding, what they considered tangential, what their training or their institution or their politics or their assumptions were filtering in or filtering out. How to revise the prompt based on what came back, not just to get more material, but to get different material. How to recognize when the system was confidently returning something that looked authoritative and was actually wrong. Wrong, or partial, or twenty years stale, or quietly committed to a position the student needed to push against. How to keep your own intellectual project intact when the retrieval mechanism between you and your sources was opaque and slow and full of strangers' judgments.
That is the AI-era research skill. All of it. Every piece. I just had to teach it manually, on paper, with weeks of latency, with a warden as the API gateway. The students learned it because they had no choice. The system forced them to interrogate every step. They couldn't fall in love with the first result because the first result took three weeks to arrive and they had a paper due. They couldn't trust the retrieval because they could see, on the page, that the outside researcher was a human being with limits. They had to be the smartest reader in the room because they were the only person in the chain who knew what their project actually was.
This. This is what students using AI need to learn. The model is the outside researcher. The prompt is the question. The output is the printed packet. The student is still the only person in the chain who knows what the project actually is, or should know, if anyone has taught them how to know it. Most haven't. Most haven't even been told the project is theirs. Most have been told the project is to produce a thing that looks like the kind of thing the professor wants, and the machine will do that for them, and the professor will grade it, and everyone goes home, and nothing has happened.
I learned other things at San Quentin too. I learned to walk into a classroom with as few assumptions as possible. About who the students are. What they've read. What they can do. What they're afraid of. What they secretly already know. Every assumption you bring is a door you're closing on a person whose room you haven't entered yet. Every one. I learned that the job of a teacher is to push students past their own internal limits. Not the limits the institution sets for them. Not the limits the syllabus describes. The limits they have set for themselves about what they're capable of. Those are the only limits worth pushing against. Everything else is administration.
The men in that room demanded I push them. The pressure went the wrong direction. They were teaching me how to teach them. (Read that sentence again. Read it slowly. That's the part most teachers never figure out. The students are the ones holding the syllabus. You are the visitor in their lives. Act accordingly.)
Both of those lessons land harder now, with AI in the room, than they ever did before. Because AI is, structurally, the perfect tool for not pushing past your internal limits. The perfect tool. Designed for it, if you squint. It hands you a competent version of whatever you would have produced, with less effort, before you've had to find out what you're capable of. If you let students use it the way they want to use it, by default, they will use it to avoid the very thing the class is supposed to teach them. They will. They will every time. It's not their fault. Nobody is going to choose the harder version when the easier version is right there and looks the same to anyone who isn't paying attention. The job is to make somebody pay attention.
So I don't ban it. And I don't pretend it's a calculator. I tell my students on day one, the first day, before we've talked about anything else, I tell them this:
This machine will write your assignment for you. It will do it badly, in a way that looks competent. Your job in this class is to learn what badly feels like, even when it looks competent, so that you can recognize it in your own work and in the work of every institution that is going to lie to you for the rest of your life with sentences this machine helped write. We will use the machine in this room. Together. We will compare what it produces with what you produce. We will figure out where it falls flat on its face, where it sounds smart but says nothing, where it confidently lies, and where, occasionally, it surprises you. And by the end of the quarter you will be a better reader and a better writer than a person who didn't take this class, because you will know what the machine cannot do, and that knowledge is one of the few durable skills left in the wreckage of the credentialing economy that paid for this room.
That's day one.
History, language, philosophy, psychology, politics, the whole interdisciplinary tangle, are still the materials. Still on the syllabus. Still being read. But the actual subject of the class, increasingly, the question underneath every assignment, is what it means to do the work yourself when a machine is offering, every minute, every second, to do it for you. That's the question. That's the only question, in some sense, anymore.
The students at San Quentin taught me the answer before I knew there was a question. Before there was a question. They wanted the hard version. They wanted to be pushed. They wanted a teacher who assumed they could handle more, not less. And they showed me, without ever using the word, what real prompt engineering looks like, what real source interrogation looks like, what doing research inside a slow and opaque system looks like, because they had no other option. The system was the lesson. The constraint was the curriculum.
Most students, given a real choice, want the same thing those men wanted. They just don't know they want it yet, because the culture around them is selling the easy version of everything, and the machine is the easy version made literal. The machine is the culture's promise made operational. You don't have to think. We can do it for you. You can be done in thirty seconds. And most students, sensibly, take the offer, because nobody has ever told them what they'd be giving up.
Adrienne Kennedy taught me to sit in a room without assumptions. The men at San Quentin taught me to demand the harder version, and they taught me the workflow that everyone now needs to learn, twenty years before the technology that requires it arrived. Twenty years. AI hasn't changed the job. It's just made the job impossible to fake.
You can ban it and pretend you're protecting something. You can embrace it and pretend you're innovating. Or you can do the actual work, which is harder than it has ever been, and worth doing for exactly that reason. Worth doing because it's hard. Worth doing because the machine can't. Worth doing because the students, the real ones, the ones who showed up to learn something, are going to ask, sooner or later, what you're for. And you'd better have an answer that isn't I supervised the use of the machine that did the thing the class was supposed to teach you.
That's not teaching. That's administration. There's a difference. Find it. Stand on it. Teach.
