Somebody sent me a thank you note last week, and I could tell about two sentences in that a machine had helped write it. The rhythm was too even. The gratitude was a shade too tidy.
Here is the thing, though. I was glad to have it. I knew the person meant every word, and I suspect they were short on time and the machine helped them say something they genuinely wanted to say. It captured the moment beautifully, better than a hurried note dashed off between meetings would have. I have spent enough years around these tools that noticing the seams does not trouble me.
But that last sentence is worth pausing on, because it describes a privilege and not a virtue. I am comfortable because I have had the time to become comfortable. It is my job. Most people have not been given that, and I have sat in enough rooms with enough faculty to know that a great many of them would have felt something quite different about that same note. Something between flattery and suspicion. A small sense of having been handled. I do not think those people are being unreasonable, which is why the question is worth sitting with.
So let me ask it plainly. What exactly is supposed to be wrong with a thank you note that a machine helped write?
Consider flowers.
When I send my wife flowers, she is happy about it every single time. I did not grow those flowers. I did not spend a season watering and pruning and fussing over the plant. I did not walk out to the garden with shears at first light and cut the stems myself. I drove to a store, looked at what was on offer, picked the ones that suited the occasion, and had them sent. She has never once held that against me, and it would be strange if she did.
The labor of production was never what made the gesture mean something. What made it mean something was that I knew the occasion, I knew what she likes, I chose accordingly, and I actually followed through. The choosing and the sending are the human parts. The growing and the cutting have been outsourced for as long as either of us has been alive.
So why does the email trouble people when the flowers do not?
I have come to believe the answer is that we are living through the awkward stretch, the years before a convention settles. Nobody pretends to have grown the roses, because everybody understands the arrangement. There is no deception at the florist. With machine assisted writing, we have not yet agreed on what the arrangement is, so noticing it feels like catching somebody at something. That is a lag in our shared conventions, not a defect in anybody’s character, and I think saying so out loud does a great deal of good.
Here is what is worth remembering: the flower convention is younger than we think. Giving cut flowers as a romantic gesture is largely a Victorian invention. A dozen red roses in February in North Carolina is a twentieth century artifact that required refrigerated shipping and cargo flights out of Colombia and Ecuador. The norm my wife and I both take entirely for granted is roughly a hundred years old, and the particular version of it is closer to fifty. It settled because the outsourced part was never the meaningful part, and eventually everybody figured that out.
The choosing and the sending are the human parts. The growing and the cutting have been outsourced for as long as either of us has been alive.
I expect something similar with AI assisted writing, though not evenly and not everywhere. In five years, most people will assume a machine helped with the routine correspondence and will not think twice about it, because what they actually wanted to know was whether you meant it. The question is already migrating from who typed this to who stands behind it. Nobody asks who grew the roses. They ask who sent them.
But I want to be careful here, because there is a real difference between the two cases, and it matters enormously for those of us who teach. A florist can grow and cut and arrange. A florist cannot decide to send my wife flowers on a Tuesday for no particular reason. The machine can generate the sentiment itself, not merely execute it. That is why I believe the convention will settle quickly for the thank you note and the meeting summary, and will stay contested for a long while in the places where the entire point is that a specific human being vouched for something. The letter of recommendation. The condolence. The performance review. The comments in the margin of a student’s paper.
Which brings me to the second thing on my mind, which is how little time we have to sort any of this out.
On the ninth of October, 1903, the New York Times ran an editorial titled “Flying Machines Which Do Not Fly.” It was written two days after Samuel Langley’s aerodrome dropped into the Potomac, and it predicted that a machine that actually flew might be developed in something between one million and ten million years. Sixty nine days later, two bicycle mechanics from Ohio flew at Kitty Hawk, about two hours from where I sit typing this in North Carolina.
I have always loved that story for the obvious reason, but the part I find instructive is not that the experts were wrong. It is where they were looking. Langley had the Smithsonian behind him and roughly seventy thousand dollars in government grants. The Times was not being foolish in doubting him; his machine genuinely did not work. What the paper could not imagine was that the answer would arrive from two men with no grant money at all, working out of a bicycle shop, on a sand dune in eastern North Carolina.
Eleven years after Kitty Hawk, aircraft were flying combat missions over the Western Front. Sixteen years after, men flew nonstop across the Atlantic. Sixty six years after, a man walked on the moon. I had a grandmother who was alive for the first of those and the last of them. One human lifetime carried us from a million years away to routine.
I have watched a smaller version of the same compression in my own life. Antenna television to streaming. No computer in the house to a computer in my pocket. Driving to an office every day of my working life to collaborate with the NC Community College System from a remote community college or home office when the situation called for it. And now this, which arrived faster than any of it. Generative AI became available to ordinary people in late 2022. It is 2026, and I cannot name a single college in our system where it is not already a daily fact of life.
One human lifetime carried us from a million years away to routine.
I used to think the pushback I got on AI was mostly a failure to keep up. I no longer think that, and I would caution any colleague against thinking it either. Consider again what happened with aviation. The people who were uneasy about airplanes in 1914 were not slow adopters who needed a workshop. They were reading the situation accurately. Within eleven years the technology was being put to uses its inventors never had in mind, and nobody had to be a pessimist to see it coming.
The speed of a technology tells us nothing whatsoever about whether it is good. What it tells us is that we will not be granted the usual few decades to work out the norms before the thing is everywhere. Every previous technology of consequence arrived slowly enough that a society could absorb it by living alongside it. A generation grew up with the automobile and simply knew things about it that their parents had to learn on purpose. That unhurried absorption is what we have lost, and it is the whole of the problem.
Which raises a question I do not think we ask nearly often enough. If a society no longer has time to absorb a technology by osmosis, whose job is it to help everybody catch up?
I want to answer that carefully, because the easy answer is that it is everybody’s job, and that is the same as saying it is nobody’s.
It is not going to be parents. I say that with no criticism whatsoever intended. The parents I know in eastern North Carolina are working, often at more than one thing, and the notion that they will develop a working understanding of machine learning between a second shift and a ballgame is not a plan. It is a wish. They will do what any reasonable person does, which is form an impression from headlines and from whatever their child shows them on a phone, and those two sources point in opposite directions.
It is not going to be supervisors. A supervisor may forward an article, and may even mean it kindly, but a supervisor’s job is to get the work out the door. Very few of them have been given the time or the standing to sit down with an employee and think carefully about when a machine’s output can be trusted and when it cannot. Where employers do take it on, and some are beginning to, they are usually training people to use one product rather than to think about a category.
It is not going to be colleagues, for the same reason it was not going to be supervisors, and with the added difficulty that a colleague who is confidently wrong is often more persuasive than one who is carefully right.
It is certainly not going to be the companies who build these systems. They will explain their product ably and at length. That is not the same as forming a person’s judgment, and we should not confuse the two.
The easy answer is that it is everybody’s job, and that is the same as saying it is nobody’s.
That leaves educators. And I do not mean that as a burden reluctantly accepted. I mean that education is the only institution a society maintains for the deliberate formation of judgment at scale, and this is not the first time the job has landed here.
When the automobile became ordinary, schools took on driver education. When household chemistry and nutrition and household finance became things a person could get badly wrong, schools took those on too, whatever we have since decided about the name. When card catalogs gave way to databases and then to a search box that would return ten thousand results of wildly varying honesty, librarians and faculty built information literacy into the curriculum, and it is difficult now to imagine a college without it. In each case something arrived that ordinary people had to be able to evaluate, and in each case the work of teaching that evaluation came to rest with us. Not because we asked for it. Because there was no one else standing there.
Community colleges sit in a particular spot with respect to this one, and I would ask my colleagues not to be modest about it. We are open door. We teach the eighteen year old and the forty three year old in the same section. We train the nurse, the lineman, the welder, the paralegal, the early childhood teacher, and the person who will run the front office at a small business where nobody else has thought about any of this at all. When a student leaves us and goes to work, whatever we managed to teach them about judgment travels into a workplace that may have no other source for it. Our graduates become the person in the room who knows to check.
That is a considerable responsibility and I do not want to oversell what any one instructor can do about it. But it does change what the work is. If we treat AI as a specialization, something for the technology faculty and the innovation committee, we will have declined the assignment. It is not a specialization. It is the current form of a thing we have always done, which is to prepare people to exercise judgment about the world they are actually going to live in.
Now let me say a word about tools, because I hear a particular phrase often and I no longer find it adequate. The phrase is that AI is just a tool, and it all depends on who is holding it.
There is truth in that, but it is too thin to lean on, and the historians of technology worked this out long before any of us were arguing about chatbots. Melvin Kranzberg put it about as economically as it can be put: technology is neither good nor bad, nor is it neutral. Langdon Winner made the longer case that the things we build carry particular arrangements of power inside them whether their makers intended that or not.
What they mean, in plainer language, is that tools have grain, the way a board has grain, and the grain makes certain things easy and other things hard. Aviation did not simply happen to find military use. A machine that carries weight quickly over ground nobody can defend is going to be used that way, and every major power on earth worked that out within a decade. The technology did not determine the outcome, but it certainly determined which outcomes were cheap.
AI has grain too. It is far cheaper to generate plausible text at scale than it is to verify any of it. So the flood of confident, unfootnoted, wrong material is not a misuse of a neutral system. It is the system working exactly as built, aimed somewhere unhelpful.
Our graduates become the person in the room who knows to check.
I raise this because it makes the case for teaching ethics stronger, not weaker. If tools truly were neutral, then ethics would be nothing but a matter of personal virtue, and our job would end at telling students to be good people. Because tools have grain, students need to learn to read the grain. That is a skill. Skills can be taught. And it does not belong in a standalone module bolted onto an already full semester.
I want to be clear about that last point, because when I say ethics belongs in every course regardless of the curriculum, faculty hear one more thing added to a syllabus with no room in it. That is not what I mean. Every discipline already contains this question.
- The nursing instructor is asking when a clinician accepts a machine generated recommendation, and who is liable when it is wrong.
- The welding instructor is asking what a student does when the software’s spec sheet does not match what the eye sees in the metal.
- Accounting is asking what professional judgment means when the reconciliation is automated.
- Criminal justice has been arguing about risk scores and predictive policing for years now.
- English is asking what authorship means, which is my flower question with a grade attached to it.
None of that requires cutting content. It requires asking, inside whatever you already teach, where the machine’s judgment stops and the professional’s judgment starts. Framed that way, it is not an ethics unit. It is a discipline specific competency, and it may be the most durable thing we can hand a graduate.
And it takes both halves, the personal and the institutional. Aviation did not become safe because pilots were told to be careful. It became safe because good practice was built into checklists, into training, into the habit of reporting a near miss rather than burying it. Individual judgment, supported by structures that make the right thing the easy thing.
A college has its own versions of those structures, and they are more within our reach than we sometimes act like. Course design review. Assignment prompts written so that the thinking is visible and not merely the product. Program advisory committees that include employers who can say plainly what they now expect a graduate to be able to verify. Departmental agreement on what disclosure looks like, so that a student is not guessing at five different standards across five different courses. None of that requires waiting on anybody.
The last thing I want to offer is the one our students would probably remember longest, and it comes from a cartoon.
Aladdin finds a genie who can do remarkable things, and his trouble is never the genie’s power. It is knowing what to ask for. The genie in that story operates under stated constraints: he cannot kill, cannot make anyone fall in love, cannot raise the dead. Somebody thought carefully about guardrails, and the story is better for it. And Jafar does not fail because he asks for too little. He fails because he asks for the most powerful thing he can imagine without reading the fine print, and he receives precisely what he specified. Phenomenal cosmic power, in a very small space.
I had thought this was a small insight of my own until I went looking, which is a lesson in itself. That gap between what you asked for and what you meant is a recognized problem in AI safety, and it has a name. A team of DeepMind researchers calls it specification gaming, by which they mean a system that does precisely what it was told and entirely misses what was wanted. Their paper opens with King Midas and the golden touch. Elsewhere in that same literature you will find the genie in the lamp and the sorcerer’s apprentice doing identical explanatory work.
I take some comfort in that. It means the old stories we already know are not a simplification of the technical problem. They are a reasonably faithful account of it, which is convenient for those of us whose job is to make hard things teachable. The citation is Krakovna and colleagues, and it is short enough to assign.
But here is the difference that matters for a classroom, and it is the reason I find the analogy hopeful rather than ominous. Aladdin got three wishes. One shot each, irreversible, no correcting course. That is not what working with these systems is like. You ask, you look at what comes back, you tell it where it went wrong, you ask again. It is a conversation, not an incantation.
That single distinction should reshape what we teach. If prompting were wish making, the skill would be getting the magic words exactly right the first time, and we would be running workshops on secret phrasings. Because it is iterative, the actual skill is evaluation. Looking hard at what came back and knowing whether it is any good. That is critical thinking with a faster feedback loop, and we have been in the business of teaching it since long before anybody had a computer.
It is a conversation, not an incantation.
Which is, I think, the most reassuring thing in this entire essay. The competency the moment demands is not new to us and it is not outside our expertise. We do not have to become computer scientists. We have to do more insistently the thing we were already trained to do, in a setting where the consequences of not doing it have gotten sharper.
One warning about the genie, though. A genie is never wrong; he is only too literal. These systems fail differently and more often. They will hand you a citation to an article that was never written, in a journal that exists, by an author who is real, and it will look perfect. Aladdin never had to verify anything. Our students do, every single time, and if we do not teach them that habit, I am not confident anyone else will.
So here is where a few ordinary things have left me.
The flowers tell me that conventions settle, and that when they do, we will care about the intent behind a message far more than the mechanism that produced it. The flying machine tells me two things at once. It tells me we will not be given the century the flowers got, because the unhurried way a society absorbs a new thing is not available to us this time. And it tells me the tool has grain, that some of the worry is warranted, and that judgment and structure have to arrive together. The genie tells me the skill worth teaching is not the asking but the looking, and that we are fortunate to get more than three tries.
Taken together they point at one conclusion, which is that the gap between how fast this arrived and how slowly people can reasonably be expected to adjust has to be closed by somebody, and the only institution built for that work is ours. Not because we are wiser than the parent working a double or the supervisor with a deadline. Because forming judgment is the thing we do, and they have been given neither the time nor the standing to do it.
The genie is out of the bottle. I do not say that with any fatalism, and I would ask my colleagues not to hear it that way. The question before us was never whether wishes would be made. It is whether anyone will have taught the person making them how to word it, and how to look hard at what comes back.
I would like that to be a teacher.