The Students Were Right Wrong

IWhat I Heard in the Booing

In May, graduating students at the University of Central Florida booed when their commencement speaker praised artificial intelligence. The ceremony was supposed to mark something they had accomplished. Instead, they were hearing about the next industrial revolution.

I work in the industry being booed at. When I first wrote about that moment, I called the students right.

I understood the anger. I still do. A sales pitch about the future can sound different when you are the person expected to find a place in it. You can resent being told to welcome a disruption whose benefits someone else gets to describe and whose costs you may have to discover yourself.

But I did something with that sympathy that I should not have done. I gave the crowd my own argument. They were booing the euphemisms. They understood the apprenticeship being dismantled. They could hear the dishonesty beneath the reassurance. By the end of the essay, a reaction at a ceremony had become evidence that the students saw everything I thought I saw.

I had not established any of that. The booing told me that people objected. It did not tell me that they objected for the same reasons, or that my account of those reasons was correct. I was pleased to hear people refuse a story I distrusted. That pleasure made me less careful.

The students were entitled to object. My earlier essay did not advocate refusing every use of AI. I am taking up that separate question here: which uses should we accept, change, or refuse? Blanket rejection is one possible answer, but not a position I can attribute to the people who booed.

IIThe Job at the End of the Degree

There is a bargain inside a great deal of educational advice. Work hard. Acquire the qualification. Become useful. There will be a place for you.

The bargain need not be written down to shape a life. Years, money, and expectations can be committed to it before the person making the commitment has much experience of how employers decide. When the promised place becomes harder to find, telling that person to be adaptable can sound like changing the terms after payment.

The difficulty is real. In the second quarter of 2026, the New York Fed reported unemployment of about 5.6 percent among recent college graduates and underemployment of 42 percent. That second measure describes graduates in occupations that typically do not require a degree. It is not a count of people doing no useful work, but it is a reason to take the disappointment seriously.

Those figures do not identify AI as their cause. Nor does the existence of a capable model tell us how much of a difficult job market to attribute to it. Interest rates, hiring decisions, the organization of work, and changes that began before generative AI became widely used also deserve examination. Natalia Emanuel and her coauthors have, for example, investigated remote work and its consequences for younger workers' access to learning and employment.

There is no comfort in giving someone an explanation merely because it fits the shape of their distress. An incorrect diagnosis can leave them preparing for the wrong problem.

That does not make the institutions selling education innocent. If a university encourages students to borrow against an expected career, it should be willing to explain the basis of that expectation, what has changed, and what support it will provide. Calling education intrinsically valuable does not answer a question about a loan payment. Calling the labor market uncertain does not excuse pretending otherwise while recruiting the next class.

But a degree cannot guarantee that a particular arrangement of tasks will remain valuable to employers. The effort was real. The knowledge may remain valuable. Neither fact can make the original route stay open.

That is a painful distinction, especially when the people announcing it have secure jobs. It is also one we have to make if the point is to help someone decide what to do next.

IIIThe Tool Does Not Choose Who Benefits

A tool that makes an experienced worker faster may also help a beginner learn the work.

In a published study of customer-support agents, AI assistance increased issues resolved per hour by about 15 percent on average, with gains of about 30 percent among less experienced and less skilled workers. The researchers also found evidence of learning. The tool did not reserve its help for people who were already experts. In this setting, it helped beginners especially.

That does not settle the employment question. A tool that helps a beginner perform can make the beginner more attractive to an employer. It can also let an employer meet the same demand with fewer people. The productivity gain does not tell us which decision will be made.

There is evidence that makes the second possibility difficult to dismiss. Stanford's updated Canaries in the Coal Mine? paper reports a widening relative employment gap between young workers in occupations with greater and lesser exposure to AI, driven primarily by reduced hiring. The authors describe their evidence as descriptive, not a causal estimate of AI's effect, and report no widespread displacement across the economy. That is a reason to investigate the loss of opportunities for beginners. It is not permission to declare their future already gone.

Who gets the tool? Who can check its answers? Who receives training? Who keeps the time or money it saves? Who is still allowed to learn by doing work slowly?

Consider a system that helps someone understand a benefits letter. It might make an institution easier to reach. If the institution then uses the same system to remove access to a person who can resolve an unusual case, the experience may become worse precisely where help matters most. The explanation could also be wrong. The possibility of access does not excuse those failures; the possibility of failure does not make the access worthless.

This is why I think blanket rejection is wrong. It treats the label as though it has already answered the questions that matter. So does blanket enthusiasm. Both spare us the less satisfying work of examining a particular use and deciding what to permit, change, or refuse.

The important disagreement is over the arrangements people make around the tool. Those arrangements can be challenged. They can also be built differently.

IVRefusal Has to Be Specific

Nobody owes an AI company their enthusiasm. Nobody owes it their private information, their creative work, or a subscription bought as proof that they are keeping up. A student can have a sound reason to refuse a particular product or use. A university can have a sound reason to keep a task human.

But disliking the companies does not tell us everything about the tools. If a system is changing work you want to do, learning enough to judge it may serve you better than allowing either its advocates or its opponents to supply your entire understanding. That need not mean adopting it. Inspection can give a refusal a firmer basis.

I would rather ask a graduate to examine one piece of actual work than tell them to prepare for an abstraction called the AI future. What does the task require? What can a tool do reliably? Where does it fail? Can you detect the failure without asking the same tool to certify its own answer? What would you still need to understand if it were unavailable?

These are questions for learning, not a promise of employment. Someone can answer them well and still struggle to get hired. I do not want to replace earn a degree and you will be safe with learn to use AI and you will be safe. The new sentence would be another way of selling certainty to the person who can least afford a bad purchase.

And individual preparation cannot take over an employer's responsibility to develop people. If the tasks once given to junior workers are automated, the employer needs another account of how a person becomes senior. That might involve different work, supervised practice, or deliberate training. It requires more than telling applicants to arrive experienced.

There is a political distinction here, too. Refusing one tool, organizing against a deployment, demanding compensation, and rejecting a whole field are different actions. A specific demand gives people something they can agree to, resist, negotiate, or be held to. General disgust may begin the conversation. It cannot do all of that work by itself.

The students do not have to turn their commencement ceremony into a policy seminar. A boo is allowed to be a boo. But when I turn it into an argument, I owe the reader more than the sound.

VThe Honesty I Was Asking For

Useful questions can disappear inside polished language.

Augmenting human capability. Reshaping the nature of work. Unlocking opportunities we cannot yet imagine. These phrases can describe real possibilities. They can also let a speaker avoid saying what they know about the work being changed now.

I have watched uncomfortable answers get sanded down before they left the room. I have done some of the sanding myself. Someone asks whether their work is safe, or whether their child should study a particular subject, and a broad answer is easier to give than a limited one. It keeps the conversation pleasant. It lets me remain useful without admitting how little of the future I can actually promise.

But I also made the darker answer too easy. In the previous essay, I treated a prominent executive's warning about job losses as an example of someone telling the truth. A warning can be candid and still be mistaken. A forecast does not become evidence of its own accuracy because it frightens the person making it, or because I think their colleagues would prefer it unsaid.

I cannot demand that optimists distinguish their hopes from findings and then exempt predictions that confirm my fears.

What I want from the people building and buying these systems is more specific. If a company has measured a result, say what was measured and where. If it is predicting a result, say what would make the prediction fail. If it expects to remove tasks, explain whether it also plans to remove positions. If it promises new opportunities, identify what it is doing to make them available to people who do not already have a place inside.

There will be things a person cannot responsibly disclose, and there will be things nobody yet knows. Neither requires a reassuring substitute. I cannot tell you that and I do not know leave a reader with a different problem from a promise that everything will work out. They are harder answers to give partly because they leave the uncertainty visible.

I still work in this industry. I still have a family to support. Those facts explain my position; they do not decide whether it is justified. The argument that caring people need to remain in the room has to show up in decisions: a claim narrowed, a failure admitted, a person trained, a use challenged. Otherwise it is something comfortable to say about staying.

I do not think the graduates owe us a warmer reception. I think we owe them enough specificity to make disagreement useful. They should be able to question a result, reject a deployment, or demand a different distribution of its benefits without first accepting our account of inevitability.

And I owe them the same care when I agree with them. Their fear is not my evidence. Their anger is not my mandate to describe a whole future on their behalf.

The next time someone asks me whether their work is safe, I want to be able to tell them what I know, what I am guessing, and what I would need to learn before saying more. It will be a smaller answer than the one I gave the graduating class. It may also be one they can use.