Exams will outlive everything else

If anything in education survives the next twenty years in a form we would recognize, it will be the exam. Lectures, homework, and the degree will all change shape. The exam will change shape too, but it will still be there, and it will matter more than it does now, because it is the one piece of the system whose value is tied directly to whether humans have value at all.

We are dealing with a technology that is smarter than us. Whether human knowledge will still be needed is a bigger question than I can answer. But if it is needed, it will be needed in specific places: work humans are still better at, and work humans must do even if a machine could do it. The competition for those jobs will be intense, and every applicant will have the same tools to get through anything that can be gotten through with tools. So the people hiring will need an evaluation a human has to pass, and they will need to know the human passed it alone. I am using "exam" to mean that: any evaluation a person has to pass by themselves, under observation, with something riding on the result. Whatever it looks like, that becomes the artifact everything else is measured against, and the stakes on it rise with the stakes on the job.

What changes

Teaching changes completely, and I'm not going to pretend I know how. Homework changes because unproctored homework stops meaning anything: if an assignment can be generated it will be, and a degree built on generated assignments loses value every year. So homework will be verified in some way, or it won't count.

The degree survives, and I think it becomes more valuable, but for a different reason than now. Its value won't come from the college being the best way to teach. It will come from the degree saying that a specific person went through a process, passed a series of exams independently, and can be relied on. The institutions that verify carefully will have a degree worth something, and the ones that don't will find out.

The exam itself gets more rigorous and more like the work. Multiple choice measured what could be measured when grading anything else at scale was too expensive. In 2024, when the first reasoning models came out, I recorded a note to myself about what that meant for assessment, and the example I used was small. What is three times three, with four choices, and one of them is nine. You pick nine, but the answer isn't complete until you show how you got there: three groups of three, or three plus three plus three, spoken out loud or typed alongside. Before, nobody could grade that at scale. Now a model can follow the reasoning and see where it went. The same goes for an essay. It shouldn't be graded on the text alone but on how the text came to be: what you read, where you went, your notepad. An exam that captures the process measures more, and it is harder to hand to a machine.

The objection

The strongest argument against all of this is that work itself is changing. If the future of work is a person directing AI agents while the agents do the work, then test the output, or evaluate people continuously on the job, and stop testing them in rooms.

I think that's half right. Tests of what a person can recall are going to lose most of their signal, and the market for them will shrink. And on-the-job evaluation is real; once someone has a track record, the track record is the credential. But somebody has to get the first job, the first license, the first visa, and for that there is no track record yet. There is only a gate. And when every applicant at the gate has the same agents, the difference between them is the part the agents didn't do, which is the hardest part to see. So what has to be verified shifts from "does this person know X" to "did this person, and not their tools, do the part that mattered." That is still an exam. It watches someone work instead of asking them to remember.

One exam, a few years out

For a nursing candidate, say, or an interpreter, or anyone whose work is a series of judgments under time pressure, I think it looks like this. The candidate sits down, alone, with a verified identity and a locked machine. Instead of a question bank, the exam describes a situation and asks them to handle it, in speech, the way an oral proficiency interview does now for language. The system asks follow-up questions based on what they said, the way a good examiner would, so no two candidates get the same script and there is nothing to memorize. Everything is captured: the answers, the reasoning out loud, the pauses, the environment. A model scores it against a rubric that was calibrated on thousands of human-rated sessions, and a human rater checks a blind sample. What the institution gets is not a score but a file: the score, the evidence behind it, and a record that this person, alone, produced it. Comparability comes from the protocol and the scale, not from everyone answering the same items, which is how language testing has worked for decades. It is harder to cheat because there is nothing to look up, and it measures the thing the job actually requires.

Why security survives with it

If the exam is the artifact, the exam will be attacked. An exam is the difference between a job and no job, between emigrating and staying, and breaching one is getting very easy: a second device, a person in the next room, remote access to the machine, and now a model that can answer anything in the time it takes to type it. People are very motivated, and if they can game the system to get ahead, they will.

So there are two futures. In one, the stakes get so high that we give up on remote exams, bring everyone into a room, and watch them. That will happen for some exams, and it should. But a room for every exam is expensive and slow and doesn't fit the world we've built, and rooms leak too. In the other future, remote exam security gets very good, and I mean the whole job, not a human watching a webcam. Identity confirmed, the machine locked, every signal analyzed at once (keyboard, mouse, video, audio, gaze), and a person with judgment making the call when the signals say something is wrong, because a model should never be the one accusing someone. It also has to be fair, which remote proctoring has not always been: it has to work for people with disabilities, on bad internet, and in rooms that aren't offices, and it has to treat the test-taker's experience as a hard limit rather than a cost. If it gets there, it survives.

And the part of it that lasts is not the catching. Catching cheaters is an arms race, and every detector loses value once people learn what it looks for. What lasts, and what gets more valuable as the models get better, is proof that a specific human demonstrated a specific competence, by themselves, in a way that would hold up if challenged. In a world where any answer and any credential can be generated, that proof is the scarce thing, and producing it reliably, at scale, without making the test-taker miserable, is the whole job.

Three predictions

By 2030, the highest-stakes licensing exams, the ones that gate medicine and law, will have moved back toward test centers, while professional and workforce certification will have moved the other way, toward remote exams with real verification, and will be the fastest-growing part of the market. If all high-stakes exams end up in rooms, I was wrong about remote security.

By 2030, the cost of writing an exam question will be close to zero and the money spent per exam on verifying who took it will have gone up. If verification spend falls as item cost falls, I was wrong about what's scarce.

And by 2030, most degrees will still be awarded on the basis of exams passed under observation, not on coursework, because coursework will no longer be trusted. This is the one I'm least sure of, because institutions change slowly and coursework is where the tuition is.

If humans keep having value in the workforce, the exam is what tells us which humans, and that is why it outlives everything else. The shape will change and the need won't.