Table of contents
ARTIFICIAL INTELLIGENCE & SOCIETY
PART 12
People are having the wrong fight about college and AI. Cheating and showing the work. Bans and hidden text. A clean file and a closed-book exam. Sometimes those words are close enough for a faculty meeting. Sometimes using them completely changes what a degree is for.
If I am going to talk about whether four years and a pile of debt are still worth it, I need to keep the assignment straight. A student replaying a step with the tool turned off is one test. Building real skills for today’s workplace is a different test.
Direct Answer
The point of college is to prepare students to use the tools available to them, and to prepare them for the real world after college.
A degree stays worth it when it builds real skills for today’s workplace.
These skills currently include planning, execution, delegation, awareness of what a group is doing, and coordination. They include leadership, communication another person can act on, creativity, and judgment when the information is incomplete. Fluent use of the tools (the ones a workplace actually has) belongs on that list, and so does the willingness to say the machine is wrong.
Discernment is telling a useful result from a confident wrong one. Decision-making is choosing a direction when more than one option is open.
Will that list stay fixed? No. These are the skills for today’s workplace, but they will move when the work moves.
They can be built in a lab, a shop, a studio, or a placement when people have to make something together, but the result still has to be real.
They can be built when a professor teaches the next tool because it raises the output, but the grade is still what got finished.
Is a shared class project the only way to build them? No. I think it is a highly effective way, but a lab, a shop, or a placement can build the same skills when the work calls for that instead. The class picks one goal, gets a budget, and has to delegate.
A lecture plus a diploma is a weak offer once the lecture is free and the first routine jobs are thinning. Showing the work does not build these skills.
Showing the work and finishing the work
Showing the work and finishing the work are not the same assignment.
A finished file can look clean and still tell you nothing about whether a group can plan, split the jobs, stay inside a budget, and finish. Schools have started treating the path as the product: who typed the paragraph, who opened the chatbot, and who can redo the problem after the window closes. Did the student do it alone? The path they took is not the grade.
If a tool understands a step better than we do, and that step does not actually need to live in a person’s head, hand the step to the tool. Keep the human work on what the tool does not carry: planning, execution, delegation, building awareness, coordination, discernment, choosing a direction, and the other skills that decide whether a project gets finished. Policing the scratch work does not teach those skills.
Earlier pieces in this series asked who gets the money when machines do more of the work, and what happens if a lot of required jobs thin out. This one is about the classroom, because a lot of people are still being told to buy a lecture and call it preparation.
A shared project is a highly effective way
Labs, shops, studios, and placements can all build the skills above, when the alternative is a semester of lectures with a paper at the end. A shared class project is one version of that work. I think it is a highly effective one.
The class should be asked to finish something.
Give the whole class one project. Make it big enough that nobody finishes it alone. They have to delegate, and they have to complete it together. Give them a budget. The budget has to balance against the goal they actually want, which means cutting something when the numbers do not.
Who picks the goal? The class does. Start with a broad list. Narrow it. Keep narrowing until the class can vote. That vote is the class goal, and it is practice at choosing a direction, but more than one option can look right when the list is still broad. Discernment is telling which option is worth the budget. Then use the tools available to accomplish it.
Through the course, the professor teaches new tools as they go, the ones that can raise the output of this project, not a tool unit that exists so it can be tested later.
Nothing is forced.
A student can use the tool or not. The grade is what got finished, not a reconstruction of how they got there.
The sequence is ordinary.
- A broad list of goals, then a narrower list, then a vote.
- One class project, with delegation, not sixteen private papers.
- A budget that has to balance.
- Tools taught when they help the work.
- A grade on what got finished.
Does every course have to copy that design? No. A lab can do this work when the task is physical, but the design is still very different from a lecture, a chatbot ban, and a final that asks who can still do step three alone.
For example, a class could start with a wide list: a campus service, a local reporting project, a product, a clinic workflow, a small build. They cut the list until one goal can win a vote. They price it. They split the roles. When a tool would make the drafting, the scheduling, the research, or the budget math faster, the professor shows that tool and gets out of the way. At the end, the thing either exists or it does not.
A YouTube lecture cannot run that room. A chatbot will not notice that the budget is fantasy or that two people did all the work. A campus can, if someone is in the room and the grade is the outcome.
A closed exam is a different test
Researchers gave nearly a thousand high school math students access to GPT-4 tutors, then took the tool away for the exam. [1]
During practice, an unrestricted chatbot (the kind that behaves like a normal chat window) raised scores about 48 percent above students who had no AI. A second tutor, built to give hints instead of answers, raised practice scores about 127 percent. [1]
On the later exam, with the tool gone, the students who had used the open chatbot scored about 17 percent worse than students who never had it. The hint-based tutor mostly erased that damage. It did not produce a clear gain on the exam either. Students on the open tool often asked for the solution and copied it, and a lot of them did not feel that their learning had slipped. [1]
Read those two results next to each other. The practice looked better. The closed exam looked worse for the open tool.
So the closed exam and the project are not the same assignment.
That study measured a test with the tool removed. I think a university should not organize itself around that test. If you only count the score after the tool is taken away, you are grading a different job. So a degree built around that job is aimed at the wrong task. Hand off the step that does not need to be understood by hand, and grade whether the people produced the outcome.
Grading the path spends the relationship
Harvard’s Center for Digital Thriving put the classroom version of the cheating fight in a report called “An AI Policy Isn’t a Playbook.” Coverage describes a two-way suspicion. Teachers suspect undisclosed AI use and often cannot prove it. Students fear a false accusation and cannot prove they did the work themselves. [2]
The chase gets technical fast. A teacher hides nonsense instructions in white text so a pasted prompt will confess. Students photograph the page so the hidden text never reaches the bot. [2] The time goes into proving who typed the paragraph, and nobody has built anything. The same reporting says fewer than one in five teachers work in a school with a specific generative-AI policy. About a quarter of teachers have already faced an AI dilemma, and nearly three-quarters of those dilemmas were about cheating. More rules, by themselves, are not a teaching plan. [3]
Beck Tench, a co-author, has argued that educators do not need to be AI experts to coach students. They need a conversation about goals, pressure, and when the tool helps. Bans tend to push the use underground. [3]
That chase exists because the grade is the path. If the grade is the result, the hidden-text trick is beside the point. The professor’s job is to teach the next tool when it would help the project, and to judge what the class finished.
The first job was the classroom after the diploma
The labor market is already sorting some majors by a different measure. Economists at the Federal Reserve Bank of Dallas linked Texas university records to in-state employment and pay. [4]
After ChatGPT’s release in 2022, a major with a 10-percentage-point higher share of automatable tasks saw about a 1.7-percentage-point relative drop in graduates finding Texas work within a year. Among those who did find work, first-year pay in the more exposed majors fell about 5 percent relative to less exposed fields, comparing 2021 with 2024. [4]
Computer science, computer engineering, and languages sat toward the exposed end. Nursing, education, and psychology sat toward the sheltered end. [4]
Graduates from more exposed majors in 2024 were also more likely to go straight to graduate school, about 1.4 percentage points for that same 10-point gap. Most of them did not switch fields. About two-thirds of the computer science graduates who went back enrolled in computer science again. [4]
New workers with master’s degrees in exposed fields saw relative pay declines similar to the four-year graduates. From fall 2024 to fall 2025, Texas undergraduate enrollment in more exposed majors fell about 4.8 percent for each 10-point gap in automatable tasks. [4]
I wouldn’t turn a Texas administrative study into a claim about the whole country. There is a limit to what one state’s file can prove, but the pattern is still stronger than a pile of anecdotes. It does not follow every graduate who left the state, and it does not prove that AI explains every soft spot in hiring.
The first place this shows up is often a job that is never opened, and another year in the same exposed subject is weak protection. If that first job is never posted, the graduate does not get a place to practice delegation. Entry work used to be where people learned judgment on a real product. A class project with a budget is one way to get that practice before the diploma, instead of hoping a first job still exists to teach it.
A job on a chart is not a job you can reach
McKinsey Global Institute’s September 2026 report, “Workforce in motion,” models a large reshuffle rather than a simple shortage of jobs. [5]
In the base case, automation could reduce labor demand by the equivalent of about 36 million jobs over the next decade, while growth in the AI supply chain and the wider economy could create demand for more than 40 million. About 25 million workers in roles touched by automation might stay in place because their industries expand. [5]
Roughly 11 million people (about 7 percent of current employees) might need to change occupations by 2035. The range runs from about 6 million to 16 million depending on how fast the tools spread. Quartz’s account of the report puts the useful line in plain speech: the next decade’s challenge is mobility, not scarcity. [5]
These are scenarios, not a headcount of layoffs that already happened.
But only about one in seven workers who need to move has what McKinsey calls a direct path: little retraining and no pay cut. About four in ten face a winding path that needs real reskilling. Nearly half face an unpaved path. About 85 percent of growing jobs ask for some credential or certification. [5]
More than three-quarters of the people who may need a new occupation are in office and administrative support, retail and sales, or transportation and logistics. Workers in lower-wage jobs are about 7.6 times as likely as higher-wage workers to need a full occupation change. Meeting that pace would mean about 770,000 occupational moves a year, roughly 3.6 times the recent historical rate. [5]
A replacement job in another column of a spreadsheet does not get a person into that job. Can the person actually reach it? If you are the person who has to move, there may be a growing job somewhere, and the route can still require a license, a pay cut, a move, or months you do not have. The rent is still due while that pathway is “winding.”
The first wave is already visible in work people used to call safe because it happened at a desk. For example, bookkeeping and routine accounting tasks, lead lists, reception, executive-assistant scheduling, storefront tax preparation, and a large share of paralegal document work are built from repeatable steps. Dallas Fed researchers describe the early hit as a pullback in hiring for white-collar tasks the new tools can do, which lands first on people trying to enter. [4]
I expect that list to widen. Calling it the whole story would be a guess. Calling it a rounding error would ignore the evidence we already have.
Don’t aim the degree at the fallback counter
When a professional door sticks, someone usually offers the old backup: go work the counter, flip burgers, take the late shift. I don’t take much comfort from that advice.
The counter is a stack of repeatable tasks. Take the order, cook the same items, bag them, wipe the same stations, say the same script. Because the steps repeat and the exceptions are rare, companies have spent years trying to move pieces of that stack onto kiosks and machines.
I think that backup repeats the same problem as the desk job. I’m not claiming every restaurant job ends on a date certain. I am claiming it is a bad plan to spend four years and a large debt on a diploma whose unspoken backup is that stack of tasks. If the professional job and the fallback job are both made of tasks a trained system can absorb, the student did not buy safety. The student bought time that will not hold.
Staying ahead means aiming at work that still needs a person who can be trusted, blamed, or physically present, and aiming there with the current tools in hand. Care that depends on reading a real body, skilled work on a site that cannot be fully scripted, leadership of other people, and roles where the law still wants a human name on the decision are harder to hand off entirely. Parts of each will still be automated. The remainder is where planning, coordination, and responsibility live.
Use the tools. Teach them while the work is happening.
I use these systems every day.
Limiting students to protect a lecture format, while the same tools sit open on their phones, does not preserve learning. It keeps the lecture in place.
AI can take a group from an idea to a working version very quickly. Schools should not block students from that. They should teach them how to use it and aim it at the outcome the work is supposed to reach.
Anyone trained to use the tools can raise the level of output: faster drafts, more passes at the same problem, clearer explanations, a budget that can be rebuilt in an afternoon. The university’s job is to teach that in the room, while the work is happening, when it helps. Not to require a confession about which sentences were touched by a model.
The skills that hold up are the ones people file under soft and then underpay: planning, execution, delegation, awareness of what the group is actually doing, and coordination across people who are not on the same screen. Add discernment, which is telling a useful result from a confident wrong one, and decision-making, which is choosing a direction when more than one option is open. Add leadership, communication another person can act on, and creativity that solves a problem the template did not anticipate. Add judgment when the information is incomplete, and the willingness to say the machine is wrong. Care for a person in the room belongs there too, because responsibility does not upload cleanly.
The job market is already pricing a version of this. In McKinsey’s reading of postings, demand for AI fluency rose about elevenfold between 2022 and 2026. Demand for adaptability rose about fivefold. Demand for resilience, curiosity, and willingness to learn roughly tripled. [5]
PwC’s 2026 survey of 49,364 workers across 48 countries puts a social shape on the same split. About 64 percent said they had used AI at work in the past year. PwC still classifies only 14 percent as front-runners: people with scarce skills and strong AI capability. The largest group, 56 percent, is what it calls the engine room: less scarce skills, further behind on AI, and less likely to be rewarded for learning. Only about two in five of that majority said they can get the learning resources they need. [6]
Across the whole sample, access to learning and development fell from 59 percent to 51 percent in a year, even as AI use rose. Daily users were more likely to trust top management and to feel able to learn. [6]
That survey is self-reported. It measures experience and perception, not a lab test of productivity. It still matches the classroom problem. The advantage stays with people who already have someone willing to teach the tool. A campus that withholds practice leaves everyone else further behind. A campus that teaches the tool while real work is underway, in a lab or on a project, and still grades the result, gives ordinary students a way in.
What the tuition still has to buy
None of this means a motivated person on a couch with free courses equals a student in a serious program. Free video does not give you a lab, a shop, a clinic placement, or a professor who will tell you the build failed.
It does not give you a class that has to pick one goal, split the work, and live with a budget.
It does not give you a credential that employers still use as a filter, and McKinsey’s pathway work says most growing jobs still ask for one. [5] A determined person can make themselves do the work without a class. Most people will not. A video does not notice if you skipped a week, if the budget is wrong, or if two people did all the work.
The question for a student is whether a given university delivers practice at those skills, and the tools, at a price they can justify. Updating a syllabus every time a vendor ships a feature is not that delivery. Neither is a prestige name on a transcript after four years of lectures the student could have watched for free.
If the program never puts students in a setting where they have to use the tools and finish real work, the diploma mostly shows that the student had access to lectures. Those lectures are already free.
The grade should be the result. The tools are allowed. The work that still belongs to people is choosing the direction, planning it, delegating it, using discernment on what the tools produce, and standing behind what got built.
We do not need to spend four years proving a student can redo a step after the tool is turned off. We should spend that time building the skills the workplace still needs. A shared project is a highly effective way to do that, and so are a lab, a shop, and a placement. If the degree never asks students to choose a direction, use the tools, and finish real work, we can spend the whole time talking past the jobs they will actually have to do.
Frequently Asked Questions
Should colleges ban AI so students still learn?
No. A ban pushes use out of sight. Teach the tools when they raise the output of real work, do not force them, and grade the result. In a high school math experiment, an unrestricted chatbot improved practice scores and left students worse on a later exam with the tool removed. That exam measured the work with the tool taken away. If the job includes the tool, that is a different test. [1]
Are these job numbers observed layoffs?
No. McKinsey’s figures are modeled scenarios through 2035. The base case has automation reducing demand by the equivalent of about 36 million jobs while other forces create demand for more than 40 million, with roughly 11 million workers needing a new occupation. The range is about 6 million to 16 million. The useful warning is about mobility. Most of the people who must move do not have an easy path. [5]
Did AI already cut hiring for some college majors?
In Texas public universities, graduates from majors with more automatable tasks became less likely to find in-state work within a year, and those who worked earned less in the first year, relative to less exposed majors, after ChatGPT’s 2022 release. Graduate degrees in the same exposed fields showed a similar relative pay pattern. The data do not cover every state or prove AI caused the entire soft patch in youth hiring. [4]
Which skills are worth building on purpose?
A degree stays worth it when it builds skills for today’s workplace. Those currently include planning, execution, delegation, awareness, coordination, discernment, decision-making, leadership, communication, and creativity, plus judgment and fluent use of the tools. They can be built in a lab, a shop, a studio, a placement, or a shared class project. I think the project is a highly effective way, but it is not the only way. Job ads have been asking more often for AI fluency, adaptability, resilience, curiosity, and willingness to learn. [5] [6]
What can a university still offer that free courses cannot?
A room, a lab, a shop, a placement, and a professor who teaches the next tool when it helps. A shared project, with a budget and real delegation, is a highly effective version of that work, and it is not the only one. A credential many growing jobs still require. Those are real advantages only when the school delivers them. A lecture plus a diploma is a weak offer once instruction is cheap. [5]
References
-
Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman. “Generative AI without Guardrails Can Harm Learning: Evidence from High School Mathematics.” Proceedings of the National Academy of Sciences, 2025. https://www.pnas.org/doi/10.1073/pnas.2422633122
-
GovTech. “AI Cheating Concerns Harm Student-Teacher Relationships.” Coverage of the Center for Digital Thriving report “An AI Policy Isn’t a Playbook” (Beck Tench and colleagues, Harvard Graduate School of Education). https://www.govtech.com/education/k-12/ai-cheating-concerns-harm-student-teacher-relationships
-
WWMT. “AI fuels cheating concerns in schools, mistrust between teachers and students: report.” Coverage of the same Center for Digital Thriving report, including school-policy and teacher-dilemma findings and Beck Tench on coaching rather than rulebooks. https://wwmt.com/news/nation-world/ai-fuels-cheating-concerns-in-schools-mistrust-between-teachers-and-students-report
-
Samuel Dodini and Tucker Smith. “AI Plays a Role in Weak Labor Market for College Graduates.” Federal Reserve Bank of Dallas, September 22, 2026. https://www.dallasfed.org/research/economics/2026/0922
-
Cris Tolomia. “McKinsey says AI might push 11 million U.S. workers into entirely new careers by 2035.” Quartz, September 30, 2026. Report of McKinsey Global Institute, “Workforce in motion: Skills and pathways to future jobs in the United States,” September 29, 2026. https://qz.com/mckinsey-ai-automation-workers-career-transitions-2035-093026
-
PwC. “Companies risk losing their most AI-savvy employees while leaving majority behind.” Press release for the 2026 Global Workforce Hopes and Fears Survey, September 29, 2026. https://www.prnewswire.com/news-releases/companies-risk-losing-their-most-ai-savvy-employees-while-leaving-majority-behind-pwc-302891362.html