What Will Universities Look Like When AI Can Teach?

An AI tutor can explain calculus in three different ways, mark a practice essay, translate a paper, and draft working code before a lecturer has answered one email. The cost of another explanation is close to zero. The student can ask again without embarrassment and at two in the morning.

So what is a university still selling?

I began with the obvious answer: less than it used to. Recorded lectures already made information portable. Search made facts easy to retrieve. AI now handles the part that still required a patient human: an explanation shaped around the person asking. Add rising prices, student debt, fewer young people, and employers who say they care about skills, and the physical campus can look like an expensive delivery system for something the internet gives away.

That answer did not survive the evidence intact.

The short answer: universities will survive AI, but many lectures, assignments, and business models will not. AI makes routine teaching and routine production cheaper. It also makes several scarce things more valuable: credible proof that a student can work without assistance, supervised practice, expert judgment, laboratories, clinical settings, and relationships with people who can open a door.

The weak university sells content and a credential. The stronger one will sell proof, practice, people, and access to places software cannot reproduce.

The trouble started before AI

AI arrived in a system that was carrying old cracks.

US undergraduate enrollment grew from 14.48 million in 2003 to a recession-driven peak of 18.08 million in 2010, then fell to 15.40 million by 2022. Over the same span, the average published package of tuition, fees, room, and board rose from $20,860 to $27,673 in constant 2022–23 dollars. The student population grew 6.3%. The published charge grew 32.7% after inflation. The chart follows every year on the same index, so the rise, peak, and reversal are visible instead of being compressed into two endpoints.

• Undergraduate enrollment
• Published charges

Index: 2003 = 100 · Charges are in constant 2022–23 dollars

Undergraduate enrollment. 2003: 100.0, 2004: 102.1, 2005: 103.3, 2006: 104.8, 2007: 107.8, 2008: 112.9, 2009: 120.6, 2010: 124.9, 2011: 124.8, 2012: 122.5, 2013: 120.7, 2014: 119.4, 2015: 117.7, 2016: 116.5, 2017: 115.8, 2018: 114.8, 2019: 114.3, 2020: 109.7, 2021: 106.7, 2022: 106.3. Published charges. 2003: 100.0, 2004: 103.4, 2005: 105.7, 2006: 109.0, 2007: 110.1, 2008: 114.1, 2009: 117.0, 2010: 120.1, 2011: 122.5, 2012: 125.6, 2013: 128.4, 2014: 131.9, 2015: 135.3, 2016: 136.7, 2017: 138.0, 2018: 139.7, 2019: 141.2, 2020: 141.4, 2021: 136.8, 2022: 132.7.

Published price is not the same as net price. Grants cut the bill, and College Board reports that average net tuition for first-time, full-time students at public four-year colleges has fallen since its 2012–13 peak. That correction matters. It does not erase living costs, uneven aid, years spent outside full-time work, or the risk carried by borrowers. The New York Fed put US student-loan balances at $1.651 trillion in the second quarter of 2026.

The next pressure is already visible in high schools. WICHE's national projection peaks at about 3.86 million graduates in 2025 and falls to 3.37 million by 2041. That is a 12.5% decline, with much steeper changes in some states and regions. A college that depends on a steady stream of local 18-year-olds cannot fix arithmetic with a better admissions slogan.

2023: 97.50. 2025: 100.00. 2030: 94.50. 2035: 94.20. 2040: 89.30. 2041: 87.50.

Students have more routes around a four-year program too. Registered apprenticeships combine paid work with instruction. Community colleges issue short certificates. Large platforms sell certificates tied to software, data, and cloud work. The numbers do not show a clean replacement. They show an expanding edge around the degree.

STEM certificates below the associate level rose 58.7% between 2012–13 and 2021–22. STEM bachelor's awards rose 44.0%. Yet bachelor’s degrees still outnumbered those certificates by more than four to one in 2021–22.

• STEM certificates
• STEM bachelor’s degrees

Index: 2012-13 = 100. Latest awards: 96,637 certificates; 435,506 bachelor’s degrees

STEM certificates. 2012-13: 100.0, 2015-16: 124.9, 2018-19: 146.3, 2021-22: 158.7. STEM bachelor’s degrees. 2012-13: 100.0, 2015-16: 117.3, 2018-19: 136.6, 2021-22: 144.0.

Employer language has changed faster than employer behaviour. A Harvard Business School and Burning Glass Institute study, funded by the Schultz Family Foundation, examined firms that removed degree requirements. The net effect on non-degree hiring was only 0.14 percentage points. Fewer than one in 700 hires benefited from the change. “Skills first” can describe a job advertisement while the recruiter, interview loop, and promotion system keep sorting by degree.

This is the starting position. College remains valuable on average, but the price is harder to defend, the customer pool will shrink, and some alternatives are becoming credible. AI does not create those pressures. It raises their speed.

A chatbot can answer back

Universities have survived educational technology before. Radio courses, television, personal computers, the web, and massive open online courses all promised to separate learning from campus. They widened access to material. They did much less to replace the institution.

A recorded lecture solves distribution. It gives every viewer the same explanation and leaves the hard part to the viewer. When the student gets stuck at minute 23, the video continues talking.

AI can stop. It can ask what the student already knows, produce another example, turn a mistake into a new exercise, and repeat the cycle immediately. That changes the economics of practice and feedback, not just access to content.

The distinction between a general chatbot and a designed tutor is important. In a randomized study published in Scientific Reports, 194 Harvard undergraduates studied two physics topics either in an active-learning class or with a carefully built GPT-4 tutor. The tutor group had a median post-test score of 4.5 out of 5, compared with 3.5 in class, and reported higher engagement. The result is striking. Its limits are equally clear: two lessons, one course, a structured interface, and extensive prompts written to keep the tutor on task.

Another randomized study in PNAS followed nearly 1,000 Turkish high-school students learning mathematics. Access to ordinary GPT-4 improved practice performance by 48%. A guarded tutor designed to give hints instead of answers improved it by 127%. On a later exam with AI removed, the ordinary-chatbot group scored 17% below the control group. The guarded-tutor group performed about the same as the control.

That result broke the easiest version of the AI tutor story. A student can complete more work with a machine while learning less. Productive friction is part of education. If software removes every pause, wrong turn, and blank page, it can remove the moments when knowledge becomes the student's own.

The best use is therefore designed assistance. Let AI generate another problem, diagnose a misconception, translate difficult language, or give a first round of low-stakes feedback. Keep the student responsible for the reasoning. The quality of the course moves from “Did we upload the lecture?” to “What does the system let the student outsource, and what must the student still learn to do?” UNESCO's guidance makes the same basic demand: validate the tools, protect student privacy, and keep human agency in the design.

This also changes academic work behind the course. A lecturer can draft examples, build practice sets, simplify a reading, compare rubrics, and answer routine questions faster. That should free time for seminars, mentoring, and feedback on difficult work. It may also give managers a reason to increase class sizes and keep the savings. The technology does not choose between those outcomes. Budgets do.

A take-home essay no longer proves who did the work

The assessment problem is more immediate than the teaching opportunity.

In HEPI's annual UK undergraduate survey, reported use of generative AI for assessed work rose from 53% in 2024 to 88% in 2025 and 94% in 2026. The 2026 survey was sponsored by Kortext, and all three years rely on students describing their own behaviour. It is still hard to read the direction as noise.

2024: 53.00%. 2025: 88.00%. 2026: 94.00%.

A broader US study in Science covered more than 95,000 students at 20 public research universities. Thirty-seven percent reported using AI for coursework at least monthly. Nine percent said they had used it in ways they considered cheating. Different questions and samples produce different percentages, but both studies point to the same operational fact: AI is already inside assessed work.

The old take-home essay asked a grader to infer ability from a finished document. That inference was never perfect. Friends edited papers, parents helped, and paid services existed. AI makes outside production instant, cheap, and difficult to separate from ordinary spell-checking or feedback.

Detection will not restore the old bargain. Text detectors make errors, can be evaded through revision, and tend to punish unusual language patterns. TEQSA's 2025 assessment-reform guidance tells Australian institutions to design for evidentiary certainty instead of relying on detection. A polished file is no longer enough evidence by itself.

The practical response is already visible: oral defense, supervised work, live demonstrations, staged projects, process logs, and short questions about choices made along the way. A nursing student must still make a sound judgment with a patient present. An engineer should explain a design under questioning. A programmer can use an assistant, then diagnose a broken system when the assistant's answer fails.

This does not require banning AI. A good assessment can have two parts. First, let the student use the tools they will meet at work. Then ask for a separate demonstration of the knowledge needed to check, correct, and defend the result. The institution has to show both competent tool use and independent ability.

That is more expensive than collecting 300 essays. It requires staff time. It also produces the thing a degree will increasingly need to sell: trusted evidence.

The first job is getting harder to reach

Universities prepare students for jobs that are changing while the students are still enrolled.

Generative AI is strongest at the sort of bounded production often assigned to junior workers: draft the memo, summarize the documents, write the first version of the code, prepare the slides, classify the support tickets. Those tasks were useful to employers, and they taught new workers how the organization operates. If a senior employee can do more of them with software, the first rung gets thinner.

The early evidence needs careful language. A 2026 Stanford Digital Economy Lab analysis found no broad collapse in employment. It did find that employment for 22- to 25-year-olds in highly AI-exposed occupations was 19% below a counterfactual in which their employment had kept pace with less-exposed work. The change appeared mainly through fewer hires. The study is descriptive. It does not prove that AI caused the entire difference.

Census researchers found a similar pattern in payroll data: employment among 22- to 24-year-olds in the most exposed industry-state group fell about 12% over ten quarters after late 2022. They also found evidence of earlier divergence, which makes a simple causal claim unsafe.

Firm-level data provides a useful brake on the scary version of this story. Another Census working paper estimated that 18% of firms used AI, or 32% when weighted by employment. Two-thirds of adopters said AI assisted tasks. Only 2% reported an AI-related decrease in employment. The International Labour Organization likewise estimates that job transformation is more likely than outright replacement for most exposed workers.

The point is narrower than “AI is taking all graduate jobs.” Entry-level hiring may weaken first in some exposed occupations, and the work inside those jobs may change before occupation totals do. A program that teaches a 2022 workflow through 2030 is failing even if its subject name stays the same.

Students now need supervised contact with real work before graduation. A portfolio helps, but a portfolio can also be generated. A placement, clinic, lab, apprenticeship, or client project creates witnesses, constraints, and consequences. It turns “I made this” into “I did this work, under these conditions, and these people can explain my part.”

What universities still have

This is where the disappearance thesis runs into the university's deeper functions.

Employers still pay for degrees. In 2025, the Bureau of Labor Statistics reported median weekly earnings of $1,578 for full-time workers aged 25 and older with only a bachelor's degree, compared with $966 for high-school graduates with no college. That gap mixes education with selection, field, family background, and access to opportunity. It should not be read as the return from any degree at any price. It does show that the signal remains powerful.

At the most selective colleges, the signal is tied to networks and sorting. Opportunity Insights found that attending an Ivy-Plus college, compared with a highly selective public flagship for students near the admissions margin, increased the chance of reaching the top 1% of earnings and holding an elite position. Classroom content alone is an unlikely explanation. Peers, alumni, recruiting channels, and institutional reputation travel with the name.

Research is another moat. US universities spent $117.7 billion on research and development in fiscal year 2024, according to the National Science Foundation. The federal government supplied $64.6 billion. Much of this work depends on facilities, long time horizons, shared equipment, regulated environments, and teams built across generations of researchers. A chatbot can help write code or search papers. It cannot operate a particle accelerator, run a clinical trial, maintain a tissue bank, or accept legal responsibility for a hazardous lab.

Professional education has similar physical limits. Medicine, nursing, chemistry, civil engineering, and many trades require supervised practice and licensing. Sports add another dense system of facilities, coaching, media rights, donors, and identity. These are not side features that vanish when lecture notes become free. At many institutions, they are part of the reason students, money, and public attention gather in one place.

Then there is campus life. Moving away from home, finding friends, joining a team, working with people one did not choose, and becoming answerable to a community all carry value. The experience is uneven and sometimes romanticized. It can also be a serious part of growing up.

Software can imitate conversation. It cannot supply a friend who notices an absence, a supervisor who has watched a student work for six months, or a room where an argument changes because everyone has to answer in public. The internet weakened the university's control over information. It did not remove the demand for affiliation, status, and trust.

Some students will build their own education

The disruption case is real for institutions that mostly sell lectures and a generic credential.

By 2040, a student could assemble an education from an AI tutor, open course material, vendor certificates, paid projects, and examinations offered by trusted third parties. Employers in software, design, sales, and some business roles may care more about verified work than where the student sat. A local learning club could provide peers without carrying the cost of a full campus.

In this future, many regional colleges close or shrink into assessment and placement organizations. Degrees lose ground where licensing does not require them and work samples can prove ability quickly. The course becomes a service bought when needed, not a four-year sequence bought in advance.

The constraint is trust. Someone still has to verify identity, set a fair standard, secure the examination, hear the oral defense, and take reputational responsibility for the result. Unbundling removes the university only if another institution picks up those jobs.

Most universities will change what they sell

This is the base case.

The lecture moves into an AI-supported layer that students can use before and after class. Large introductory courses need fewer hours of repeated explanation. Staff spend more time on small-group problems, feedback, research, mentoring, and supervised practice. Students learn with AI and are examined both with it and without it.

Degrees become more modular. A student can stop with a certificate, return for another block, or combine online theory with local lab and workplace sessions. Universities keep responsibility for standards, records, research, licensing partnerships, and the learning community. They may teach more people across a lifetime while housing fewer of them for four continuous years.

This model is plausible because it lets universities keep their hard-to-copy assets while using software on the expensive repetitive work. It will still force difficult choices. Faculty roles change. Assessment consumes more time. Institutions must publish clearer evidence about completion, learning, debt, and employment. The schools that use AI only to cut headcount may preserve the brand for a while and hollow out the reason to pay for it.

The bad version gives rich students the humans

The darkest future is also easy to imagine.

Wealthy students get AI plus seminars, tutors, laboratories, travel, alumni introductions, and long conversations with respected scholars. Everyone else gets AI plus a crowded room and automated marking. Both products may carry the word “college.” They will not create the same opportunities.

This split would reverse the access promise of cheap tutoring. Software lowers the cost of explanation for everyone, yet the savings can make human attention look optional. The people who already have strong networks would keep intensive teaching and add better tools. Other students would receive efficient content with weak verification, thin relationships, and little practical experience.

That system could survive financially and fail socially. It would also make prestige more powerful, since the scarce product would be admission to the human network rather than access to information.

Five questions before paying for a degree

The question for a student is no longer “Is university worth it?” in the abstract. The answer depends on the program, the price after aid, and what happens there that cannot be downloaded.

I would ask five things. Does the program teach students to use AI without surrendering judgment? Does it verify what they can do independently? Does it provide work in labs, clinics, companies, or real projects? Has the curriculum changed as entry-level work changes? Does the likely outcome justify the net price and time away from paid work?

A strong program should answer with evidence: assessment design, placement records, licensing results, named partnerships, access to facilities, and graduates whose work can be checked. A weak one will point to course titles, campus photographs, and the average earnings of all degree holders.

The university will probably survive because education was never its only product. The lecture is less secure. Once explanation becomes abundant, the institution has to earn its price by doing the parts that remain scarce: judging work honestly, putting students in consequential situations, and connecting them to people who know what they can do.

The numbers behind the charts

The tables below reproduce the exact values behind each chart. The first table includes every annual observation from 2003 through 2022, not just the start and end points.

Sources: NCES Digest table 303.70 for fall undergraduate enrollment and NCES Digest table 330.10 for average charges at all degree-granting institutions. Charges include tuition, required fees, room, and board and are stated in constant 2022–23 dollars. Each index sets the 2003 or 2003–04 value to 100.

Source: NCES Digest table 318.45. Counts cover STEM credentials conferred by US postsecondary institutions. “Certificates” here means awards below the associate degree; it does not include every private or employer-issued credential.

Source: WICHE, Knocking at the College Door, 11th edition. These are projections, not observed enrollment. Counts include public and private high-school graduates in the 50 states and Washington, DC. Values are rounded only in the calculated index and change columns.

Source: HEPI Student Generative AI Survey 2026, which reports the comparable headline figures from the 2024, 2025, and 2026 surveys. These are self-reported UK undergraduate survey results. Kortext sponsored the 2026 report.