The Classroom Meets the Machine: AI and the University

Three fronts, one machine

Higher education is not being disrupted by AI in one place. It is being disrupted in three places at once, and they are usually discussed as separate stories when they are the same story told from different rooms. The first is how students learn: what they do, unsupervised, at a laptop at eleven at night, which has shifted from occasional AI use to near-universal AI use inside two academic years. The second is how they are assessed: the take-home essay, the default instrument of undergraduate evaluation for a century, has lost much of its evidentiary value, and the institutional response is a partial retreat to the room, the pen, and the proctor. The third is what students choose to study in the first place, and here the picture cuts against the easy AI narrative in one direction and confirms it in another.

This study belongs to a series on how the generative-AI shift is repricing learning and work, and academia is the clearest laboratory for the series’ wider claim: this technology wave lands hardest on educated, cognitive, white-collar work, the opposite of where robots and early software automation landed. A university is where that cognitive work is taught, tested, and priced into a credential. If the inversion is real, the classroom should show it before the labor market fully does, in what students are told to master and in which of those disciplines still pay off once they graduate.

The throughline argued here is specific. The decline of the humanities is real, large, and structural, but it predates generative AI by more than a decade and should not be laid at AI’s door. What is genuinely new, and genuinely surprising, is the reversal happening at the top of the STEM funnel: after fifteen years of "learn to code" as the default safe advice, computer-science enrollment turned over in the same year that new computer-science graduates began posting unemployment rates above several humanities majors. That reversal, not the humanities story, is the AI-era signal this study is built to isolate.

The decline of the humanities is real and structural, but it predates the chatbot by a decade. What is new is the reversal at the top of the STEM funnel.

How students actually work now

The clearest, best-corroborated number in this entire study is also the simplest. HEPI’s annual survey of UK undergraduates found that 53% had used AI for assessed work in 2024. A year later that was 88%. In 2026 it reached 94%. Overall AI use, including unassessed study, tracks close behind: 66%, then 92%, then 95%. In three admissions cycles, using AI in some form for university work went from a majority practice to something close to a universal one.

The finer detail is where the interesting movement is. The share of students specifically using AI to generate text for them, the closest thing to "AI wrote my essay", rose from 30% in 2024 to 64% in 2025, then fell to 56% in 2026 even as overall use kept climbing. Read alongside qualitative reporting from the same survey, the most defensible interpretation is a shift in what AI is used for: less as a ghostwriter producing a final draft, more as an explainer and summarizer working through source material, drafting arguments, and checking reasoning before the student writes. Institutional support for that shift is growing in parallel. The share of students using AI tools that their institution actually provides, rather than a personal ChatGPT account, rose from 9% in 2024 to 23% in 2025 to 38% in 2026, evidence that universities are trying to move from prohibition to provisioning.

This is not a UK-only pattern. The Digital Education Council’s survey of students across 16 countries found 86% using AI in their studies and 54% at least weekly, the same order of magnitude as the HEPI numbers from a different instrument and a different population. Pew found something similar lower down the pipeline: 26% of US teens said they had used ChatGPT for schoolwork in 2024, double the 13% who said so in 2023, even though only 18% of teens think using it to write an essay is acceptable, against 54% who think using it to research a topic is fine. Students, in other words, are drawing their own line between AI as a research tool and AI as an author, roughly where the HEPI shift in usage pattern also points.

The clearest evidence that AI use clusters by discipline, rather than spreading evenly, comes from Anthropic’s own usage logs. Computer-science students accounted for 36.8% of student conversations on Claude, against a 5.4% share of US bachelor’s degrees, a roughly seven-fold overrepresentation. That is a second, quieter data point for the study’s throughline: the students most fluent with AI are disproportionately the ones studying the discipline the technology is now most directly repricing.

The assessment crisis

If AI use climbed this fast, the instrument built to catch it had to fail this fast too, and it did. Vanderbilt University disabled Turnitin’s AI-detection feature indefinitely in August 2023, within a year of ChatGPT’s release, citing the tool’s unreliability. Australian Catholic University went the other direction first and then reversed: it logged roughly 6,000 academic-misconduct cases in 2024, about 90% of them AI-related, then abandoned the Turnitin AI-detection tool as ineffective once the case volume made clear the tool could not be trusted at scale.

The technical reason detection keeps failing is documented, not anecdotal, and it has a bias problem layered on top of an accuracy problem. A widely cited Stanford study found that AI-detection tools flagged 61% of essays written by non-native English speakers as AI-written, against a rate near zero for essays written by native speakers, because non-native writing patterns statistically resemble the more formulaic sentence structures the detectors were trained to catch. Even a detector with a genuinely low 1% false-positive rate, applied across a body of roughly 75,000 papers a mid-size institution might process in a year, implies on the order of 750 wrongful accusations annually, a number worth sitting with before any institution recommits to a detection-first strategy.

The tension here has two real sides. One side says the technical case against detection is closed: the tools are unreliable, biased against exactly the students who can least afford a false accusation, and easy to defeat with light paraphrasing regardless. The other side says that abandoning detection does not make the underlying problem disappear, it only removes the deterrent, and institutions still need some mechanism, however imperfect, to preserve the meaning of an unsupervised written assignment. Neither Vanderbilt’s withdrawal nor ACU’s reversal resolved that tension. They mainly demonstrated that the first-generation answer, software that scores a paper’s probability of AI authorship, did not survive contact with real student writing.

A detector with a 1% false-positive rate, run across 75,000 papers a year, implies roughly 750 wrongful accusations. That is the cost of getting detection wrong.

Blue books are back, and the argument over whether that helps

With detection discredited, the institutional response has largely been to change the assessment rather than police it, and the most visible version of that is a return to the room. Reporting from 2024 through 2026 describes surging demand for physical blue books, the handwritten examination booklets that had been fading out of American higher education, at Texas A&M, the University of Florida, and UC Berkeley among others. The shift is broader than blue books specifically: oral examinations, handwritten in-class work, and device-free testing rooms are all reported to be expanding, and UK accountancy qualifications moved a portion of their assessment back to in-person exam halls for the same reason.

The logic is straightforward. A laptop with an internet connection cannot be trusted to certify that the writing in front of an instructor is the student’s own reasoning, so the response removes the laptop. What the logic does not settle is whether the substitute is actually better. The counter-evidence, held with equal weight here, has two parts. First, students adapt to in-person constraints too: institutions that have tightened proctoring report new forms of circumvention, from smuggled devices to memorized AI-assisted outlines, so the move does not eliminate the underlying incentive to use AI, it only changes where the workaround happens. Second, and more fundamentally, a three-hour handwritten exam tests recall and composition speed under time pressure, not the research, synthesis, and iterative-revision skills that most of the disciplines value in the world outside the exam hall, and that AI itself is reshaping the value of. Whether the blue book is a genuine fix or a nostalgic stopgap that measures the wrong thing well is not resolved by any of the sourcing behind this reversal. It is a live institutional bet, not a settled result.

Does an AI tutor actually teach?

Set against the assessment crisis is a second, more optimistic body of evidence: AI, deployed as a tutor rather than a ghostwriter, appears to teach faster than conventional instruction in at least some controlled settings. A randomized trial run by the World Bank and Stanford researchers in Nigeria gave roughly 800 students six weeks of access to a GPT-4-based tutoring tool and measured a 0.31 standard-deviation learning gain, a result the researchers framed as equivalent to roughly one and a half to two years of typical schooling compressed into six weeks. A separate randomized trial at Harvard, led by Kestin and colleagues in 2025, found that students using an AI tutor learned roughly twice as much per hour of study time as students in an active-learning classroom, the gold-standard comparison condition in education research for the past two decades.

Those two results, taken together, look like the long-sought "two-sigma" effect first described by the educational psychologist Benjamin Bloom in 1984, the finding that one-to-one human tutoring could move an average student from the 50th to roughly the 98th percentile, an effect no scalable classroom intervention had matched in forty years. The AI-tutoring literature is not there yet, and a 2025 systematic review of the field is explicit about why: effect sizes shrink substantially once an AI tutor is compared against a strong modern human-tutoring or intensive small-group control, rather than against an ordinary lecture. Bloom’s original two-sigma claim has itself faced replication difficulty for decades. The fair summary is that AI tutoring produces real, trial-measured gains against realistic classroom baselines, but the more dramatic two-sigma framing depends heavily on which control condition a study picks, and the field has not converged on an answer.

The scale question matters as much as the effect-size question. Khanmigo, Khan Academy’s AI tutor, reached approximately 795 US school districts and about 770,000 US students by the 2024-25 school year, evidence that the deployment, not just the research trial, is now large enough to matter at a system level. Whether that deployment reproduces the trial-scale gains once it runs at the scale of a full district, with uneven teacher support and uneven student engagement, is the next open question, and it is one none of the cited studies were designed to answer.

Faculty, policy, and the retreat from outright bans

Faculty sentiment has not caught up to student behavior, and the gap between the two is itself a data point. Surveys put roughly 45% of faculty holding an overall negative view of AI in higher education, against 34% positive, and a joint AAC&U and College Board survey found 95% of faculty fear student overreliance on AI and a diminished capacity for critical thinking among students who use it. That is close to consensus concern from the group closest to the classroom, and it sits uneasily next to the HEPI finding that 94% of the students those same faculty teach are already using AI for assessed work.

Institutional policy is nonetheless moving away from prohibition, not toward it. An analysis of more than 31,000 course syllabi from 2021 through 2025 found faculty shifting toward more permissive AI-use policies by autumn 2025: 65% of syllabi still prohibited AI for higher-order reasoning tasks, but only 20% prohibited it for coding assistance and just 17% for proofreading, a clear hierarchy in which faculty are comfortable with AI on mechanical tasks and wary of it on tasks closer to the discipline’s core reasoning. Ohio State went further, making AI fluency a graduation requirement starting with the fall 2025 cohort, targeting full rollout by 2029, an explicit bet that AI literacy belongs alongside writing and quantitative reasoning as a baseline competency rather than a banned shortcut.

95% of faculty fear student overreliance on AI. 94% of students already use it for assessed work. The policy gap between those two numbers is where the crisis actually lives.

The humanities decline: real, and older than the chatbot

The most commonly cited casualty of AI in higher education is the humanities, and the underlying numbers are genuinely stark. Humanities bachelor’s degrees fell to 165,489 in 2024, the fewest conferred since 1991, and roughly 30% below the field’s 2012 peak. The humanities’ share of all bachelor’s degrees awarded fell from 13.1% in 2012 to 8.4% in 2024. Individual disciplines show the trend even more sharply: English degrees fell 33.1% over the decade, and history fell roughly 25%.

The critical fact about this decline is its timing. It has been continuous since 2012, a full decade before ChatGPT existed as a public product, tracking a combination of forces that predate generative AI entirely: rising tuition and the resulting pressure to major in something with a legible career payoff, a post-2008 shift in how families weigh the return on a degree, and a long relative decline in the cultural and institutional prestige of humanities disciplines. Attributing the 2012-to-2024 slide to a technology that reached mass adoption only in 2023 gets the chronology backward. AI may well be compounding the pressure now, by making the argument for a vocationally legible major feel more urgent, but the decline itself was already twelve years old when generative AI arrived, and the data does not support treating it as AI's doing.

The new signal: the top of the STEM funnel turns over

If the humanities decline is the old story wrongly relabeled as AI's doing, the genuinely new AI-era signal sits somewhere else entirely, and it is the least intuitive number in this study. After more than a decade of continuous growth, undergraduate enrollment in computer and information science fell 8.1% in fall 2025, down to approximately 606,000 students, the field’s first broad enrollment decline in ten years. Data science, a closely related but distinct major, kept growing over the same period, reaching roughly 80,000 students, suggesting the pullback is concentrated in the traditional CS major rather than in quantitative and data-adjacent study generally.

The labor-market data helps explain why. The New York Fed’s early-career outcomes series found computer-science new graduates carrying a 6.1% unemployment rate and computer-engineering new graduates roughly 7.5%, both above the roughly 4.8% average across all college graduates, and both above several humanities majors, including art history at about 3.0%. For a decade, "learn to code" was the default advice given to anxious undergraduates and their parents as the safe, AI-proof, recession-proof major. The 2025 data inverts that advice for the first time on record, at exactly the moment employers are reporting that AI coding assistants can substitute for a meaningful share of the routine software-development tasks that entry-level computer-science graduates used to be hired to do.

This figure carries a real caveat and this study states it plainly. The New York Fed’s major-level unemployment breakdown is a relatively small, survey-based sample with wide confidence intervals at the individual-major level, and a single year of data is directional evidence, not proof of a durable trend. The CS enrollment decline is a firmer number, drawn from a broader institutional survey, but even it is one data point after a decade of growth and could reflect a temporary correction rather than a structural reversal. What can be said with more confidence is the direction: two independent data sources, one on what students are choosing and one on what happens to them after graduation, both point the same way in the same year, and that year is the first year AI coding tools were broadly deployed inside professional software teams.

The limits of this reading

Several cautions bound everything argued above. The HEPI trend line, the strongest and most consistent dataset in this study, describes UK undergraduates specifically; the Digital Education Council and Pew figures corroborate the direction internationally and among US teens, but the precise year-over-year magnitude is a UK finding generalized with care, not a global measurement. The institutional case studies, Vanderbilt’s withdrawal of AI detection and Australian Catholic University’s reversal, are drawn from secondary reporting on named institutions rather than from primary university disclosures, and are tiered accordingly. The AI-tutoring effect sizes are real trial results but genuinely contested in magnitude: the same intervention looks close to Bloom’s two-sigma effect against a weak control and far more modest against a strong one, and the field has not settled which comparison is the right one to generalize from.

The newest and most consequential finding in this study, the reversal in computer-science enrollment and early-career outcomes, is also the least mature. It rests on one year of enrollment data after a decade of growth and on a small-sample, wide-interval unemployment series. Both deserve at least another admissions cycle and another graduating cohort before the reversal can be called durable rather than a single-year correction. And the WEF skills projection and the India AISHE gap are included deliberately as bounded evidence: one a transparent scenario, not a measurement, and the other a plain acknowledgment that this study could not confirm what it set out to check.

What survives all of that qualification is still a coherent picture. Student AI use for coursework is not a niche behavior anymore; it is close to universal and well corroborated across independent sources. The instrument built to police it has failed on its own technical terms, and the institutional response is still being improvised in real time. And the discipline that spent a decade as the default safe advice for an uncertain labor market is, for the first time on record, no longer behaving like one.