The AI Campus Experiment

A student can now produce a persuasive essay without having struggled to decide what the essay should argue. A few revisions can make the prose sound personal. The finished work may satisfy the assignment while revealing very little about what the student understands.

For a university, that is an uncomfortable problem. Some of its most familiar evidence of learning can now be produced by something that is not enrolled.

The temptation is to treat this primarily as a disciplinary matter: decide which uses count as cheating, require disclosure and punish violations. Colleges need clear rules. But even perfectly honest AI use can leave a student undereducated. Someone who acknowledges every machine-generated explanation may still have surrendered the practice that would have made those explanations intelligible. The harder question is what a college education should demand when completing the work and learning from it can so easily come apart.

That question is becoming part of the curriculum. Beginning with its fall 2026 incoming class, Purdue is requiring AI competency across undergraduate programs at its main campuses. Its approach emphasizes critical evaluation and knowing when AI is appropriate, with courses adapted to different fields. This is a sensible ambition. Graduates should understand the tools they will encounter at work, including their limitations. 1

Yet evaluating an AI answer requires something beyond familiarity with the interface. A business student reviewing a forecast needs to understand the assumptions behind it. An English student needs enough command of a text to recognize when an elegant interpretation has ignored an inconvenient passage. Telling students to exercise judgment does little unless their education gives them the knowledge and practice from which judgment grows.

AI could help provide that practice. Imagine a student who has understood most of a lecture but lost the thread at one crucial step. An explanation tailored to that misunderstanding could make the next hour of study useful. So could a patient exchange that requires the student to attempt a solution, explain a mistake and try again. There is a serious educational opportunity here, particularly when individual help is otherwise difficult to obtain.

A Harvard experiment offers evidence of that possibility. In an undergraduate physics course, researchers tested a custom AI tutor built around established teaching principles against classroom lessons using active learning. Students using the tutor made greater immediate learning gains in less time. Published in Scientific Reports in 2025, the study gives colleges a reason to invest in carefully designed instruction. It does not establish that an ordinary chatbot can replace a professor, or that gains on the tested material will persist months later. 2

The enthusiasm becomes harder to sustain when a tool moves into the untidy conditions of an ordinary semester. An October 2026 working paper from University of Maryland researchers describes a randomized trial involving 2,379 undergraduates. In comparisons between sections of the same course, access to a course-specific AI assistant produced lower final grades and less participation in online course activities. 3

The details matter. Only about 15 percent of students offered the assistant used it at least once, and both groups could access other AI tools. Most participating instructors left the assistant in a mode that supplied answers directly; only two ended the term using its guided tutoring mode. The paper remains preliminary, and its semester-long outcomes are different from the immediate tests in the Harvard experiment. These are findings about particular implementations, not a controlled comparison of two universities or a verdict on every AI tutor. 3

Still, colleges should be interested in what happens when a promising technology meets the incentives of daily academic life. Students have deadlines. Professors have crowded schedules. A system that makes answers readily available will be used by people trying to get through the week. Universities have to design for that reality.

The difference between completing a task and acquiring a skill has been tested more directly outside higher education. In a study of nearly 1,000 high school mathematics students in Turkey, published in PNAS in 2025, students using a basic GPT-4 assistant did better during practice but scored 17 percent lower than the control group when later tested without it. A version designed to give hints largely avoided that harm, although it did not produce a significant advantage on the unaided exam. The students were younger, but the experiment supplies a useful question for colleges: what can someone do after the assistance is withdrawn? 4

That question should sit beside another: what can the student accomplish with AI that would otherwise have been beyond reach? A serious education needs evidence of both. It should develop independent competence while teaching students to use powerful assistance responsibly. Requiring one without the other would leave graduates poorly prepared for a workplace in which they may have excellent tools and still be accountable for the result.

This calls for more deliberate assignments. A history student might use AI to develop competing interpretations, then defend a thesis against an unfamiliar source in class. A programming student could submit an AI-assisted project and be asked to explain its design, identify a failure and repair it. Such exercises would give instructors evidence of understanding that a polished submission alone cannot provide. They would also make the purpose of studying clearer to students.

Some work should still require students to begin unaided. An initial attempt exposes the point at which understanding fails, giving feedback something useful to address. Other work should invite AI use and demand a careful account of the choices that followed. The appropriate boundary will differ by discipline and stage of study. An introductory course teaching a foundational method has a different obligation from a senior project asking students to apply it.

None of this is an argument for making college needlessly difficult. An inaccessible explanation or a long wait for feedback can waste effort without teaching anything. AI may remove those obstacles. The challenge for faculty is to identify which demands of an assignment build understanding and which merely consume time. That requires attention to the subject and the student, as well as the software.

It also requires money and faculty time. Discussing a student’s reasoning, redesigning an assessment and evaluating a tutoring system all create work. A university that introduces AI while expecting teaching staff to absorb those tasks without support is making a consequential choice about educational quality. If the promise is better individual instruction, colleges should budget for the people who will determine whether students are receiving it.

They should measure the result with equal care. A subscription count reveals access. A satisfaction survey reveals whether students liked the experience. Neither establishes how much they learned. Colleges should examine whether students retain knowledge, apply it to unfamiliar problems and recognize a plausible error. They should follow those outcomes over time and look for differences among student groups, rather than allowing a favorable campus average to settle the question.

The labor market adds urgency without settling what colleges should teach. A September 2026 Dallas Fed analysis of Texas public university records found that graduates from majors more exposed to AI automation experienced relative declines in employment and earnings after ChatGPT’s release. Those findings concern a particular population and period; they do not establish the fate of any major. They do strengthen the case for preparing students to adapt as the work available to beginners changes. 5

Adaptation, however, requires enough understanding to recognize what has changed. A graduate who can produce an analysis but cannot interrogate its assumptions may find it difficult to improve on the machine’s first answer. Universities should be teaching students to notice what is missing, decide which evidence matters and take responsibility for a conclusion. Those habits deserve explicit practice in the curriculum.

Families evaluating colleges can ask to see what that practice looks like. An admissions presentation about AI access should lead to a conversation about an actual course: how assignments have changed, how understanding is assessed and what evidence supports the approach. Students should be able to ask the same questions. Their time is too valuable to spend four years accumulating work they cannot explain.

AI gives colleges a chance to make individual help more available and ambitious work more accessible. Earning those benefits will require universities to be as serious about evaluating the learning as they are about adopting the technology. When a student submits an impressive piece of work, the institution should be able to say what that student can now understand and do. That is the promise a degree ought to make.

Sources

1. Purdue University. Purdue’s AI competency graduation requirement begins in every undergraduate degree program. August 19, 2026. Main-campus scope and the incoming-cohort start date are also specified in Purdue’s December 12, 2025 announcement.
2. Greg Kestin and colleagues. AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, June 3, 2025. Peer-reviewed experiment measuring learning on specific physics material; it does not establish long-term retention.
3. Jing Liu and colleagues. The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial. EdWorkingPaper 26-1598, October 2026. Working paper. The trial included 2,379 students overall; comparisons between sections of the same course used a subsample of 1,353. Findings concern the assistant as deployed in this trial.
4. Hamsa Bastani and colleagues. Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, June 25, 2025. Peer-reviewed high school study, not a college trial. The August 2025 correction concerns an author affiliation, not the findings.
5. Samuel Dodini and Tucker Smith. AI plays a role in weak labor market for college graduates. Dallas Fed Economics, September 22, 2026. Analysis of Texas public university education records and employment and earnings observed in Texas.