A lead instructor at a UK university taught the same two undergraduate economics and business modules across three academic years: one before ChatGPT, one with students using AI on their own, and one with AI built into tutorials and workshops. A 2026 study in Education and Information Technologies tracked grades, attendance, and independent study hours in all three. It is one of the clearest comparisons yet of AI in higher education.

Access to AI changed little on its own. Course design changed what students did with it. Four other recent studies, from Ireland, China, a review of academic libraries, and a conceptual paper on academic integrity, show where the decisive variables sit.

What the UK Modules Show About AI in Higher Education

In the smaller module, which contributed 31 student-level observations, the average overall grade went from 18 in the pre-AI year to 14 in the unguided year and 16 in the guided year. That recovery was not statistically significant. Independent study hours moved from 23 to 26 and then to 57, and that jump was significant.

The larger module, with 103 observations, showed the clearest gain. Overall grades held at 16 in the first two years and rose to 18 in the guided year, a statistically significant change, while study hours went from 19 and 20 to 44. These are three cohorts, not a controlled experiment.

The skill-level results varied. In the smaller module, the critical-thinking measure (the coursework essay) fell from 20 to 17 in the unguided year and recovered to 19 in the guided year, a marginally significant change (p = 0.056). The presentation grade in the larger module, the study’s proxy for communication skills, rose only from 17 to 18, which was not significant. The gains were larger in written and analytical work than in presentations.

The 27 students who answered the end-of-year survey were positive. Of them, 72.2 percent called AI helpful or extremely helpful in decision-tree design workshops, and 66.6 percent rated it effective or highly effective for problem-solving in tutorials. Self-reports from a small group show perception, not performance, but they match the direction of the grades. Three cohorts cannot settle the question for AI in higher education generally, which is why the other four studies matter.

What Students Do When AI Arrives Without Rules

Violeta Morari and colleagues surveyed 608 students at an Irish university during the first semester of 2023 to 2024 and found a student body waiting for guidance on AI in higher education. Only 43 percent reported familiarity with the institution’s integrity policies. Asked whether using generative AI for assignments is cheating, 43.9 percent agreed and 32 percent stayed neutral.

Students also reported hearing little about it. While 71.6 percent said lecturers often discussed plagiarism, only 23.1 percent said the same of chatbots such as ChatGPT. Lecturers taught the integrity problems they already knew how to name. Only 22.1 percent agreed that generative AI enhances learning, though many stayed neutral, and acceptance rose with academic stage.

Enforcement signals were mixed. A majority, 352 against 251, wanted a formal reporting tool, and 58 percent supported one. Yet only 28 percent said they would report a peer, while 258 said they would not and 176 were unsure. Students want enforcement to exist and refuse to be its instrument. A policy that leans on peer reporting asks for conduct the cohort says it will not supply.

The history of educational technology explains why unstructured adoption disappoints. In Failure to Disrupt, Justin Reich reviewed MOOCs, automated graders, and computerized tutors. He concluded that technologies built for learning at scale fell short of their promises, that their benefits went mostly to well-resourced students, and that successful reform depends on incremental institutional change. Generative AI adds a complication: the older tools graded and tutored, while this one can produce the assignment itself.

Critical Thinking Decides Who Gains From AI

A study of 744 university students across mainland China, published by Yingchun Wang in Frontiers in Psychology, tested what separates students who gain from AI from those who do not. Students rated their AI literacy on four dimensions: awareness, usage, evaluation, and ethics. Higher literacy went with higher academic self-efficacy, meaning confidence in one’s ability to succeed at coursework.

For AI in higher education, the mechanism matters. Self-efficacy carried 26 to 29 percent of the link between AI literacy and achievement, depending on the dimension. Critical thinking then changed the result. Among students one standard deviation above the mean on critical thinking, awareness of AI predicted self-efficacy with a coefficient of 0.50. Among students one standard deviation below, the coefficient was 0.09 and statistically indistinguishable from zero.

The other three dimensions split the same way: 0.47 to 0.51 for the high group and 0.11 to 0.15 for the low group. Literacy raised confidence only where students could judge what the machine produced. A tool amplifies the judgment of whoever holds it.

A tool amplifies the judgment of whoever holds it.

The data come from one point in time and from self-report, and the achievement measure is a six-item rating of how fully students complete coursework, so the causal order remains open.

Even so, the result suggests that teaching students to operate AI does not by itself build confidence, because the gain depended on a separate capacity.

Judgment of this kind is slow to build and does not arrive with the output. The experience-based understanding that scholars call tacit knowledge belongs to the person who has done the work, which is why a first-year student and a doctoral candidate can read the same model-generated paragraph and do entirely different things with it. Teaching students to evaluate AI output comes before teaching them to rely on it.

Why Assessment Design Beats Detection

Detection software cannot carry the weight universities place on it when they govern AI in higher education. Alireza Maleki’s conceptual paper in Discover Artificial Intelligence argues that plagiarism detection becomes insufficient because AI-generated content is often original by design. Governance must shift from asking whether the student produced a text to asking how the student produced it and with which tools.

Maleki organizes academic integrity into four interdependent layers: governance and policy, assessment design, institutional and cultural norms, and individual moral agency. The paper reports no data. Its value lies in the sequence it describes, in which policy shapes assessment, assessment shapes cultural norms, and norms shape the moral choices individual students make.

The UK modules illustrate the same point. In the smaller module, a regression controlling for attendance gave unguided AI use a significant negative coefficient on overall grades and independent study hours a significant positive one. In the larger module, the unguided effect was small and changed with the estimation method. Where the course gave students a structured reason to study, hours and grades rose.

The practical alternative to detection is to assess the process. A multi-stage assignment with a proposal, a draft, peer feedback, and a final version gives every stage a reason for the student to think. A short oral defense confirms that the author understands the argument. A written statement of which tools were used, for what, and why leaves a trail the instructor can read.

Neil Selwyn made the underlying point in Should Robots Replace Teachers? before ChatGPT existed. After examining classroom robots, intelligent tutoring systems, learning analytics, and automated decision-making, he argued that bringing AI into teaching is a choice that belongs in debates about values, judgment, and politics. Process-based assessment is that choice made concrete.

Capacity Gaps Decide Who Gets Guidance

Structure costs staff time, and staff time costs money. That makes funding a central question for AI in higher education. A systematic review by Lateef Ayinde, Raphael Ebiefung, and Bolaji David Oladokun in The Journal of Academic Librarianship examined 29 studies from 2020 to 2024 on AI in academic libraries. Funding was the most frequently cited obstacle, appearing in 14 of the 29 studies, or 48 percent.

The rest of the list reads like a map of unequal capacity. Eight studies reported limited IT skills among librarians, six reported inadequate infrastructure, three reported an unreliable power supply, and four reported fear of job loss. The authors note that institutions in developing countries struggle in particular to allocate dedicated funds for AI. The review covers English-language studies only, so scholarship in other languages is missing from the picture.

A policy written at a well-funded research university will not transfer to an institution that cannot pay for training. Access to guided AI use then depends on a budget decision. Institutions are adopting AI in higher education amid broader shifts in how AI is used and governed, and the capacity gap decides which students benefit.

A Workable Policy for AI in Higher Education

Taken together, the five studies support four commitments for institutions. They form a proposal drawn from the evidence, and no single paper tests the package as a whole. Each commitment answers a gap the studies expose: unclear rules, unobservable process, uneven judgment, and unequal capacity.

• Write a rule for each assignment. State which uses of AI are permitted, which are prohibited, and why, in the assignment brief itself rather than in a university-wide document.
• Assess the process. Use staged submissions, brief oral defenses, and short statements of AI use so that the route to the final text is visible.
• Teach evaluation before use. Wang proposes a model in which students generate content with AI, then submit a critical analysis report that examines factual accuracy, argumentative logic, potential bias, and source reliability, and finally rewrite the work while separating their own contribution from the machine’s. Making that report a graded first-year skill puts evaluation ahead of reliance.
• Fund the staff. Structured use requires someone to structure it, which means training time, library capacity, and instructor hours.

Two Objections to Structured AI Use

The strongest objection is that structure may be doing all the work. In the UK modules, the guided year also brought higher attendance and far more independent study, so the data cannot separate the effect of AI from the effect of any demanding redesign. That reading is fair, and it supports the argument: if a well-structured course without AI produced the same result, design would still be what drives outcomes.

A second objection comes from instructors who favor prohibition: any use of a model removes the effortful struggle that produces learning. The unguided year supports part of that worry: final exam averages in the smaller module fell from 22 to 14. A ban needs a way to enforce it, and detection fails when the text is original by design. Prohibition then binds only the students who choose to honor it.

Public resistance to AI now extends beyond technical debates about regulation, and universities will be asked what their students actually gain. The evidence needed to guide AI in higher education is already within reach: staged assignments, tracked study time, and comparison groups that hold workload fixed. Institutions that run those comparisons will know what AI adds to learning. The rest will keep debating a tool whose use they cannot see.