Teaching and Learning in the Age of AI: What University Leaders Must Get Right Now*

The debate about whether universities should "allow" AI has ended—not because institutions resolved it, but because students moved on without them.

According to HEPI's 2025 Student Generative AI Survey, the proportion of UK undergraduates using AI for assessments jumped from 53% to 88% in a single year, while overall AI tool adoption rose from 66% to 92%. In the United States, a 2025 RAND survey of nationally representative samples found that 54% of students used AI for school—an increase of more than 15 percentage points in just one to two years—while over 80% reported that teachers had not explicitly taught them how to use it.

Students are not waiting for policy. They are using AI now, without guidance, without training, and often without any clear sense of where institutional boundaries lie.

That is a governance failure, not a technology problem. The operational question for provosts, deans, and teaching and learning leaders in 2026 is no longer whether but how: how do we redesign teaching, learning, and assessment so that AI improves learning quality and integrity rather than quietly eroding both?

This entry examines four levers that matter most: AI course copilots, instructional design workflows, academic integrity modernization, and assessment redesign.

I. AI Course Copilots: The Difference Between a Practice Layer and an Answer Engine

What a well-designed copilot actually does

A course copilot is not a generic chatbot dropped into a LMS. A serious implementation is course-bounded and pedagogically intentional. It supports learning through Socratic tutoring—prompting reasoning steps rather than delivering answers—targeted practice aligned to specific course outcomes, rubric-based feedback coaching, spaced-retrieval study planning, and differentiated support for language learners and students with accessibility needs.

The pedagogical rationale for this design is now empirically grounded. A 2025 randomized controlled trial from Harvard—published in Scientific Reports—found that students using a custom AI tutor learned significantly more in less time than students in active learning classrooms, and reported greater engagement and motivation. The study involved 194 undergraduate physics students in a crossover design where each student experienced both conditions, ensuring equivalence; the active learning control was itself a best-practice classroom, not passive lecturing.

That is a striking result, and it requires careful interpretation. The study's AI tutor was purpose-built using established pedagogical principles—not a general-purpose chatbot. The lesson is not "replace instructors with AI"; it is that a well-designed AI practice environment, grounded in sound instructional theory, can dramatically increase time-on-task and learning depth in ways that casual tool adoption cannot.

The counterevidence deserves equal attention. A systematic review published in 2025 found that heavy reliance on AI tutoring systems was associated in some studies with cognitive offloading—students passively accepting AI outputs rather than engaging with material—and with declines in critical thinking, information retention, and the capacity for independent analysis. The critical variable is design. AI that guides thinking produces different outcomes than AI that replaces it.

The "copilot as practice layer" model

The most reliable pattern emerging in practice is straightforward:

Core instruction stays human. Lectures, seminar discussion, laboratory work, studio critique, and clinical supervision remain faculty-led. These are the environments where disciplinary thinking, professional modeling, and human relationship develop.

The copilot adds a practice layer. Retrieval practice, problem coaching, iterative drafting feedback, and exam preparation occur between class sessions, mediated by AI that is constrained to course content and oriented toward guided reasoning.

This design achieves two things simultaneously: it increases learning time without multiplying instructor workload, and it channels student AI use toward learning behaviors—practice, reflection, revision—rather than submission behaviors—outsourcing finished work.

Guardrails that are non-negotiable

For copilots to function as learning tools rather than answer dispensaries, three institutional defaults matter:

First, the copilot must draw on and cite course materials rather than offering unmoored general responses. Second, it must surface reasoning steps as coaching prompts, not terminal outputs—the goal is always the next step the student should take, not the answer they are seeking. Third, aggregate usage patterns should be logged for course improvement purposes, with individual privacy protected. The data question—what engagement data is being collected, by whom, and how it is used—requires explicit governance before any deployment, not after.

II. Instructional Design: Raising Quality, Not Just Throughput

Most universities are seriously underinvested in instructional design capacity. AI can help close this gap—but only when it is used to improve design rigor, not to accelerate the production of templated content that looks polished and teaches poorly.

Where AI actually adds value for designers and faculty

The most productive applications cluster around four areas.

Outcome clarity. AI can draft learning outcomes using Bloom-aligned verbs, check alignment between outcomes, activities, and assessments, and surface "hidden curriculum" gaps—skills that are implicitly assumed but never explicitly taught.

Content architecture. Converting a syllabus into a clear learning pathway, building concept maps and prerequisite chains, generating candidate micro-lectures and common misconceptions lists for faculty review: these are time-consuming tasks where AI can produce strong drafts that faculty then validate and refine.

Active learning at scale. Generating discussion prompts across cognitive levels, drafting case scenarios and debrief questions, producing formative quizzes with rationales—this is labor-intensive work that AI can accelerate substantially when the underlying outcomes are clear.

Rubrics and exemplars. Drafting analytic rubrics aligned to stated outcomes, generating exemplar/near-miss pairs that teach students what quality looks like before they attempt a task: AI can produce serviceable first drafts that save significant faculty time and improve consistency.

The design principle that should govern all of this

If you standardize one thing institution-wide, make it this: AI produces draft options; faculty and instructional designers validate alignment, equity, and workload realism; students encounter AI only where the pedagogy explicitly justifies it. Everything else follows from that sequencing.

Skipping the validation step is how "AI-designed courses" happen—courses that are structurally coherent and pedagogically empty.

III. Academic Integrity: Stop Outsourcing Governance to Detectors

The detection trap

The instinct to solve AI use with AI detection is understandable. It is also increasingly untenable.

Research has shown that while AI detectors identified AI-generated text with 74% accuracy in controlled conditions, that accuracy dropped to 42% when students made minor modifications to the generated content. Vendor-claimed accuracy rates—typically 98–99%—reflect optimal laboratory conditions; real-world performance is substantially lower and highly variable. Detection tools produce false positives at meaningful rates and show particular bias against multilingual and non-native English speakers, raising serious equity and fairness concerns.

Turnitin reports approximately a 4% false positive rate at the sentence level—meaning a 500-word essay could have one or more sentences incorrectly flagged as AI-generated. At scale across thousands of students, that rate produces a substantial number of wrongful accusations with real disciplinary consequences.

In the 2023–24 academic year, 63% of teachers reported students for AI use on schoolwork—up from 48% the year before—while faculty report spending more time investigating detection flags than teaching. The MLA-CCCC Joint Task Force on Writing and AI has explicitly cautioned against detection tools, noting that false accusations may disproportionately affect marginalized students. Several major institutions, including UCLA and UC San Diego, disabled their AI detection tools in 2024–25 specifically due to false positive rates and escalating costs.

Detection is not a governance system. It is a pressure valve that creates adversarial dynamics, disproportionate harm, and the illusion of control.

What a modern integrity system looks like

Step 1: Make AI policies explicit and standardizable. The single most common source of academic integrity conflict in 2025 is ambiguity. In HEPI's survey, students described their institutions as "dancing around the subject"—AI use is neither banned nor permitted, described as academic misconduct in one context and modeled by faculty in another.

The administrative solution is straightforward: provide faculty with a small set of policy templates they can adopt and adapt—one for courses where AI is encouraged with disclosure requirements, one for courses where AI is permitted for specific steps only, and one for courses where AI is prohibited with clear justification. Clarity at the course level, sustained across the institution by shared language, is the foundation of any functioning integrity system.

Step 2: Treat detection outputs as signals, not verdicts. When detection tools are used at all, their outputs must trigger human review and direct conversation with students—not automatic escalation. Process evidence (drafts, notes, citations, revision history) should be required before any formal proceeding. Appeals pathways and burden-of-proof standards must be published before they are needed.

Step 3: Build integrity by design. The most durable integrity systems do not focus on policing. They focus on designing assessments that make outsourcing genuinely difficult and authentic engagement genuinely necessary—a point developed in detail in the next section.

IV. Assessment Redesign: The Validity Crisis Is the Real Problem

The urgency of this section is often misframed as a cheating problem. It is actually a validity problem.

When AI can generate plausible outputs for most standard assessment formats—take-home essays, generic problem sets, discussion posts, case analyses—those formats stop reliably measuring what faculty intend them to measure. The question is not just whether a student submitted their own work; it is whether the assessment, even when completed honestly, is actually evidence of the learning it claims to evaluate.

Assessment redesign is the most consequential lever available to teaching and learning leaders in 2026. It addresses integrity and validity simultaneously, and it is the only intervention that remains effective regardless of how AI capabilities evolve.

The "secure + authentic" portfolio model

The most credible emerging direction combines three elements: at least one secure assessment per course that validates core learning outcomes in a supervised or controlled environment; authentic assessments that mirror real-world performance demands; and portfolio evidence accumulated across time—drafts, reflections, revision responses, documented feedback cycles.

None of this requires abandoning existing course structures entirely. It requires adding accountability checkpoints that make the learning process itself visible and evaluable.

Practical redesign patterns that scale

Lightweight oral defenses. A 5–8 minute oral defense of a submitted artifact—a randomized prompt asking the student to explain a key decision or walk through their reasoning—is among the most efficient validity tools available. It requires no additional surveillance infrastructure, scales to large courses with modest TA investment, and is extraordinarily difficult to prepare for with AI alone.

Process-first writing. Anchor the grade not to the final product but to the documented process: topic proposal, annotated bibliography, outline with thesis rationale, Draft 1 with peer feedback response, Draft 2 with reflection on revisions. AI may assist at specified stages—brainstorming, critique generation—but the grade is earned through evidence of thinking, not polish of prose.

Local and unique data. Instead of "analyze the causes of X," ask students to analyze a dataset from a partner organization, a case from the institution's own context, field observations, interview transcripts, or lab results specific to the course. AI cannot engage with material it has not encountered, and students must.

Studio, clinical, and practicum demonstration. For professional programs, expand direct performance assessment: OSCE-style stations in health programs, live client simulations in business and counseling, real-time problem-solving in engineering. These have always been the gold standard of professional education; AI raises their relative priority.

Multi-modal outputs. Requiring a combination—short written brief, annotated diagram, two-minute explanation video—makes authorship easier to validate, differentiates the assessment from purely text-based AI capabilities, and supports Universal Design for Learning by giving students multiple modes of expression.

Designing for AI-permitted assessments

For many courses, prohibition is neither realistic nor pedagogically appropriate. The workforce these students are entering uses AI daily. The right response is not to pretend otherwise but to assess responsible use: require disclosure of what AI was used, where, and why; require students to verify claims and cite sources independently; require a demonstration of independent reasoning through reflection, oral defense, or critique; and build rubrics that score the quality of judgment and verification, not the surface polish of the output.

This is a more demanding standard than "did they write it themselves." It is also a more honest preparation for professional practice.

The Institutional Operating Model: What Provosts Should Do Now

The four levers above require institutional infrastructure to function at scale. That infrastructure has three components.

A Teaching and Learning AI Playbook—not a one-time workshop, but a maintained, living document that includes course policy templates by assessment type, disclosure and citation guidance, privacy and data-handling rules, approved use cases, and discipline-specific examples of redesigned assessments. UNESCO's guidance for generative AI in education consistently emphasizes this kind of systematic capacity-building as the prerequisite for human-centered implementation.

Governance structured as risk management, not tool procurement. The NIST AI Risk Management Framework offers a voluntary but widely adopted structure that translates cleanly to higher education: identify the risks (privacy, bias, hallucination, inequity, validity), specify the controls that mitigate them (policy, training, assessment design, vendor review), assign ownership, and review periodically. Buying a platform is not a governance strategy.

Faculty development that is practical, not philosophical. The most effective format is the redesign studio: faculty bring a syllabus or a specific assessment challenge and leave with a revised assessment map and at least one new instrument ready to pilot. Discipline-specific cohorts—lab sciences operate differently than philosophy programs, which operate differently than nursing—outperform generic training consistently. TA training is non-negotiable: teaching assistants are the front line of both learning support and integrity response, and they are routinely the least-prepared people in the system.

The Bottom Line

Over 80% of students report that teachers did not explicitly teach them how to use AI for schoolwork—yet they are using it anyway, at rapidly increasing rates, in an environment of profound policy ambiguity. The institutions that navigate this well will not be those that ban AI or those that simply buy an AI platform. They will be those that do the harder work: redesigning what they ask students to demonstrate, building the faculty capacity to assess it, and governing the technology with the rigor the moment requires.

That is not primarily an AI challenge. It is a leadership challenge that happens to involve AI.

*Text developed with AI assistance.