Why AI in Education Is Four Things, Not One

A faculty member uses artificial intelligence (AI) to draft an email to departmental colleagues about a schedule change. A student uses an AI tutor to work through a calculus problem. A computer science instructor teaches students how to use AI to generate code that can pass an enterprise security review.

Three uses of AI. One line item in the strategic plan. All three are simply AI, and absent a shared vocabulary, institutions tend to respond to them in the same way.

Many institutional AI policies treat these different uses of AI as though they were interchangeable, grouping them all under "AI in education" with one policy, one task force, and one set of guidelines. But a single rule cannot effectively govern all three. The email is an administrative task that requires no pedagogical competency. The tutoring session is a pedagogical interaction that could build or erode a student's mathematical reasoning, depending on its design. The computer science instructor is teaching a professional skill that may influence whether her student gets hired after graduation.

Three uses of AI. Three different purposes. Three different decision-makers. Yet at many institutions, all three uses appear in the same policy document, which may explain why a single policy can seem simultaneously too restrictive in one context and insufficiently specific in another. In a June 2026 EDUCAUSE Review article, Tanya Gamby and colleagues make a related argument: before writing an AI policy, institutions need a working theory of what thinking students should do themselves and what competence looks like in each discipline.^Footnote1

The Association of American Medical Colleges (AAMC) offers a useful starting point for distinguishing the different uses of AI across higher education. In 2025, it published "Principles for the Responsible Use of Artificial Intelligence in and for Medical Education," and the prepositions in that title are doing important work.^Footnote2 AAMC does not define the two prepositions in its principles framework, but the title points to a useful distinction: AI in medicine and AI for medicine are structurally different categories that call for different governance, expertise, and urgency.

That distinction works well for medicine, but colleges and universities, especially those serving dozens of disciplines, need a broader vocabulary.

A cross-disciplinary framework for higher education distinguishes four layers. "AI for education" supports the operation and administration of education, including educators' routine tasks. "AI in education" is intentionally built into teaching and learning by instructors or learners to develop understanding and reasoning. "AI of the profession" prepares students to use and evaluate AI in their fields, and "AI literacy" means understanding how AI works and where its limitations lie. This framework can help higher education leaders identify which forms of AI their policies address and where gaps may remain.

The four layers represent different institutional responsibilities, and success in one does not establish success in the others. A college or university may use AI extensively in its administrative operations, provide faculty with sophisticated AI teaching tools, and offer general AI literacy training, yet still fail to prepare students for how AI is transforming their disciplines and professions. The four-layer framework makes those differences visible. It allows institutional leaders to distinguish operational efficiency from educational effectiveness, foundational understanding from professional competence, and technological adoption from graduate readiness.

What Medicine's Prepositions Suggest

The AAMC distinction is elegant. AI in medicine refers to clinical applications: a physician using AI to read a CT scan, a predictive model flagging sepsis risk, or an ambient scribe documenting a patient encounter. This is AI as a professional competency. Future physicians must learn to interpret algorithmic outputs, evaluate them for bias, and make final decisions. That competency is increasingly nonoptional.

AI for medicine refers to institutional machinery: AI screening 2,000 residency applications, generating practice test questions, or summarizing student evaluations for recommendation letters. This is AI as infrastructure. It makes the system run faster and more efficiently. Removing AI from these processes would increase the administrative workload, but the medical school would still function.

AAMC-convened work reinforces the distinction. The International Advisory Committee for Artificial Intelligence (IACAI), which the AAMC convenes with international partners, published a 2026 AI Integration Framework: Educator Matrix with twelve domains. Among them, the matrix distinguishes foundational AI skills, instructional applications, clinical training, and workforce readiness.^Footnote3 While these categories are more granular, they point to the same conclusion: using AI to support education is not the same responsibility as preparing learners to practice in an AI-enabled profession.

The AAMC's implicit logic is worth stating plainly: AI for medicine is a turbocharger. AI in medicine is a new literacy. One is about efficiency. The other is about competence. Delaying the turbocharger carries administrative costs and is a scheduling decision. Delaying literacy carries curricular costs: graduates enter fields where AI competency is increasingly expected.

This logic works well for medical education. The reason it works is structural: in medical school, the classroom is the clinic. Students see patients during their training. The AI students learn about in class is the same AI they will use to treat a patient on Tuesday. There is relatively little gap between classroom AI and professional AI.

In many other academic disciplines, the connection between classroom AI and professional AI is less direct. Students may learn about and use AI in the classroom in ways that differ substantially from how they will encounter it in the workforce. That gap is precisely why the framework needs to expand.

The four-layer framework proposed here offers a simpler, cross-disciplinary vocabulary for distinguishing institutional operations, teaching and learning, foundational literacy, and professional practice across higher education.

The Instructor Thought Experiment

Consider the following thought experiment. Imagine an effective full-time instructor who left industry well before generative AI (GenAI) reached the workplace. This instructor has deep teaching experience, genuinely cares about students, and has decided to engage with AI.

The instructor learns how large language models work and understands probability engines, bias, and hallucination. That's AI literacy. The instructor uses AI to draft lesson plans, generate quiz banks, and format learning management system (LMS) pages. That's AI for education, the operational layer. The instructor also redesigns assignments so students must use AI, verify its output, and develop editorial judgment when working with AI-generated text. That's AI in education, the pedagogical layer.

This instructor has done everything right, and an entire category is still missing.

Teaching loads and the work of adapting courses to AI leave little time to monitor how AI is changing the workplace. Few institutional leaders provide faculty members with channels or incentives to stay connected to industry. As a result, the instructor may not know how AI is automating the supply chain decisions business students will encounter in their first jobs, how contract review software is transforming the legal field paralegal students are entering, or how clinical decision-support systems are used in the hospitals where nursing students will work. The instructor's industry experience predates these tools, but students primarily need to know about the last three years.

Now consider an adjunct instructor who manages supply chain operations during the day and teaches one evening course. This instructor sees firsthand how AI is changing the field but has no seat on the curriculum committee and little influence on what the program teaches. The full-time instructor has pedagogical depth but dated field currency. The adjunct instructor has current field knowledge but no voice in the curriculum. Together, they expose the gap between teaching experience and current field knowledge.

This is not a failure of effort or intelligence. It is a structural gap. Medical schools may experience less of a gap because many clinical faculty members remain active practitioners. But a business instructor may teach about AI in supply chain management without having worked in that setting. That is a fundamentally different challenge than using AI to teach.

Both instructors are example personas. In practice, the institutional ecosystem is messier. Acknowledging that complexity matters, partly because oversimplifying it can put higher education faculty members on the defensive, and partly because the structures behind that gap are where the actual work needs to happen. Adjunct faculty members may have more current field knowledge but little influence on the curriculum. Advisory boards can recommend changes but may lack the authority to implement them. The curriculum process itself is slow by design.

Student surveys provide further evidence of why this work is needed. In the Digital Education Council's 2026 survey of more than 27,000 students in thirty-five countries, half of student respondents said they felt underprepared to use AI professionally. Among respondents in the United States and Canada, that proportion reached 66 percent. Across all respondents, only 29 percent of students agreed that their instructors were well equipped to guide them on AI use. In the United States and Canada, however, only 17 percent agreed.^Footnote4 These findings measure student perception rather than instruction directly, and the instructor figure says as much about institutional support as about faculty. However, these perceptions reflect the downstream effect one would expect from the structural gap described above.

Four Layers, Not Two

The proposed framework distinguishes four layers of AI integration. While each answers a different question, serves a different audience, and has a different primary lead, all four should be coordinated under one institutional strategy.

AI for education answers this question: Is AI improving institutional operations and educators' routine work? Examples include back-office processes such as room scheduling and invoice processing, as well as LMS automation and degree audits. Routine educator tasks, such as formatting course pages, also belong here. This is the infrastructure layer. These applications address institutional needs, although the underlying technologies are not specific to education. From an instructional perspective, AI is primarily an efficiency tool rather than a requirement: these functions can be performed without AI, and the cost of going without it is easy to understand. This is also the layer IT teams are best equipped to evaluate because vendors across industries test these tools and their results are easy to benchmark. IT and administrative leadership lead this layer, with faculty input on tools that touch students.

AI in education answers this question: How can AI deepen learning rather than replace it? This is AI as a pedagogical instrument. In this layer, AI can serve as a coach, enhance assignments, incorporate progressive overload to increase challenge rather than remove it, and use verification protocols that teach students to critically evaluate AI output.^Footnote5 The student's interaction with AI is the learning activity. Faculty and curriculum committees lead this layer.

AI of the profession answers this question: Do students know about the AI they'll encounter in their careers? This is often job-specific AI that may be required in the workplace or embedded in industry software tools. Research on AI literacy distinguishes domain-specific AI competency from general AI literacy.^Footnote6 The nursing student must manage AI-generated clinical alerts, including the alert fatigue those alerts can cause, and work alongside ambient AI scribes. The accounting student must verify AI audit trails. The computer information systems student must document AI-agent orchestration frameworks. The art student must understand not only that people who use AI tools may compete for commissions but also how AI can support creative work. This layer applies to faculty as well as students. Instructors teach disciplines that AI is changing, and their own disciplinary currency, including the AI they use in professional practice and scholarship, belongs here. This layer has no natural lead. Program faculty, advisory boards, career services, and accreditation liaisons each hold part of it, which may be why it is the layer most often left out of institutional AI strategies. Developments in this layer may outpace accreditation processes, licensing requirements, and curriculum cycles.

AI literacy answers this question: Do people understand how AI works? This foundational layer includes probability engines, bias, ethical considerations, and hallucinations. Its core concepts apply across disciplines. A nursing student and a graphic design student need the same conceptual foundation about how training data shapes AI output, even though each student will apply it differently. General education committees lead this layer, which is the equivalent of quantitative reasoning for the AI age.

The four layers are not sequential stages of AI maturity, nor are they mutually exclusive. They distinguish responsibilities and desired outcomes, not technologies, and AI literacy provides a foundation that informs the other three. The same activity can occupy several layers, and the same person can work in all four in a single week.

Consider a faculty member, for example. Using AI to summarize department meeting notes is AI for education. Developing an AI-supported classroom exercise is AI in education. Using AI to analyze genomic sequences in the faculty member's own research is AI of the profession. Teaching graduate students to evaluate AI-generated scientific findings belongs to two layers at once: in and of the profession. Beneath all of it, understanding model bias and hallucinations is AI literacy.

The four-layer framework focuses on educational and disciplinary responsibilities rather than cataloging every institutional function. Research and scholarly practice fall within AI of the profession when AI changes the methods and judgments of a discipline. Teaching research methods may span AI in education and AI of the profession. Research administration, such as grants management or compliance review, works much like AI for education: it is operational, and administrators lead it.

The Literacy That Can Test Itself

AI literacy has a structural property that no previous form of literacy has possessed: the technology students are learning to evaluate can also take part in assessing their understanding. Digital literacy, information literacy, and quantitative reasoning are general education requirements that are embedded in a single gateway course and then never assessed again. The student who passes the information literacy module in a first-year composition course may never apply that skill to evaluating sources in their senior capstone. That gap may go undetected unless those skills are assessed again. The gateway-course-and-forget model fails because the assessment loop breaks after the initial course.

The AI literacy layer offers a way to avoid that mistake: AI can be both the subject of learning and a tool for assessment. A library database cannot assess whether a student knows how to use it critically. A calculator cannot test whether its user understands what the numbers mean. AI, however, can hold a conversation that probes whether a user understands how it works, and it can do this at scale, repeatedly, across a range of courses, and at any point in the curriculum. A five-minute AI-supported exchange could ask a student to identify an AI hallucination, explain why a probability engine generates different answers to the same question, and articulate when AI output needs human verification. Rather than a multiple-choice quiz, this would be a conversational exchange in which the AI adapts its follow-up questions based on the student's reasoning. Research on conversation-based assessment describes how such systems can generate adaptive questions, surface misconceptions, and give instructors formative evidence. Early applications are appearing in clinical education.^Footnote7 That evidence still requires appropriate validation and human oversight. AI literacy is the only layer that can be taught without AI. It is also the layer where AI most directly tests whether the teaching worked.

One question can help identify which layer an AI activity belongs to: Who evaluates whether it worked? AI literacy is evaluated by the institution or, in many cases, discovered through failure in the other layers. AI for education is evaluated by administrators watching efficiency metrics and peer benchmarks. AI in education is evaluated by instructors measuring learning outcomes. AI of the profession is evaluated by employers, clients, accreditors, licensing bodies, and practitioners in the field. Some of that evidence arrives before graduation, through clinical placements and internships. Much of it arrives after graduation, when institutional options for responding have narrowed considerably. Differences in feedback timing may help explain why institutional leaders prioritize AI for education, with its readily available performance metrics, over AI of the profession, where evidence of gaps in preparation may emerge later.

The Art Problem

The AI of the profession layer warrants a closer look because naming it almost derailed the entire framework. Early naming candidates included "AI workforce education" and "AI industry education." Both align with language used in workforce development, career and technical education funding, and advisory board structures. But both failed a critical test.

In the arts, AI may be less of a workforce tool than a source of competition. A visual arts student does not need to learn how to wield AI in a future job in the way a nursing student needs to learn to use clinical decision systems. A visual arts student needs to understand that they may be competing with people whose work is created wholly or partly using GenAI. The student also needs to understand that copyright law is in flux and that their aesthetic originality is now their primary market differentiator. "AI workforce education" assumes AI is an ally. For some disciplines, it is an adversary.

The phrasing "AI of the profession" won out because "profession" is neutral as to whether AI is a tool or a threat. It covers the nursing student learning to interpret a sepsis prediction algorithm and the art student learning that GenAI can quickly produce a commercial illustration. This distinction matters across polytechnic institutions, comprehensive universities, and community colleges, where a shared vocabulary must work across dozens of programs, not only those in which AI is obviously helpful.

The Framework in Practice

This framework has practical implications for policy, investment, teaching, and institutional planning across higher education.

Policy clarity. Institutional AI policies may mix everything together in a single document, covering chatbot procurement, assignment redesign, student AI competencies, and data privacy. The four-layer framework helps explain why these policies may appear incoherent: they govern four different responsibilities with one set of rules. AI in education requires curricular governance (learning outcomes, assessment redesign, accreditation alignment). AI for education requires operational governance (procurement, vendor management, data privacy). These layers involve different decision-makers, different timelines, and different risk profiles. Separating them clarifies who owns what.

The investment trap. Some colleges and universities invest heavily in AI-powered efficiency tools for back-office administrative processes such as room scheduling and invoice processing (AI for education), while neglecting curricular integration and professional preparation. The result is efficiency without readiness: an institution that streamlines operations without changing what graduates can do.

The opposite trap exists too. Some institutional leaders mandate AI-enhanced teaching (AI in education) without investing in the tools that could free faculty time to do it (AI for education). Academic leaders should not expect an adjunct faculty member teaching five sections to redesign assignments for AI collaboration without providing adequate time, tools, and support.

A four-layer audit. Leaders can review existing AI initiatives against four questions, each with its own evidence. Do we understand AI? The evidence is demonstrable AI competency among students, faculty, and staff. Are we using AI to work more effectively? The evidence is measurable operational benefit, with risks managed. Are we using AI to educate more effectively? The evidence is AI-supported teaching that improves defined learning outcomes. Are we preparing people for how AI is transforming their disciplines? The evidence is academic programs that identify and assess the AI competencies and impacts relevant to their fields.

Each initiative should identify an accountable leader, a desired outcome, and evidence of success. One initiative may contribute to several layers. High adoption in one layer is not evidence of maturity across all four.

The FLUFF/SPARK shortcut. Faculty do not need the four-layer vocabulary to start making better decisions today. A simpler framework already works for daily decision-making. Tasks that are formatting, layouts, under-the-hood, filing, and filtering (FLUFF) can be handed to AI because the work is not the learning. Work that is specific, persuasive, authentic, rigorous, and keen-insight (SPARK) is the learning, so AI should make the student think rather than think for the student.^Footnote8

FLUFF/SPARK addresses the nature of the work. The four layers address the institutional purpose of that work. The two operate at different levels, but they reinforce each other. A college or university can invest heavily in tools that eliminate FLUFF and make educators more productive without investing in the SPARK experiences that make students more capable. Conversely, educators who excel at SPARK teaching may work within academic programs whose curricula remain misaligned with changes in the professions.

A Shared Vocabulary for Conversations Happening in Silos

The AAMC's title gets something important right: prepositions matter. The difference between AI in medicine and AI for medicine is not merely semantic. The distinction has implications for who governs the technology, what competencies are required, and whether the investment is optional or existential.

The existential stakes are both institutional and personal. For institutions, the stakes concern the relevance and credibility of the education they provide. For faculty, the stakes reach professional identity. Professor and profession share a Latin root, profiteri, "to declare publicly." A professor is someone who publicly claims expertise in a field. When AI changes that field and complicates the evidence used to assess student learning, that claim becomes harder to make. The challenge is to sustain the expertise and educational judgment that make the role meaningful. The responsibility is not to know more than every machine, but to remain qualified to judge what matters, what students understand, and what they need to learn next.

Higher education needs to extend the AAMC's insight. Right now, the CIO is buying tools (for), the faculty senate is debating assignments (in), career services staff members are advising students about a job market being reshaped by AI (of the profession), and the general education committee is wondering whether every student needs to understand how large language models work (literacy). They are all talking about AI in education. They may not be talking to each other.

Four layers. Four questions. Four accountable leads. One institutional strategy. The vocabulary is simple. The institutional discipline to use it is hard: it means no longer conflating infrastructure decisions with curricular ones and no longer pretending that buying a chatbot is the same as teaching a student.

The distinction has precedent in business strategy. Michael Porter's value chain analysis separates primary activities (the work that creates the product) from support activities that keep the organization running.^Footnote9 In education, the "product" is the graduate. Abernathy and Utterback make a similar distinction between the technology embedded in processes and the technology embedded in products.^Footnote10 Applied to higher education, that distinction suggests that how instructors teach is not the same as what graduates know. The unique challenge for colleges and universities is that their product is a person who leaves and enters a different value chain. What graduates carry with them is what matters.

The four-layer framework addresses the vocabulary problem: making sure that when institutional stakeholders talk about AI in education, they know which of the four things they mean. A college or university can become highly sophisticated in using AI without producing graduates who are prepared for professions being transformed by AI. Institutional AI adoption is not the same as institutional AI readiness. When institutional leaders discuss the need for an AI strategy, someone should ask, "Which of the four layers is the strategy for?"

Notes

1. Tanya Gamby et al., "The Current State of Play: AI in Higher Education and the Road Ahead," EDUCAUSE Review, June 16, 2026. Jump back to footnote 1 in the text.↩
2. Association of American Medical Colleges, "Principles for the Responsible Use of Artificial Intelligence in and for Medical Education," version 2.0, July 2025. Jump back to footnote 2 in the text.↩
3. International Advisory Committee for Artificial Intelligence, IACAI AI Integration Framework: Educator Matrix (MedBiquitous, 2026). Jump back to footnote 3 in the text.↩
4. Digital Education Council, AI in Higher Education Global Survey 2026 (Digital Education Council, July 2026). Jump back to footnote 4 in the text.↩
5. Szymon Machajewski, The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI, version 2 (Zenodo, March 2026). Jump back to footnote 5 in the text.↩
6. Nils Knoth et al., "Developing a Holistic AI Literacy Assessment Matrix: Bridging Generic, Domain-Specific, and Ethical Competencies," Computers & Education Open 6 (2024): 100177. Jump back to footnote 6 in the text.↩
7. Seyma N. Yildirim-Erbasli et al., "Conversation-Based Assessments in Education: Design, Implementation, and Cognitive Walkthroughs for Usability Testing," Journal of Educational Technology Systems 52, no. 1 (2023): 27–51; Q. T. Nguyen et al., "AI-Assisted Formative Assessment in Clinical Education: From Algorithms to Agency," JMIR Medical Education 12 (2026): e93710. Jump back to footnote 7 in the text.↩
8. Machajewski, The Learn-It-All Educator. Jump back to footnote 8 in the text.↩
9. Michael E. Porter, Competitive Advantage: Creating and Sustaining Superior Performance (Free Press, 1985). Jump back to footnote 9 in the text.↩
10. William J. Abernathy and James M. Utterback, "Patterns of Industrial Innovation," Technology Review 80, no. 7 (June/July 1978): 40–47. Jump back to footnote 10 in the text.↩

---

Szymon Machajewski is Director of Academic Enterprise Applications and Operations at University of Illinois Chicago.