AI in Higher Education Is Growing Faster Than Policies

The Year AI Research in Higher Education Tripled and Nobody Was Ready

In 2021, something broke loose. Publications on artificial intelligence in higher education nearly doubled. In 2022, they did it again. By the end of that year, researchers had published two to three times as many papers on AI in higher education as they had in any prior year (Crompton & Burke, 2023). The field had been simmering since 2016, but suddenly it was boiling over.

The timing is not coincidental. These are the years just before ChatGPT went public, when large language models were still mostly inside research labs and tech companies. Yet the academic literature was already swarming with studies about AI tools that could assess student writing, predict dropout rates, and act as teaching assistants. The infrastructure for an AI revolution in higher education was being built in plain sight, and almost nobody outside the research community noticed.

Helen Crompton and Diane Burke, both at Old Dominion University, wanted to know exactly what that research landscape looked like. So they did what any good scientists do when a field explodes: they stepped back and mapped it. Their systematic review, published in the International Journal of Educational Technology in Higher Education and now cited over 1,200 times, examined 138 articles from 2016 to 2022. What they found is not just a summary of trends. It is a warning shot across the bow of every university administrator who thinks they have time to figure this out later.

The research is growing faster than the policies that govern it. Faster than the ethics frameworks that should constrain it. Faster than most faculty even know it is happening. And the gap between what researchers are building and what institutions are ready for is widening by the month.

Where Is the Research Actually Happening?

For years, the conventional wisdom was that AI in education was an American story. Silicon Valley would build the tools, American universities would test them, and the rest of the world would follow. That is no longer true.

Crompton and Burke found that research on AI in higher education has been conducted across six of the seven continents. But the center of gravity has shifted. China now leads the world in the number of publications, displacing the United States from the top spot (Crompton & Burke, 2023). This is not a small shift. It reflects a deliberate national strategy. China has been investing heavily in AI education research, and the output shows it.

The geographic spread matters for a practical reason. AI systems are not culturally neutral. A predictive model trained on Chinese university students may not work the same way for a diverse student body in Brazil or Nigeria or Germany. If the research literature is becoming increasingly dominated by one country's educational context, the tools that emerge from that research will carry the assumptions of that context. Universities elsewhere need to pay attention to where their AI tools are coming from, not just what they claim to do.

The People Behind the Research Changed Too

Here is a finding that surprised even the researchers themselves. Earlier reviews had noted a troubling pattern: most AI in education research was being done by computer scientists and engineers, not by people who actually study how humans learn. The technology was being built without deep educational expertise.

That has changed. Crompton and Burke found that the most common affiliation for researchers in this field is now departments of education (Crompton & Burke, 2023). This is a meaningful shift. It means that pedagogical questions are starting to drive the research, not just technical ones. The people asking "Can we build this?" are increasingly the same people asking "Should we use this, and how?"

But this shift also creates a new tension. Education researchers are typically not trained in the technical details of machine learning. They may not fully understand the limitations of the models they are studying. The field is becoming more educationally grounded, but it may be losing some technical rigor in the process. The ideal researcher in this space probably does not exist yet: someone who understands both the architecture of a transformer model and the cognitive science of how students actually learn.

Who Are These Tools Actually For?

The answer, overwhelmingly, is students. Crompton and Burke found that 72 percent of the studies focused on AI systems intended for students themselves (Crompton & Burke, 2023). Another 17 percent focused on tools for instructors, and only 11 percent targeted managers or administrators.

This distribution tells you something about the implicit theory of change in the field. The assumption seems to be that AI will improve higher education primarily by changing how students interact with content. Give students an AI tutor, an AI writing assistant, an AI assessment tool, and learning will improve.

But that assumption is worth questioning. The most powerful uses of AI in higher education may not be student facing at all. They might be behind the scenes: systems that help instructors design better courses, that flag students who are about to drop out before they disappear, that help administrators allocate resources more effectively. The fact that only 11 percent of the research focuses on managers suggests that the field is not yet thinking seriously about systemic change. It is focused on the classroom, one student at a time.

Undergraduate students were the most studied group, making up 72 percent of the subjects (Crompton & Burke, 2023). Graduate students, professional students, and adult learners were largely ignored. This is a gap that matters. The needs of a first year undergraduate taking introductory biology are very different from those of a doctoral candidate writing a dissertation or a working professional taking an online course. The field is building tools for the most visible student population, but it may be neglecting the ones who could benefit most.

What Are These Tools Actually Doing?

Crompton and Burke used grounded coding to identify five distinct categories of AI use in higher education. Each one represents a different theory of how AI should intervene in the learning process.

Assessment and Evaluation

The largest category was assessment and evaluation. These are systems that grade student work, provide feedback, or evaluate performance. Language learning was the most common subject domain, including writing, reading, and vocabulary acquisition (Crompton & Burke, 2023). This makes intuitive sense. Language tasks are structured enough that AI can handle them reasonably well, but open ended enough that the feedback is genuinely useful.

But there is a risk here that the field has not fully confronted. If AI assessment tools are built primarily for language learning, they may encode assumptions about what "good" writing looks like that are culturally specific. A system trained on standard academic English may penalize students who write in dialects or who come from different rhetorical traditions. The research is growing, but questions of algorithmic bias in educational assessment remain largely unanswered.

Predicting

The second category was predicting. These are systems that try to forecast student outcomes: who will pass, who will drop out, who needs intervention. Predictive models are seductive because they promise to let universities intervene before problems become crises. But they also raise uncomfortable questions. If a model predicts that a student is likely to fail, what do you do with that information? Do you offer them extra tutoring, or do you steer them away from challenging courses? The research shows that these systems are being built, but it does not yet show that they are being used wisely.

AI Assistant

The third category was AI assistant. These are chatbots, virtual teaching assistants, and other systems that interact directly with students to answer questions or provide guidance. The appeal is obvious: a student can get help at 2 AM when no human instructor is available. But the quality of that help varies enormously depending on the underlying model. A simple rule based chatbot is very different from a large language model, and the research literature does not always distinguish carefully between them.

Intelligent Tutoring Systems

The fourth category was intelligent tutoring systems, or ITS. These are the most ambitious AI tools in education: systems that adapt to individual students, provide personalized instruction, and adjust difficulty based on performance. They have been studied for decades, long before the current AI boom. But recent advances in machine learning have made them more powerful and more practical. The research shows that ITS are being deployed in higher education, but mostly in well defined domains like mathematics and programming where the content can be structured clearly.

Managing Student Learning

The fifth category was managing student learning. These are systems that track student progress, recommend resources, or organize course materials. They are the least glamorous of the five categories, but potentially the most impactful. A system that helps students manage their own learning across multiple courses could be more valuable than any single AI tutor. Yet this category received the least attention in the research literature.

What the Research Does Not Tell Us

This review is comprehensive, but it has limits that are worth understanding. The authors only included articles published in English. Given that China is now the leading producer of research in this field, a significant portion of the work may be happening in Chinese language journals that this review did not capture. The geographic shift toward China may be even more dramatic than the numbers suggest.

The review also stopped in 2022. That was the year before ChatGPT launched publicly and changed the conversation entirely. The landscape in 2024 looks very different from the landscape in 2022. The explosion of large language models has made AI tools more accessible, more powerful, and more controversial. The trends that Crompton and Burke identified have probably accelerated, but the nature of the research may have shifted in ways that this review cannot capture.

There is also a deeper question that the review does not address. How much of this research is actually being used in practice? A publication is not a deployment. The fact that 138 articles were published does not mean that 138 AI systems are being used in real classrooms. The gap between research and practice in education is notoriously wide, and there is no reason to think AI is any different.

What This Actually Means

• ▸If you work in higher education, assume that AI tools are already being tested on students at your institution, possibly without your knowledge. The research is happening faster than the oversight. Ask your department what AI pilots are running and what data they are collecting.

• ▸The shift from computer science to education departments is good news, but it creates a new problem. Education researchers need technical literacy to evaluate the tools they study. If your institution offers cross training between education and computer science, take it. If it does not, push for it.

• ▸Language learning dominates the research, but that is partly because it is easier. The hard problems in AI education are in domains like critical thinking, creative writing, and complex problem solving. Those are the areas where the research is thinnest and the need is greatest.

• ▸Predictive models are being built faster than ethical frameworks for using them. Before deploying any predictive system, ask: What happens to the students the model flags? What happens to the students it misses? What happens if the model is wrong? If your institution cannot answer these questions, it is not ready for predictive AI.

• ▸The concentration of research in China and the United States means that most AI education tools are being designed for students in those countries. If you work in a different context, be skeptical of tools that claim to be universally applicable. Test them on your students before trusting them.