The biggest risk of AI in education is not cheating, but the erosion of learning itself.

A photo of a university lecturer giving a lecture in an empty classroom is becoming a metaphor. There is a person standing on the podium, but no one in the seats — not because of passive isolation, but because students may be interacting with chatbots in front of their screens. The wave of artificial intelligence is sweeping through higher education, yet the biggest risk is not the cheating issue we have repeatedly discussed, but the slow erosion of the very ecosystem of learning itself.

This is the core viewpoint put forward by Neil Ethington, Professor of Philosophy at the University of Massachusetts Boston and Director of the Center for Applied Ethics, and Jacob Burley, a researcher at the center. In a newly released white paper, the two scholars systematically sort out the increasingly widespread application of artificial intelligence in higher education, and warn that as machines become better and better at "pursuing academic work", the students, teachers and researchers that universities depend on for survival, as well as the inheritance relationship between them, are facing the danger of being hollowed out.

Cheating Is Just the Tip of the Iceberg

The vast majority of the public debate on artificial intelligence and universities focuses on one question: Will students use chatbots to write papers for them? Can teachers spot it? Should it be completely banned?

In Ethington and Burley's view, these concerns, while understandable, are seriously deviating from more core issues. Artificial intelligence has long penetrated into all aspects of institutional life in universities — from allocating resources, labeling "high-risk" students, optimizing course schedules, to automating daily administrative decisions. Students use it to summarize literature and assist their learning; teachers use it to assign homework and compile syllabi; researchers use it to write code and scan papers, compressing hours of tedious work into minutes.

The problem is not whether a few individuals cheat, but that when machines become more and more competent in scientific research and learning itself, what irreplaceable meaning do universities still have?

Three Types of Artificial Intelligence, Three Levels of Risk

To clarify this issue, the two authors divide the current applications of artificial intelligence in higher education into three levels.

Level 1: Non-autonomous Systems

Admission review, procurement, academic advising, risk assessment — artificial intelligence software has long been deployed in these fields. They are called "non-autonomous systems" because although they can perform tasks automatically, they still require humans to operate them as tools.

The problems brought by this level are relatively traditional: student privacy, data security, algorithmic bias, and lack of transparency. Who has access to student data? How is the "risk score" generated? Will the system exacerbate inequality and treat some students as "problems" that need to be managed?

Important as these issues are, universities do have governance mechanisms such as compliance offices and ethics review committees. The really tricky part lies ahead.

Level 2: Hybrid Systems

Hybrid systems include AI tutoring chatbots, personalized feedback tools, automated writing support and more. Most of them rely on generative artificial intelligence, especially large language models. Humans set the overall goals, but the intermediate steps taken by the system to achieve the goals are often not clearly specified.

It is right here that Ethington and Burley find the proper place for the "cheating" discussion. But they immediately point out that the problems caused by hybrid systems go far beyond that.

The first problem is transparency. The natural language interface of chatbots makes it difficult for people to tell whether they are talking to a real person or a machine. When students review for exams, they should know whether they are asking a teaching assistant or a robot; when they receive feedback on their final papers, they need to know whether the comments are from the teacher. If all this becomes blurred, research from the University of Pittsburgh has shown that students will experience uncertainty, anxiety and distrust.

The second issue is accountability and authorship. If teachers use artificial intelligence to generate assignments, and students also use artificial intelligence to complete assignments, then who will evaluate them? What will be evaluated? If the feedback is partially generated by the machine, who is responsible when it misleads students, demotivates them, or contains implicit bias? When artificial intelligence makes a substantial contribution to the synthesis or writing of research results, universities urgently need to formulate clearer authorship norms — not only for students, but also for teachers.

The third issue is the most critical: cognitive offloading. Artificial intelligence does reduce tedious work, which is not a bad thing in itself. But it may also lead users to skip the links that truly develop their abilities — generating ideas, overcoming confusion, revising clumsy drafts, and discovering their own mistakes. And these are precisely the core of learning.

Level 3: Autonomous Agents

The most disruptive change may come from systems that are more like "agents" than "assistants". Although truly autonomous artificial intelligence is still a vision, "a researcher built into a machine" — an intelligent system capable of carrying out research independently — is becoming more and more realistic.

In the field of scientific research, some robotic laboratories can already run continuously, automate most experimental steps, and even independently select new experiments based on previous results. In the field of teaching, it is expected that these tools can "liberate time", allow teachers to focus on more humanistic work, and hand over daily teaching to more efficient systems.

At first glance, this seems to be a huge leap in productivity. But Ethington and Burley put forward a profound warning: Universities are not information factories, but practice systems.

Universities rely on a steady stream of graduate students and young scholars — they learn by participating in research and teaching. If autonomous systems take over more of the "routine" responsibilities that used to be the entry path for academic careers, universities may still be able to offer courses and publish publications, but they are quietly disintegrating the opportunity structures that have long sustained professional skills.

The same logic applies to undergraduates. When artificial intelligence can provide explanations, drafts, solutions and learning plans on demand, it is easy for people to leave the most challenging parts of learning to machines. For the industry that promotes artificial intelligence into universities, this kind of work seems "inefficient", and students had better let machines do it.

However, cognitive psychology has long shown that students' intelligence grows exactly in the process of drafting, revising, failing, trying again, dealing with confusion, and revising weak arguments. This is the process of learning how to learn.

University: A Certificate Factory, or a Practice Ecosystem?

Combining the above analysis, Ethington and Burley conclude that the biggest risk brought by the automation of higher education is not that machines replace some specific tasks, but the erosion of the entire practice ecosystem that has long supported teaching, research and learning.

This leads to a disturbing question: In today's era when knowledge work is increasingly automated, what is the meaning of the existence of universities?

One possible answer is that universities are primarily engines for producing certificates and knowledge. The core issue is output — do students graduate? Are papers published? If autonomous systems can produce these results more efficiently, universities have every reason to adopt them.

But another view holds that the value of universities lies partially in the ecosystem itself. This model provides access channels for novices to grow into experts, a mentorship structure that cultivates judgment and sense of responsibility, and an educational design that encourages active exploration rather than blindly pursuing efficiency. What matters is not only whether knowledge and degrees are produced, but also how they are produced, and what kind of talents, capabilities and communities are shaped in the process.

The two authors clearly state that they prefer the latter understanding. In their view, the role of universities is no less than to become a reliable ecosystem that can cultivate human expertise and judgment.

A Question That Must Be Answered

In an era when knowledge work itself is increasingly automated, Ethington and Burley urge every university to ask itself: What on earth is the responsibility of higher education to students, to young scholars, and to the society it serves?

The answers to these questions will not only determine the way artificial intelligence is applied in universities, but also determine the future direction of modern universities. That photo of a lecturer in an empty classroom may not just be a nostalgia for the past, but a warning signal about the future.

This article is from the WeChat official account "Edu Guide", author: Neil, published with authorization from 36Kr.