From Agent Labels to Accountable Deployment: Evidence Maturity and Meaningful Human Oversight for Agentic AI in Higher Education

Higher education is standing at a quiet but significant turning point. For years, artificial intelligence in classrooms has functioned much like a very fast, very knowledgeable librarian. A student asks a question, and the system provides an answer. If the answer is wrong, the student simply asks again or checks another source. The risk is limited to a moment of confusion. But a new generation of software is beginning to emerge that does more than just talk. These systems, often called "agentic" artificial intelligence, can do things. They can break a large goal into smaller steps, choose which tools to use, gather information from different places, and then execute a plan without constant human direction. Imagine a system that doesn't just write an essay for a student, but one that can draft the essay, check the sources, submit it to a grading portal, and then email the professor with a summary of the work done. This shift from answering questions to taking actions changes the nature of the risk. When a system only speaks, a mistake is just a wrong sentence. When a system acts, a mistake can alter a student's record, expose private data, or trigger a chain of events that is difficult to undo.

The question facing universities today is not whether these powerful tools work, but whether we know enough to let them run on their own. A new critical review by researcher Abhishek Dey at South Ural State University examines the current state of this technology in higher education. The study does not look at the technology in isolation; instead, it asks a difficult ethical question: do we have enough proof that these systems are safe, fair, and beneficial before we hand them the keys to important decisions? The review sifts through dozens of existing studies and reports to find out what the evidence actually supports and where the gaps remain. The findings suggest that while these systems can improve specific, small tasks, the evidence is not yet strong enough to justify letting them make major decisions about students' futures.

The review begins by clarifying what counts as an "agent." Not every chatbot or automated tool fits this description. To be considered a true agent, a system must be able to plan a sequence of actions, remember context from earlier in a conversation, use external tools, and check its own work. Many studies in the field mix up these advanced systems with simpler ones that just generate text. This confusion has led to what the author calls "label inflation," where any automated tool is described as an agent, and "outcome inflation," where small improvements in how a system looks or sounds are mistaken for real improvements in how much students learn. The review finds that the literature is crowded with these broad claims, but the evidence is often shaky. The strongest proof exists for how well these systems produce specific artifacts, like a piece of feedback on a student's writing. However, the evidence becomes much weaker when looking at whether students actually learn more, whether the system treats everyone fairly, or whether it reduces the workload for teachers in the long run.

One of the most striking findings is that the most promising studies are often short, single-site experiments that do not tell the whole story. For example, one study tested a system that gave feedback to middle school science students. The system, using two different agents to check each other's work, significantly reduced the number of times it gave over-the-top praise or made up facts. This is a clear technical success. Yet, the study did not show whether the students actually learned more science because of this feedback, or if the teachers saved time after having to verify the agent's work. Another study looked at a system used in a university admissions office to answer questions from applicants. The system answered thousands of questions quickly and accurately, with very few errors. But this was a two-week trial. It did not show how the system would handle a full semester of admissions, how it would affect the fairness of who gets accepted, or how it would change the workload for the staff who have to manage the process. The review concludes that while these systems are technically impressive, we are still in the early stages of understanding their real-world impact.

The review also highlights a critical gap in how we manage these systems. Many reports talk about "human oversight," suggesting that a teacher or administrator is watching the system. But the review argues that having a human present is not the same as having effective control. True oversight means the human can see what the system is doing, stop it if it goes wrong, and reverse its actions if a mistake is made. In many current setups, the human is just a rubber stamp, approving outputs without the ability to understand the complex steps the system took to get there. This is dangerous because if a system makes a bad decision, the person in charge may not have the tools to fix it. The review points out that without strong, tested oversight, we cannot ethically let these systems make decisions that affect people's lives, such as grading exams or deciding who gets into a program.

To solve this mismatch between what the systems can do and what we know about them, the author proposes a new framework called Autonomy-Oversight-Evidence. This framework suggests that the amount of freedom we give an artificial intelligence should never exceed the strength of the proof that it is safe, nor the strength of the human control we have over it. The framework breaks down autonomy into levels, from a system that simply writes a sentence in response to a prompt, up to a system that can make consequential decisions about a student's future. It also breaks down evidence into levels, from a simple computer test to a long-term, real-world trial across many schools. The rule is simple: you cannot give a system high-level freedom unless you have high-level proof and high-level human control. If the evidence is weak, or if the human oversight is just a formality, the system's permissions must be kept narrow and reversible.

The implications of this finding are significant for how universities move forward. The review does not argue that these systems should be banned. Instead, it argues for a cautious, measured approach. Universities should not assume that a system that works well in a lab will work well in a real classroom. They need to test these tools in their own specific contexts, looking carefully at who benefits and who might be harmed. They need to ensure that teachers and staff have the power to stop the system and that students have a way to challenge decisions made by the machine. The review emphasizes that the burden of proof lies with the institution. Before handing over any real authority to an artificial agent, the school must show that they have the evidence to support it and the people in place to manage it.

Ultimately, the paper serves as a call for clarity and responsibility. The technology is moving fast, evolving from tools that answer questions to systems that take action. But the evidence of their safety and fairness has not kept pace. The review finds that while these agents can improve certain tasks, we are not yet ready to trust them with the big decisions that shape a student's education. The path forward is not to stop innovation, but to slow down and build the necessary safeguards. By matching the power of the system with the strength of the evidence and the quality of human control, universities can ensure that these powerful tools serve students without compromising the values of fairness, accountability, and learning. The goal is not to prevent the future, but to make sure that when it arrives, it is built on a foundation of truth and trust.