Trust Boundary Design for AI-Mediated Learning Support Services in Higher Education: A Conceptual Framework for Conditional Reliance
This paper proposes a conceptual framework of "trust boundary design" for AI-mediated learning support in higher education, which structures conditional reliance through three recurrent functions—entry, adoption, and repair—to ensure that AI use remains educationally accountable, revisable, and dependent on human agency rather than mere acceptance.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In universities today, artificial intelligence has become a quiet partner in learning. It can draft feedback on an essay, predict which students might struggle, or offer personalized study tips. For decades, researchers have asked a simple question: do people trust these machines? The answer, however, has proven complicated. Trust is a feeling, a general sense that a system is reliable. But in education, a feeling is not enough. A student might trust a tool and use it blindly, or distrust it and miss out on helpful guidance. The real challenge is not whether to use the technology, but knowing exactly when it is safe to rely on a specific piece of advice it gives. This distinction separates a general attitude from a practical decision: when is a specific output from an AI actually useful, and when should a human step in to check it?
A new conceptual paper by Chize Lee at Sanming University tackles this precise problem. The author argues that we need a new way to think about how humans and AI interact in the classroom. Instead of asking if students or teachers trust AI, the paper proposes a method called "trust boundary design." This is a set of rules and checks that determine when AI information is allowed to enter a learning task, how it should be checked before being used, and what happens when it leads to a mistake. The paper does not test a new software tool or run a survey of students. Instead, it builds a logical framework by bringing together existing research on feedback, learning analytics, and human oversight. It suggests that relying on AI should be a conditional process, not a simple yes-or-no choice.
The framework divides the interaction into three recurring moments. The first moment is entry. Before an AI tool is even allowed to touch a student's work, a decision must be made about whether it is appropriate for that specific situation. A tool might be fine for brainstorming ideas in a low-stakes exercise but completely unsuitable for grading a final exam or giving sensitive personal feedback. This stage is about setting the rules of the road. It asks: what is the goal of the task, how much risk is involved, and does the student have the skills to handle the AI's output? If the stakes are high or the student is vulnerable, the rules become stricter, perhaps requiring a teacher to approve the AI's use before it can even start.
The second moment is adoption, which happens when a specific piece of AI output is on the table. Just because an AI was allowed to participate does not mean every word it writes is safe to use. This stage requires "guided verification." Imagine a student receiving a paragraph of feedback from an AI. They cannot just accept it because it looks professional. They must check it against the assignment requirements, compare it to what they know, and look for errors or gaps. The paper suggests that the intensity of this check should match the importance of the task. A minor suggestion might need a quick glance, while a major prediction about a student's future performance requires deep scrutiny and perhaps a second opinion from a human expert. The goal here is to ensure that the decision to use the information is based on evidence, not just on how fluent the AI sounds.
The third moment is repair, which occurs when things go wrong. Sometimes, an AI gives advice that leads a student down the wrong path, or a prediction turns out to be misleading. The paper argues that fixing the immediate error is not enough. If a student follows bad advice and fails a task, the response must go beyond just correcting the grade. The system must also ask why the bad advice was allowed to influence the student in the first place. Was the entry rule too loose? Was the verification step skipped? The "repair" process involves fixing the immediate problem and then changing the rules for the future so the same mistake does not happen again. This turns a failure into a lesson about how to use the tool better next time.
Throughout these three moments, the paper emphasizes that responsibility is shared. It is not just the student's job to be smart enough to spot errors, nor is it solely the teacher's job to watch everything. The framework relies on three pillars working together: the student's ability to question and engage, the teacher's professional judgment to set standards and intervene, and the institution's support to provide clear policies and resources. If any of these pillars is missing, the system breaks down. For example, if a student lacks the skills to verify the AI's work, or if the school has no policy for what to do when AI fails, the "trust boundary" cannot function.
The author is careful to state that this framework is a proposal, not a proven fact. It is a map for thinking, not a finished destination. The paper does not claim that following these steps will automatically make students learn better or that it will eliminate all errors. Instead, it offers a way to make the process of using AI in education more transparent and accountable. It shifts the conversation from "do we trust AI?" to "under what specific conditions is it safe to rely on this specific piece of information?" By treating reliance as a series of conditional decisions rather than a fixed habit, the paper suggests a path forward where AI can be a powerful tool without taking over the human judgment that education requires. The work remains a conceptual guide, waiting for real-world studies to test whether these boundaries actually help students and teachers navigate the complex landscape of artificial intelligence.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.