From Generative AI to Agentic AI in Higher Education: A Structured Literature Synthesis of Instructional, Human Agency, and Governance Tensions
This paper synthesizes fragmented literature on the transition from reactive generative AI to proactive agentic AI in higher education to identify key instructional, agency, and governance tensions, culminating in the development of the AIRA-G framework to guide institutional readiness and accountability.
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
The Big Picture: From a Smart Calculator to a Self-Driving Car
Imagine higher education is a busy city. For a long time, universities have used Generative AI (GenAI) like a very smart, fast calculator or a personal assistant. You ask it a question ("Write me a summary of this book"), and it gives you an answer. It waits for you to push the button.
But the paper argues we are now moving to Agentic AI. This isn't just a calculator anymore; it's like a self-driving car or a roving tour guide. It doesn't just wait for questions. It can:
- Plan a route on its own.
- Check traffic (data) and change the plan.
- Talk to other cars (systems) to coordinate.
- Make decisions about where to go next without you asking every single step.
The authors say universities are not ready for this shift. They have rules for how students use calculators, but they don't have rules for how to manage a self-driving car that is driving the whole class.
The Four Big Tensions (The "Traffic Jams")
The paper identifies four main problems (tensions) that happen when you switch from the "calculator" to the "self-driving car."
1. The Shift: From Tool to Teammate
- The Analogy: Imagine a hammer. You swing it to hit a nail. It does exactly what you tell it. Now imagine a robot construction worker that decides which nails to hit, when to take a break, and how to build the wall.
- The Problem: AI is moving from being a tool you hold to a "distributed actor" that works alongside teachers and students. The question becomes: How much of the driving should we let the robot do? If the AI decides the lesson plan or grades the test, who is actually in charge?
2. The Trap: Personalization vs. Real Learning
- The Analogy: Imagine a personal chef who cooks exactly what you crave. If you only eat what you crave, you might get sick because you aren't eating the vegetables you need.
- The Problem: Agentic AI is great at making learning "personalized" (giving you exactly what you want). But if it makes learning too easy or only gives you what you already know, you might stop struggling with hard concepts. The paper warns that if AI makes everything feel easy, students might stop doing the "mental lifting" required to actually learn. We need to make sure the AI isn't just a shortcut, but a real teacher.
3. The Balance: Helping vs. Crutches
- The Analogy: Think of training wheels on a bike. They help you balance until you are ready to ride alone. But if you keep the training wheels on forever, you never learn to balance yourself.
- The Problem: AI can help students learn faster (training wheels). But there is a risk of "cognitive offloading"—using the AI to do the thinking for you. If students rely on the AI to write, solve, or think, they might lose their own ability to do those things. The goal is "calibrated agency": using the AI as a helper, not a replacement for your own brain.
4. The Risk: One Broken Window vs. The Whole Building
- The Analogy: If a student uses a bad calculator, they get the wrong math answer (a broken window). But if a self-driving car in the school system gets hacked or makes a bad decision, it could crash the whole school bus, leak student secrets, or unfairly expel a student (the whole building shaking).
- The Problem: When AI is just a tool, the risk is small (a wrong answer). When AI is an "agent" that connects to student records, grades, and school databases, the risk explodes. It's no longer just about cheating; it's about security, privacy, and who is responsible when the system messes up.
The Solution: The AIRA-G Framework
The authors propose a new way for universities to decide if they should use these new AI systems. They call it the AIRA-G Framework.
Think of this as a Traffic Light System for AI. Before a university lets a new AI "self-driving car" into their city, they have to run it through four checkpoints (lenses):
- Autonomy Check: How much can this AI do on its own? Is it just answering questions, or is it making decisions?
- Learning Check: Does this actually help students learn, or is it just making things faster?
- Human Check: Does this make students smarter and more independent, or does it make them lazy and dependent?
- Safety Check: Is the university ready to handle the risks? Can they protect data? Can they fix it if it breaks?
Based on these checks, the AI gets sorted into one of four Zones:
- Zone A (Green Light - Low Risk): The AI is just a helper (like a FAQ bot). Action: Let it go, but tell students what it is.
- Zone B (Yellow Light - Guided): The AI helps teachers plan or give feedback, but a human checks it. Action: Let it go, but keep a teacher in the loop.
- Zone C (Orange Light - Controlled Pilot): The AI is making big suggestions (like advising students on classes). Action: Test it carefully in a small group with strict rules and logs.
- Zone D (Red Light - Stop): The AI is making high-stakes decisions (like grading final exams or deciding who graduates) without enough safety nets. Action: Do not use it yet. It's too dangerous.
The Bottom Line
The paper isn't saying "Ban AI" or "Use everything." It's saying: "Stop treating AI like a simple tool and start treating it like a partner with its own power."
Universities need to stop asking, "Is this tool cool?" and start asking, "Is our school ready to govern this partner?" They need to check if the AI is safe, if it actually helps learning, and if humans are still in the driver's seat. The AIRA-G Framework is their map to make those decisions safely.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.