A Practical Framework to Help Students Use AI in Learning

You have seen it happen: A student opens an AI tool, gets a polished essay outline in minutes, submits the assignment and walks away feeling productive. They do well on the exam. The grade is real. But ask them to explain the same concept three months later, and the room goes quiet.

This is not a new problem—shortcuts in learning have existed forever—but AI has amplified it significantly. The issue is not that AI competes with textbooks, but that it competes with friction. Textbooks are slow by design. They force you to sit with a concept for hours, trace its intellectual history, make your own connections, take notes and struggle toward understanding. That struggle is not a flaw in the learning process. It is the process.

Think of the brain as a muscle, one that you need to push through resistance to build strength that lasts. The same is true for learning. Neuroscientists call the mechanism synaptic plasticity: the brain’s ability to form and strengthen connections between neurons when it is actively engaged with new material. When you work through a difficult concept and arrive at understanding, your brain doesn’t just store information; it actually rewires itself. That rewiring is what makes knowledge durable, transferable and genuinely yours.

Research on learning science distinguishes between surface learning and deep learning. Surface learning relies on memorization and formulaic recall of information absorbed in isolation, with little to no personal investment or connection to broader understanding. Deep learning forges meaningful links between new and existing knowledge, and the difference in outcomes between these two levels of learning is stark. Passive, surface-level engagement is associated with forgetting 50 to 70 percent of new material within the first 24 hours. A randomized controlled trial at Georgetown University found that medical students who used ChatGPT during study sessions scored higher on immediate assessments but that this advantage disappeared in assessments one week later.

The danger of relying on chatbots for thinking has a name: cognitive outsourcing. It manifests when a student asks AI to brainstorm ideas before contributing any of their own, pastes a research paper into a chat window and asks for an interpretation without examining the figures, or accepts a literature summary without verifying it against the original source. In each case, the brain is bypassed, leaving the student with merely the illusion of mastery.

The issue boils down to how you use AI, not whether you use AI. Using AI tools during learning is very powerful. It can speed the process and enhance the quality of it by working like a tutor or a coach that provides immediate personalized feedback, spots learning gaps and presents concepts in multiple ways.

But users beware. In a research context, cognitive outsourcing carries additional risks. AI tools are prone to generating fabricated content that is presented as established fact, including invented citations to peer-reviewed articles that do not exist. Beyond academic integrity, there is a subtler cost: the loss of intellectual individuality. Original research depends on unique perspectives and novel framings. When AI replaces that process, it diminishes the creative core of what makes your work distinctively yours.

What Good Practice Actually Looks Like

In my (Jacob’s) undergraduate physics courses, I don’t ban AI tools, nor do I treat them as a shortcut to avoid. Instead, I try to structure AI into the learning process in a way that makes its strengths and limitations visible to students. What I’ve found is that AI doesn’t replace learning, but it does fundamentally reshape how students approach it, and whether that reshaping helps or hurts depends almost entirely on how we guide them.

One of the clearest places this shows up is in how students read scientific literature. In a research and writing course sequence that I teach, first-year undergraduates use AI tools early on to help identify papers and generate summaries around a research question. For most of these students, this is their first exposure to reading any scientific literature at all. When they see a clean, technical-sounding summary, they initially tend to accept it at face value because they don’t yet have a framework for questioning it.

This is where the structure of the course matters. After generating summaries, students are required to go back to the original papers and analyze them directly to compare what the AI summary said to what the papers present. Students start noticing what’s missing, including caveats in the experimental design, limitations in the data or entire threads of argument that never made it into the summary at all. In many cases, they discover ideas for their own research precisely from those “missing pieces.”

What this creates is a kind of transition moment. AI goes from being an apparent authority to a starting point. Students begin to see its outputs as useful but incomplete and, more importantly, as something they are responsible for checking. The goal isn’t just to help them find papers more efficiently; it’s to help them understand that reading a paper is not merely about extracting pertinent information, but about interpreting that information and data well enough to ask better research questions.

From that framework, a few concrete expectations fall into place. First, transparency is nonnegotiable. If students use AI, they must document exactly how they used it, including by submitting a transcript of their conversation(s). Second, students should avoid treating AI as a black box. For example, they should use an AI tool to generate a custom program to do data analysis rather than just prompting the AI tool to do the analysis. This way they can retain control over the tool, since the code is the tool rather than the AI itself. Third, students should never accept the first response. Whether it’s a literature summary or a block of code, it needs to be checked, tested and refined.

With this in mind, here’s my practical three-step framework to help students use AI:

1. Always start with your own attempt: Draft a rough answer, outline or plan. This practice gives you something to compare against and critique. AI feedback is most useful when you already have skin in the game and a point of view to defend.
2. Never accept the first response. Whether it is a literature summary, a block of code or an explanation of a new concept, check it against primary sources, test it and refine it. Require the AI to show its reasoning, then verify that reasoning yourself.
3. Document how you use it. Transparency is a professional habit, not just an academic policy. Keeping a record of your AI interactions builds awareness of where it genuinely helped and helps you create an iterative feedback process.

Building the right habits early is important, since what I’m really doing in the undergraduate curriculum is preparing students for the much higher-stakes environment of graduate research—where what’s at stake isn’t just coursework, but time, ownership and credibility over research. Ultimately, time is the most limited resource a graduate student has, and AI can help save time and speed up learning. But that increased efficiency must be grounded in understanding, verification and a clear sense of intellectual ownership.

That same philosophy carries over into how I frame AI more broadly in the curriculum. I describe the difference to students by naming them the lead researcher on their project and the AI tool(s) as a lab assistant, drawing the distinction for students that some uses of AI support the work they’re doing, while others replace it. If AI is helping them execute something that isn’t the central intellectual task (like generating Python code to analyze data they collected in an experiment they performed), then that’s likely an appropriate use. But if it’s doing the core thinking for them, then it undermines their learning.

I believe that there is reason for optimism, and it is driven by the fact that these habits can be taught early. When students learn to use AI as a tool they direct rather than an authority they defer to, they don’t lose the ability to think deeply; rather, they reinforce it.