Learning With AI vs. Learning From AI

Why how you use AI matters more than whether you use it?

There is a sentence I keep hearing from educators, and it bothers me every time: “My students are learning from AI.”

They usually mean it as a good thing. Students are getting explanations, feedback, answers. The AI is patient, available at midnight, never frustrated. What’s not to like?

Quite a lot, it turns out, if we are precise about what learning actually is.

Two students. Same assignment. Same tool.

The first opens ChatGPT and types: “Can you explain cognitive load theory and give me a classroom example?” She reads the response, closes the tab, and rewrites the explanation in her own words. When she gets stuck, she goes back, not for the answer, but for the part she did not understand.

The second opens ChatGPT and types: “Write me a 500-word explanation of cognitive load theory for my assignment.” He copies the response, changes a few words, submits.

Both used AI. Only one learned anything.

This is the distinction the education world keeps missing. The debate about AI in schools circles around policy, detection, and academic integrity. Rarely does anyone ask the more important question: how are students using these tools, and what is that use doing to their thinking?

Everyone is talking about AI. Nobody is asking the right question!

ChatGPT crossed one million users in five days after launch — faster than any consumer technology in history (OpenAI, 2022). Tools like Gemini, Claude, Copilot, and Perplexity now sit in the pockets of virtually every university student. Teachers are overwhelmed. Institutions are scrambling.

But most of the conversation circles the wrong question. Should AI be banned? How do we detect it? These are real concerns but they are downstream of something more fundamental.

The question that actually matters is not whether students use AI. Most do, and that is not changing. The question is what kind of relationship they are building with it. Are they using AI to support their thinking, or to replace it? That single distinction determines whether AI becomes one of the most powerful learning tools ever created, or one of the most effective ways to avoid learning ever invented.

What learning actually requires?

Before evaluating how AI fits into learning, we need to be precise about what learning is. This sounds obvious. It isn’t.

Learning is not the same as receiving information. A student can read a perfect explanation of photosynthesis, understand every word in the moment, and remember none of it a week later. That is not learning. That is exposure. Learning, in the cognitive science sense, is the process by which information becomes encoded in long-term memory in a way that can be retrieved and applied in new contexts [(Kirschner et al., 2006)](http://Kirschner, P. A., Sweller, J., & Clark, R. E. (2006). Why minimal guidance during instruction does not work. Educational Psychologist, 41(2), 75–86.).

This process has requirements. It requires the learner to actively process new information: connecting it to what they already know, reorganising it, retrieving it effortfully. Cognitive psychologists call this generative processing [(Wittrock, 1992)](http://Wittrock, M. C. (1992). Generative learning processes of the brain. Educational Psychologist, 27(4), 531–541.). Passive reception: reading, watching, being told, produces far weaker memory traces than active construction [(Roediger & Butler, 2011).](http://Roediger, H. L., & Butler, A. C. (2011). The critical role of retrieval practice in long-term retention. Trends in Cognitive Sciences, 15(1), 20–26.)

This is the foundational problem with learning from AI. When a student asks an AI to explain a concept and reads the response, the cognitive work is done by the AI. The student receives a finished product. And finished products, however clear and accurate, do not require the generative processing that learning depends on.

What learning with AI looks like?

Learning with AI means keeping yourself as the thinker. The AI is a tutor, a coach, a study partner but the cognitive work stays with you.

In practice this means asking AI to explain something you don’t understand, then restating it in your own words without looking. It means sharing a draft argument and asking what is weak in your reasoning, then deciding yourself how to respond. It means using AI to surface different perspectives on a question, then evaluating which ones hold up and why.

[Chi et al. (1989)](http://Chi, M. T. H., Bassok, M., Lewis, M. W., Reimann, P., & Glaser, R. (1989). Self-explanations: How students study and use examples in learning to solve problems. Cognitive Science, 13(2), 145–182.) found that students who habitually explain material to themselves in their own words develop significantly deeper understanding than those who don’t, even when both have access to the same content. The AI can prompt that process. It cannot do it for you.

This is what [Vygotsky (1978) ](http://Vygotsky, L. S. (1978). Mind in society: The development of higher psychological processes. Harvard University Press.)meant by scaffolding: temporary support that enables independent performance, not permanent replacement of it. AI as scaffold removes obstacles. AI as substitute removes the learning.

What learning from AI looks like and why it falls short?

Receiving AI-generated content is not automatically a problem. Reading an explanation, reviewing a summary of a paper you have already read, getting a worked example, these can all be part of a legitimate learning process.

The problem is passive consumption without doing anything cognitively demanding afterward.

Cognitive Load Theory draws a useful distinction here (Sweller, 1988). There is extraneous load: unnecessary mental effort caused by confusing design, and germane load: the effort directed toward actually building knowledge. Good instruction reduces extraneous load and preserves germane load. AI-as-content-source often reduces both. A clean, well-organised AI explanation eliminates not just the confusion but the productive cognitive work a student would have done working through that confusion themselves.

[Sweller et al. (2011)](http://Sweller, J., Ayres, P., & Kalyuga, S. (2011). Cognitive load theory. Springer.) are explicit: reducing the effort required to complete a task is not the same as reducing the effort required to learn. These are different goals. Conflating them produces work that feels easy but teaches little.

The real problem: letting AI think for you

When students use AI not to support their thinking but to replace it entirely i.e. copying generated essays, using AI to avoid reading, submitting answers to problems they never attempted, something more serious than academic integrity is lost.

[Bjork and Bjork (2011)](http://Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher et al. (Eds.), Psychology and the real world (pp. 56–64). Worth Publishers.) describe what gets bypassed: desirable difficulties. The struggles that feel uncomfortable: not knowing, getting something wrong, having to revise, are not obstacles to learning. They are the process by which learning happens. Remove them, and you remove the learning.

There is also a retrieval problem that rarely gets discussed. [Roediger and Karpicke (2006) ](http://Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255.)showed that students who studied material once and then tried to recall it performed significantly better on delayed tests than those who re-studied the same material. The act of retrieval (pulling information from memory under effort) strengthens encoding far more than re-exposure does. When a student uses AI to answer a question they were asked, they are not retrieving. They are outsourcing. The AI retrieves; the student reads. Whatever encoding existed is not reinforced. It is bypassed [(Agarwal & Bain, 2019).](http://Agarwal, P. K., & Bain, P. M. (2019). Powerful teaching: Unleash the science of learning. Jossey-Bass.)

What educational psychology tells us

My own MPhil research sits at the intersection of motivation and learning, specifically how teachers and students move from amotivation toward genuine engagement. And what the research keeps returning to is simple: learning requires the learner to be present. Cognitively and emotionally present. Curious, effortful, willing to not know yet.

Self-Determination Theory identifies three needs that must be met for deep engagement to occur: autonomy, competence, and relatedness (Ryan & Deci, 2000). When AI does the thinking, none of these are met. There is no genuine choice in prompting a machine. There is no sense of growth in reading output you did not produce. And there is no relatedness in a transaction between a student and an algorithm.

[Zimmerman (2002)](http://Zimmerman, B. J. (2002). Becoming a self-regulated learner: An overview. Theory Into Practice, 41(2), 64–70.) describes self-regulated learners as those who have developed both the skills and the motivational beliefs to direct their own learning. AI used as a content source undermines both, reducing opportunities to build skill through practice and eroding the belief that one’s own effort is meaningful.

Deep learning requires effort, reflection, curiosity, and practice. AI cannot do these things for students. It can support students who are already doing them. It cannot replace students who are not.

How students should actually use AI?

The research points to a clear principle: use AI to enhance your thinking, not to avoid it.

That means using AI to get an explanation of something you do not understand, then closing it and restating the idea in your own words. It means asking for feedback on a draft you have already written, not a draft you want AI to write. It means using AI to generate practice questions on material you are reviewing, then attempting them before checking. It means exploring different perspectives on a question you are already thinking about, then doing the evaluating yourself.

What it does not mean is using AI to produce the output you were supposed to produce. Not copying generated answers. Not letting AI read the difficult text so you do not have to. Not outsourcing the argument in your writing or the problem-solving in your assignments.

The dividing line is not complicated. Before using AI on any task, ask one question: am I about to think, or am I about to let AI think instead of me? If it is the latter, the task will be completed but the learning will not happen.

The difference between learning with AI and learning from AI is not a matter of degree. It is a matter of who is doing the cognitive work and cognitive work is what learning is made of.

AI is not inherently good or bad for education. It is a tool. And like every tool, what matters is not the tool itself but what it is being used to do.

The future of education is not about whether students use AI. That question is already settled. The future depends on whether AI becomes a genuine tool for thinking or a comfortable substitute for it. And that depends entirely on whether studentsand the educators who teach them,understand the difference.