The Difference Between Using AI and Learning With AI

Quick answer

The difference between using AI and learning with AI is the difference between offloading a task and keeping a skill. When a student uses AI to get a task done and accepts the result, the work appears - but the capability does not transfer to them. When a student learns with AI, they stay cognitively engaged: they form their own view first, direct the tool, check what it claims, and make sure they could do the task themselves afterwards. Same tool, opposite outcome. The distinction is not philosophical; it is testable. Close the laptop and ask whether the student can still do the thing. If they can, they learned with it. If only the output survives and the ability does not, they merely used it. That single test - can you carry it yourself? - is what separates a tool that makes students sharper from one that quietly makes them weaker, and it is the same distinction the economy is starting to draw between people who command AI and people who are merely near it.

Why this matters now

The evidence that how students use AI matters more than whether they use it is now strong enough to act on. A 2025 study in Societies by Gerlich - a mixed-method study of 666 participants - found AI tool use was strongly correlated with cognitive offloading (r = +0.72), and that offloading was inversely related to critical thinking (r = -0.75). The effect was most pronounced in 17-25-year-olds, the cohort using these tools most heavily. The study's conclusion is the part schools should underline: AI is not inherently detrimental - its impact depends on use. The task is to stay cognitively engaged.

The other half of the evidence is just as important, and it points the opposite way when the conditions are right. A World Bank randomised controlled trial in Nigeria, reported in From Chalkboards to Chatbots (2025), found that six weeks of structured, teacher-supported AI-assisted learning produced gains equivalent to roughly one and a half to two years of typical schooling - outperforming about 80% of rigorously evaluated education interventions. The variable that made the difference was not the model; it was the structure and the human guidance wrapped around it. Put the two studies side by side and the picture resolves: unsupervised offloading erodes thinking, while deliberate, scaffolded engagement can accelerate it dramatically. The tool is the same in both. The design is not.

This is also why the distinction has a commercial edge that should concentrate the mind. McKinsey's State of AI 2025 finds that 88% of organisations now use AI, but only around 7% have fully scaled it - the gap between access and value is enormous, and it is a capability gap, not a tooling one. The World Economic Forum's Future of Jobs Report 2025 names analytical thinking the single most important core skill and expects 39% of workers' core skills to change by 2030. The students who learn with AI are building precisely the capability the labour market is short of; the ones who merely use it are practising the dependence employers are learning to screen out.

Students themselves sense the trade-off. RAND's American Youth Panel found that 67% of students said using AI for schoolwork harms critical thinking - up from 54% earlier in 2025 - even as homework use climbed from 48% to 62% across the year. They are not naïve cheerleaders; many are using a tool they quietly suspect is weakening them. That is precisely the gap education exists to close, and it is why our explainer on what AI education is starts from the same premise: the open question is whether AI becomes a crutch or a lever.

What "using AI" really means

Using AI is the path of least resistance, and that is exactly the problem. It is the natural default - ask the confident machine, take what it gives, move on.

In practice, using AI looks like this:

• The student meets a task, opens AI, and asks it to do the task.
• The output arrives, polished and plausible, and is accepted largely as-is.
• The thinking the task was meant to develop is performed by the tool, not the student.
• The work is submitted; the capability never formed.

This is what the research calls cognitive offloading - handing the mental work to an external system. Offloading is not always bad; a calculator offloads arithmetic so you can focus on the maths. The danger is offloading the very thing the task was designed to build. Gerlich's data gives this its sharpest form: the strong positive correlation between AI use and offloading (r = +0.72), and the strong negative correlation between offloading and critical thinking (r = -0.75), describe exactly the mechanism by which an essay written by the tool leaves the student no better at writing essays. A student who offloads the reasoning has not saved effort - they have skipped the lesson. The work exists; the learning does not.

What "learning with AI" really means

Learning with AI keeps the student in the cognitive driver's seat, using the tool to go further rather than to go elsewhere. The defining feature is engagement, not avoidance.

Harvard Project Zero's Teaching for Understanding offers the sharpest lens here: real understanding is a flexible performance, not recall. A student understands something when they can do it - explain it, apply it to a new situation, use it under changed conditions - not merely retrieve or reproduce it. Learning with AI is organised entirely around protecting that performance.

In practice, learning with AI looks like this:

• The student forms their own first view before opening the tool.
• They direct AI to support the thinking - to explain, stress-test, or map - not to replace it.
• They verify what comes back against what they know and against real sources.
• They produce the work themselves and, crucially, could reproduce the reasoning without the tool.

The student still does the hard thinking. AI removes the friction that used to make them give up - but the performance stays in the student. This is also where the strongest evidence sits. The Education Endowment Foundation, working with Evidence for Learning in Australia, finds that teaching metacognition and self-regulated learning - having students plan, monitor and evaluate their own thinking - is among the highest-impact, lowest-cost strategies in education, worth roughly seven months of additional progress when taught alongside subject content. Learning with AI is metacognition with a new tool in the room: it works because it forces the student to think about their own thinking at every step, which is the habit the World Bank's structured, teacher-supported trial was effectively scaling. That is the line, and it maps directly to what AI literacy means: the judgement to use a tool without being governed by it.

The Carry test: Command Not Comply

Edison AI Academy teaches students a single, blunt test that decides which side of the line they are on. We call the framework Command Not Comply, and it runs in four moves.

1. Comprehend - understand the task and form your own first view before opening AI.
2. Command - direct the tool deliberately: clear context, clear ask, clear constraints.
3. Cross-check - verify the output against what you know and against real sources.
4. Carry - make sure you could do the thinking yourself; the skill must stay in you.

The fourth move is the one that settles the whole question. Carry asks: if the AI vanished right now, could you still do this? If yes, you learned with it. If no, you complied - and complying is how a student graduates unable to think without a subscription. The discipline is deliberately structured rather than left to chance, which is the same lesson the evidence keeps returning: UNESCO's AI Competency Framework for Students (2024) sets out a staged progression precisely because the capability has to be sequenced, not absorbed by exposure; and the AI Assessment Scale (AIAS) - a five-level spectrum from "No AI" to "Full AI", set per task and made transparent to students - exists because the appropriate amount of AI is a deliberate design decision, not a default. Command Not Comply is the student-facing version of the same principle: decide, before you start, how much of the thinking has to stay yours. The Carry test is the moment of truth, because it exposes the difference that polished output hides. It is the same discipline that runs through our AI education for teens programs, from Foundations to Innovators.

Same tool, opposite outcome: three examples

The clearest way to see the difference is to watch the same task go both ways. Three examples show offloading versus learning, side by side.

• The essay. Using AI: the student asks AI to write the essay and lightly edits it. Learning with AI: the student asks AI for the strongest objection to their thesis, then writes the rebuttal themselves. What the student does today: engages with the argument instead of skipping it. How AI assists: it stress-tests the reasoning rather than producing it. What the student must verify: that the counter-argument is real, not invented. The learning outcome: sharper reasoning the student owns - analytical thinking, the very skill the WEF ranks first. The control: the words, and the thinking, stay theirs.
• The maths concept. Using AI: the student asks for the worked solution and copies it. Learning with AI: the student asks AI to explain the concept three different ways, then attempts a fresh problem unaided. What the student does today: uses AI to get unstuck, not to get the answer. How AI assists: it reframes the idea until one version clicks - the kind of structured, guided support the World Bank trial found so effective. What the student must verify: that they can now solve a new problem without it. The learning outcome: genuine understanding as a flexible performance. The control: the unaided attempt is non-negotiable.
• The research task. Using AI: the student pastes the AI summary straight into the work. Learning with AI: the student uses AI to map an unfamiliar topic quickly, then goes to primary sources to confirm every claim. What the student does today: builds a fast scaffold, then does the verifying. How AI assists: it sketches the terrain at speed. What the student must verify: every fact, against a real source - the cross-check that keeps Gerlich's offloading curve at bay. The learning outcome: a faster start with the judgement intact. The control: nothing enters the work unchecked.

In every pair, the tool is identical. The outcome is decided entirely by whether the student stayed engaged or stepped out.

How to shift from using AI to learning with AI

The shift is a change of habit, not of tool, and it can be taught directly.

1. Think first, then open the tool. Form your own view before AI ever sees the task. This single sequence prevents most offloading - it is the "comprehend" move that the rest depends on.
2. Ask AI to support, not to supply. Use it to explain, challenge or map - verbs that keep the thinking with the student - not to produce the finished answer.
3. Verify everything. Treat every output as a claim to be checked, not an answer to be trusted. This is the habit Gerlich's data shows the youngest users most often skip.
4. Set the level before you start. Decide, the way the AIAS asks educators to decide per task, how much AI is appropriate for this particular piece of work - and be honest about it.
5. Apply the Carry test before submitting. Ask: could I do this myself now? If not, the learning has not happened yet.
6. Make independence the measure. Judge the work by what the student can do unaided afterwards, not by how quickly it appeared - the same way the World Bank trial measured genuine learning gains, not output volume.

Common mistakes

• Judging by output, not capability. Polished work can hide a complete absence of learning - the classroom version of McKinsey's gap between adopting AI and actually capturing value from it.
• Banning AI to force learning. Prohibition drives use underground and ungoverned; it does not build judgement, and it forfeits the genuine gains structured use can deliver.
• Treating all AI use as offloading. Used to get unstuck and then verify, AI strengthens thinking - the habit decides, not the tool. Gerlich's own conclusion is that AI is not inherently detrimental.
• Skipping the "think first" step. Opening AI before forming a view is where offloading begins.
• Never applying the Carry test. Without it, students - and parents - cannot tell which side of the line a piece of work fell on.

How to know which one is happening

The tell is independence, and it is observable at home. A student learning with AI can close the laptop and reproduce the reasoning, explain it, and apply it to a fresh problem. A student merely using AI produces faster, thinner work and cannot do the task once the tool is gone. RAND's data suggests many students already feel this gap before adults name it - which is the opening, not the alarm. The whole subject sits inside the broader case for AI education for teenagers in Australia: the tool is here, and the only real question is whether students command it or comply with it.

The recommendation is direct. Do not measure a student's AI use by how much it produces - measure it by what they can still do without it. Teach the sequence: think first, direct the tool, verify the output, set the level deliberately, and apply the Carry test before anything is finished. The research is unusually clear for an emerging field: Gerlich shows that unsupervised offloading erodes thinking, the World Bank shows that structured, teacher-supported engagement can accelerate it by one to two years, and the EEF shows that the metacognitive habit underneath learning with AI is among the most cost-effective interventions in education. The tool is not good or bad for thinking; the design around it decides. Build the habit of learning with AI, and the tool becomes the best study partner a student ever had. Skip it, and the same tool becomes the most persuasive way yet invented to stop learning while appearing to work.