How to use AI to study properly: a 5-step system that matches how memory works

AI can summarise your notes in seconds. It can explain a difficult chapter clearly. It can produce unlimited practice questions on demand. None of that is the same thing as learning the material.

Memory does not strengthen because something has been processed near you. It strengthens because you retrieved it, got it slightly wrong, corrected the error, connected the idea to others you already hold, and came back to it days later when it had started to fade. That sequence is over a hundred years old in the laboratory literature, and the recent meta-analytic work continues to support it. What is new is the temptation to skip every step of it because a tool can finish the surface task for you.

The most useful version of the question is not whether AI helps with study. The recent peer-reviewed evidence on this is mixed in interesting ways. Students who use AI tools often perform better on the immediate task and worse on later tests of what they actually know. The useful question is how to set the AI up so that the work that builds memory still happens in your own head. Five steps. The prompts are below each one, ready to copy.

Start with one dedicated chat per subject. Biology stays separate from psychology. Maths stays separate from history. The reason is structural. Inside a single chat, the AI carries forward what you have done together. It can track recurring mistakes, notice the kinds of question you tend to get wrong, and call back to material from earlier in the session. Mixing subjects in one chat blurs all of that.

At the top of each new subject chat, paste the setup prompt below. It does two things. First, it tells the AI to draw only from the material you provide, which sharply reduces the risk of confidently stated errors on subject content the model is shaky on. Second, it shifts the AI out of summary mode and into tutor mode. The distinction matters. The recent peer-reviewed evidence shows students who use AI in summary mode often perform better on the immediate task and worse on later tests of what they actually retained (Fan et al., 2024). Tutor mode is what produces the second outcome instead of the first.

Once the chat is set up, paste a single section of notes. One lecture. One chapter. Not the whole subject. Then ask the AI for questions, not for an explanation.

Answer those questions without scrolling back to the notes. This part is the actual learning. The phenomenon is called the testing effect. Repeated retrieval produces stronger long-term retention than equivalent time spent rereading the same material (Roediger & Karpicke, 2006).

The recent Australian-led systematic review of distributed and retrieval practice in health professions education found significant benefit across the majority of included studies, replicating the effect in present-day educational settings (Trumble et al., 2024). The size of the gain is not subtle, and it has held up across decades of replication.

If recall feels effortful, that effort is the point. It is the signal that learning is happening rather than just being witnessed. Do not move on until you have attempted every question, even the ones you are unsure about.

After you have answered, paste the correction prompt. The point is to identify what was wrong, understand why it was wrong, and then rewrite the weakest answer in your own words. The AI does not rewrite it for you. You do.

The reason for the rewrite is that correcting a recently retrieved memory is when the memory becomes most editable. The next time you retrieve it, you retrieve the corrected version. Over a few cycles of mistake, feedback, rewrite, the corrected version becomes the version that lives there. The cycle that builds durable understanding is mistake, then feedback, then rewrite in your own words, then test again.

Once you have corrected the mistakes, ask the AI to push you on how the ideas relate to each other. This is the step most students skip and where the largest gains in long-term retention sit.

Answer those relationship questions before you do anything else. The reason is that the brain does not store isolated facts well. It stores networks. Asking the AI to make you compare, contrast, and connect ideas builds multiple retrieval pathways into the same material, which is what makes it accessible later under the cognitive load of an exam or an applied task. Facts are fragile. Connections are durable.

Before you close the session, paste the spacing prompt. It produces a short revision pack you will not look at today.

Then schedule it. Not vaguely, in your calendar. Three days from today. When you return, attempt the questions before reviewing anything. The reason this matters is that memories consolidate over time, and the consolidation is faster and more durable when retrieval is spaced out rather than packed into one session (Kang, 2016).

Cramming creates familiarity. Spacing creates durability. Repeated, spaced retrieval is also one of the most consistent findings in the broader meta-analytic literature on study (Trumble et al., 2024).

The system also recommends mixing topics in those return sessions. Pulling material from two or three earlier topics into the same review block makes the brain work harder to find the right pattern, which is what locks the patterns in. The discomfort of that effort is what Bjork and Bjork (2011) called desirable difficulty: the kind of effort that feels like worse learning in the moment and produces better learning over time.

If you only take one thing from this, take this: speed is not the friend of memory. Effort that feels slightly difficult is the friend of memory. The five-step system is not slower for the sake of it. It is the shape that the cognitive science of learning has converged on across decades of research, and AI does not change that. AI just makes it easier than ever to skip the steps. Set up your chats so that you do not.