Here’s how to find the learning strategy that makes AI work for you

Generative AI can explain a difficult concept, solve a problem, correct a sentence or produce an answer in seconds. That creates enormous opportunities for learning.

People have more access than ever before to information about different learning strategies. But the strategies that feel easy or familiar are not necessarily those that produce the most durable learning.

You are in charge when you use an AI so it’s a good idea to come up with a learning strategy that works. Understanding this matters because successfully completing a task is not the same as learning from it. AI might help a person produce the right answer today without increasing their ability to independently produce a correct answer tomorrow.

Language learning provides a useful example. An app such as Duolingo can provide structure, repetition and motivation. A learner can deepen that experience by trying to produce the answer from memory before looking at or selecting the options provided by the app, repeating unfamiliar sentences aloud or revisiting difficult vocabulary.

Deliberately retrieving information can improve longer term retention compared with simply studying material again. A large language model such as ChatGPT can play a different role. It can explain things, compare sentence structures, generate further examples or help investigate something the learner has noticed.

When an app is used together with a large language model, they can create a richer learning ecosystem. A small case study involving Duolingo users, for example, found that adding reflective journalling – writing down thoughts and goals – gave learners opportunities to articulate their strategies, difficulties and developing awareness of their own learning.

Consider two people using the same AI model. One gets stuck and asks for the answer. The other attempts the problem first and then asks why their answer was wrong. They might request a clue, ask for another example, or explain what they think is happening and ask the AI to challenge them.

Both have access to the same tool. But the learning activity is very different.

Changing how we use AI

Psychologists use concepts such as metacognition – the ability to examine, understand, and regulate your own thought processes – to describe how people monitor and manage their own learning. Effective learners make judgements about what they understand, what needs more practice, which strategies are working and when they need help.

Emerging research suggests these capacities also shape how people engage with AI. In one semester-long study of university students using ChatGPT for English-language writing, metacognitive awareness was associated with different patterns of AI use. Interview data showed behaviour ranging from simply copying ChatGPT’s output to using its feedback more actively to improve written work.

The design of the technology matters too. In a large field experiment involving nearly 1,000 secondary school maths students, the researchers gave one group of students access to a standard version of GPT-4. They gave another group of students access to a different version of GPT-4 that was designed to behave like a tutor, rather than simply supply answers to prompts.

The students with access to the standard version of GPT-4 saw their performance improve. Yet, when the researchers took their access away, these same students performed 17% worse than students who had no access to AI. The group who studied with the GPT-4 version designed to tutor largely avoided this decline.

The tutor group is important because the lesson is not that learners should be denied assistance or simply made to struggle. Effective support can remove unnecessary barriers while preserving the mental activity learning requires.

AI can give an answer immediately. But it can also offer a hint, ask someone to attempt the problem first, identify a misconception, generate another example or ask the learner to explain their reasoning.

The important question is therefore not simply whether AI is being used. It is which learning activities it is allowing someone to stop doing, and which it is helping them do better.

Self-aware learning

AI literacy matters. Learners need to understand what these systems can and cannot do, how to question their outputs and when their use is appropriate. But we might also need to think more explicitly about learning literacy.

Learning literacy is not a new psychological construct, but a public-facing way of describing the knowledge and judgment people need to regulate their learning in increasingly technology-rich environments.

AI literacy asks whether we understand the tool. Learning literacy asks whether we understand learning well enough to decide how that tool should be used. There is good reason to think AI could form part of the solution.

A recent meta-analysis of 32 studies involving 3,029 participants reported substantial positive effects of AI-supported interventions on self-regulated and self-directed learning.

However, the effects varied according to how the technology and learning experience were designed. That means responsibility cannot sit with learners alone.

Schools and universities can make learning itself more explicit, helping students understand why strategies such as retrieval, reflection and feedback work. Developers can design AI that encourages attempts before assistance, offers hints rather than solutions where appropriate, and asks learners to explain or revisit their reasoning.

And learners can ask a deceptively simple question: what am I actually learning by using AI this way?

We are spending enormous effort making AI better at answering questions. The educational challenge is to become equally good at recognising when those answers are helping us learn, and when they are simply helping us finish.