Generative artificial intelligence (AI that creates text on request) will explain a complex concept, solve a problem or correct a sentence in seconds. This opens up great opportunities for study. However, easy and familiar strategies do not always produce solid knowledge.
When using AI, the main actor is the learner, so it is worth thinking through a strategy in advance. Completing a task successfully and learning from it are different things. AI can help get the right answer today, but not improve the ability to find it independently tomorrow.
An example is language learning. An app like Duolingo provides structure, repetition and motivation. The experience can be deepened by first trying to recall the answer and then looking at the options, repeating unfamiliar phrases aloud and returning to difficult vocabulary. Deliberate recall improves long-term retention compared with simple rereading.
A large language model (for example, ChatGPT) plays a different role: it explains, compares sentence structures and gives examples. A small study among Duolingo users showed that a reflective journal - notes on thoughts and goals - helped learners formulate strategies and difficulties.
Compare two people with the same model. One gets stuck and asks for the answer. The other first tries on their own and then asks why their answer is wrong, requests a hint or another example. The tool is the same, but the learning activity is different.
Psychologists use the concept of metacognition - the ability to analyze and regulate one's own thought processes. Effective learners assess what they understand, what needs practice and when help is needed. Research shows that these abilities also affect interaction with AI: in a semester-long study of students writing in English with ChatGPT, metacognitive awareness was associated with different patterns of use - from copying answers to actively working with feedback.
Technology design also matters. In a field experiment with almost 1,000 schoolchildren in mathematics, one group was given the standard version of GPT-4, and another a version that behaved like a tutor rather than giving out answers. The first group's performance improved, but after access was switched off, their results were 17% worse than those of those who had not used AI. The group with the tutor version largely avoided the decline.
The lesson is not to deprive people of help. Effective support removes unnecessary barriers while preserving mental activity. AI can give an answer immediately, but it can also offer a hint, ask the learner to try first, identify a misconception or explain the reasoning. What matters is which learning actions AI allows learners to stop and which it helps them perform better.
AI literacy is important: one needs to understand what systems can and cannot do and how to verify conclusions. But it is also worth thinking about learning literacy - an accessible description of the knowledge and judgments needed to regulate study in a technology-rich environment. AI literacy asks whether we understand the tool. Learning literacy asks whether we understand learning well enough to decide how to use the tool.
A recent meta-analysis (a statistical pooling of studies) of 32 works with 3,029 participants reported positive effects of AI-supported interventions on self-regulated learning. However, the effects depended on the design of the technology and the learning experience, so responsibility cannot rest only with learners. Schools can explain why recall, reflection and feedback work. Developers can create AI that encourages attempts before help. And learners can ask: what am I actually learning by using AI this way?