Meta-AI literacy and knowledge transfer: Serial mediation through metacognitive regulation and verification behavior in AI-assisted learning

In the modern classroom, a new kind of tutor has arrived, one that never sleeps and can draft an essay or solve a problem in seconds. This is generative artificial intelligence, a tool that has quickly become a routine part of university life. Students use it to summarize readings, draft assignments, and revise their work. Yet, a quiet worry has grown alongside this convenience: just because a student can produce a good paper with help, does that mean they have actually learned anything? The core question for educators and researchers is not whether students can access these powerful tools, but how they engage with them. When a complete answer appears on demand, the mental effort required to understand a concept can be skipped entirely. This is known as cognitive offloading, where the brain delegates the hard work to an external resource. The danger is that a student might hand in a perfect assignment while retaining no understanding of the material once the tool is turned off.

To understand how learning survives in this new environment, researchers look at a concept called transfer. Transfer is the ability to take what you have learned and apply it to a new situation. There are two main types. Near transfer happens when the new problem looks very similar to the one you practiced, sharing the same surface details. Far transfer is much harder; it requires applying the same underlying principle to a problem that looks completely different on the surface. Far transfer is the true test of deep understanding, because the learner must recognize the deep structure of a problem without the familiar clues that signal which knowledge is relevant. If a student simply copies an answer, they might pass a near transfer test that resembles their homework, but they will likely fail when faced with a novel challenge that requires flexible thinking.

A team of researchers set out to investigate how students who are skilled at using artificial intelligence actually learn. They focused on a specific set of skills called Meta-AI literacy. This is not just about knowing how to type a prompt; it is a broader competence that includes understanding how AI works, knowing how to use it, detecting when it is being used, and understanding the ethics surrounding it. The researchers wanted to know if having these skills automatically led to better learning, or if something else had to happen in between. They suspected that the key lay in the student's behavior during the task. Specifically, they looked at whether students actively checked the AI's output against other sources, a behavior known as verification. They also examined whether students planned their work and monitored their own understanding, a process called metacognitive regulation.

The study involved 483 university students from a wide range of majors, including humanities, engineering, and the sciences. The students were asked to complete a learning task using an AI tool. During this time, the researchers did not rely on what the students said they did; instead, they recorded every interaction the students had with the tool. They logged how often students went back to check the source of a claim, compared different sources, traced evidence, and corrected errors. This provided a clear, objective picture of verification behavior. After the learning phase, the AI tool was turned off, and the students took a test to see what they had learned. The test included questions that were similar to the learning task, representing near transfer, and questions that applied the same principles to entirely different, unfamiliar scenarios, representing far transfer.

The results revealed a surprising truth about how learning happens with AI. The researchers found that simply having high Meta-AI literacy did not directly lead to better test scores. A student could be very knowledgeable about AI and its ethics, yet still fail to learn the material if they did not use those skills in the right way. Instead, the path to learning was a chain of events. First, students with higher Meta-AI literacy were more likely to engage in metacognitive regulation. They planned their approach and monitored their understanding as they worked. Second, this careful planning led them to verify the AI's output. They checked the answers, looked for errors, and compared the information with other sources. Finally, it was this act of verification that directly predicted how well they performed on the test.

The most striking finding concerned the difference between near and far transfer. The act of verifying the AI's work was a strong predictor of success on the difficult, far transfer questions. In fact, the connection between checking the work and solving unfamiliar problems was nearly four times stronger than the connection between checking and solving familiar problems. This suggests that when students stop to verify an answer, they are forced to re-engage with the underlying reasoning rather than just accepting a surface-level solution. This effortful processing is what allows them to understand the deep structure of a problem, which is essential for applying knowledge to new situations. Without this step, the learning remains shallow and fragile.

The study also showed that metacognitive regulation did not directly lead to far transfer on its own. It only helped when it resulted in the concrete action of verification. This means that having a good plan or a critical mindset is not enough; those internal thoughts must be translated into the external behavior of checking and correcting. The researchers found that verification behavior explained a significant portion of the differences in how well students could transfer their knowledge to new problems. The model suggested that the benefit of being AI-literate comes not from the knowledge itself, but from the habit of using that knowledge to question and verify the tool's output.

These findings offer a clear picture of what separates successful learning from passive consumption in the age of artificial intelligence. The ability to use AI effectively is not a magic switch that guarantees learning. Instead, it is a tool that supports learning only when the student uses it to trigger a specific behavior: the active verification of information. When students check the AI's work, they are not just ensuring accuracy; they are engaging in the deep mental work required to understand complex principles. This behavior is particularly crucial when the goal is to apply knowledge to unfamiliar problems, where surface-level tricks fail. The study suggests that the future of education with AI may not be about teaching students more about the technology itself, but about designing tasks that require them to check, question, and verify the answers they receive, turning a passive interaction into an active learning experience.