Imagine a student in the library preparing an assignment. She asks ChatGPT to explain a concept and receives, in seconds, a clear, organized, and convincing response. She believes she has found a reliable explanation. However, the fluency with which it is written does not guarantee its accuracy. Does the student know how that response was generated?
Language models generate text based on patterns learned during training and the context with which they interact. They can offer useful explanations, but they can also produce errors just as easily. A lack of understanding of how they work makes it easy to mistake a convincing answer for proven knowledge. Using AI fluently doesn’t mean understanding it.
How AI is used in the classroom
Our recent research has shown that university students use AI extensively and are satisfied with the experience. Most highlighted that these tools save time, are easy to use, provide immediate answers, and help in understanding content.
AI was already part of the academic routine of the students surveyed at the beginning of 2025. However, our most recent results show that this regular use coexists with limited training and that few students reported a high level of proficiency with these tools.
Familiarity does not imply knowledge
ChatGPT is the most frequently used AI tool among the students in our study, and almost exclusively, due to its widespread recognition. However, most students valued their knowledge and proficiency with this type of chatbot when using it at intermediate or lower levels. These are distinct aspects: one refers to what they believe they know about AI; the other, to their perceived ability to use it. In any case, using it regularly does not imply knowledge or mastery.
Almost all students reported not having received any specific training. However, of those who had, the vast majority had received their training outside of the university. Furthermore, two out of three students stated that the ethical implications of AI had not been addressed during their university studies.
These results challenge the convenient label of “digital natives.” Growing up surrounded by screens allows one to use applications with ease, but it doesn’t, in itself, teach how responses are generated, what data systems are trained on, why they might fabricate information, or what social, racial, and gender biases they might reproduce. Even the appearance of competence can be misleading for the user: knowing how to obtain an answer is not the same as knowing how to evaluate it.
When the answer sounds too good to be true
The students themselves sensed part of the problem. They feared plagiarism, doubted the reliability of the information, or were concerned about academic integrity. But recognizing a danger doesn’t mean having strategies to deal with it.
Not understanding how AI works can lead us to misinterpret its responses. If we believe it functions like a search engine retrieving verified information, we might accept a fabricated reference as if it came from a real publication. If we ignore the fact that training data can contain biases, we might mistake a discriminatory response for a neutral explanation.
It’s also important to understand what happens to the information we enter. Sharing students’ personal data to obtain a tailored activity requires knowing the service’s privacy policy. These risks illustrate why digital literacy must go beyond learning how to formulate requests: we need to understand the tool’s limitations to decide how to use it and which results require verification.
Learning to use AI changes its use
The answer isn’t to ban a technology already integrated into academic life. It’s about digital literacy. That means understanding the basics of how it works, formulating requests, verifying results, comparing sources, detecting biases, protecting data, and deciding when it’s best not to use it.
The available evidence indicates that such training can make a difference. A 90-minute university workshop improved students’ confidence, their understanding of the norms, and their ability to recognize acceptable academic uses. But teachers also need training. In another study with 90 prospective teachers , generative AI literacy led to improvements in technical and ethical competence, critical evaluation, and creative application.
For example, the use of artificial intelligence tools for classroom work can be permitted, but with very precise instructions. In a practical exercise with university students in Primary Education, we asked the students to request an assessment rubric from the AI that was consistent with a learning situation previously designed in groups.
First, they must be provided with the context, objectives, knowledge, and assessment criteria; then, the result must be critically examined. It is not enough to ask them to create “a good rubric”; it is necessary to specify what learning outcomes are to be observed.
It’s also wise to be wary of expressions like “adequate” or “correct,” and of verbs that are difficult to observe, such as “understands” or “thinks.” Comparing the response with the official curriculum and checking whether the criteria are relevant, the descriptors observable, and the levels truly progressive allows for the detection of errors before implementing the proposal in the classroom.
AI literacy, a global priority
UNESCO brings these capabilities together in a competency framework that places autonomy and human judgment at its core. Universities, therefore, cannot simply publish plagiarism policies or boast about their digital skills. They must integrate AI literacy into their curricula, apply it to real-world problems in each profession, and evaluate the process followed, not just the final product.
Author Bios: Juan José Lorenzo Castiñeiras is Professor of Educational Sciences, Department of Pedagogy and Didactics, Area of Didactics and School Organization, Andrea Fernández-Sánchez is Professor of Didactics and School Organization and Isabel Fernández-Menor is Assistant Professor (PhD) in the Department of Pedagogy and Didactics all at the University of A Coruña
