Learning in the Age of Generative AI

Combining lived classroom experience with scientific evidence, this article examines how generative AI, if used wisely, can become a genuine partner in learning and critical thought.

What happens when a tool capable of writing an essay in seconds, explaining a mathematical procedure, or clearly unpacking a complex scientific paper becomes part of everyday life in the classroom? This is no longer a hypothetical question but a reality faced today by students, teachers, and parents alike. Generative artificial intelligence (AI) has entered schools and radically altered the rules of the game. Some welcome its potential for personalized learning and instant support, while others see it as a threat to critical thinking and academic integrity. Is AI an ally of knowledge, or merely a convenient shortcut that harms us in the long run?

The article avoids one-sided answers and instead offers an interpretation built on a combination of two perspectives embodied by its authors. On the one hand, it provides a direct glimpse from the school desk through the personal experience of a Ljubljana high-school student, revealing how AI is practically reshaping essay preparation, problem-solving, and language learning. On the other hand, it situates this personal perspective within a broader framework: a systematic overview of the latest scientific studies, meta-analyses, and international guidelines that highlight the measured effects, risks, and recommendations for responsible use.

Our aim is not to deliver final judgments but to describe the present state as perceived from both of these vantage points. In conclusion, we also offer our shared understanding of these developments and our conviction that mastering the dialogue with artificial intelligence has become a key literacy of the 21st century.

Artificial Intelligence as a Personal Tutor

In this chapter, we present concrete examples of how students practically use artificial intelligence in their everyday learning challenges: from preparing essays and writing homework to solving mathematical problems, learning languages, and organizing notes. This is a perspective through the eyes of a high-school student who employs AI as a tool for deeper understanding and learning—not merely for obtaining answers. Her experience serves as a starting point for broader reflection on when such use is genuinely meaningful, where it can contribute to better learning, and when there is a danger of it becoming merely a shortcut that bypasses understanding. It has become evident that AI holds the greatest value as a personal tutor precisely where it does not replace the cognitive process but rather stimulates, guides, and helps to disentangle complex concepts.

Preparing for an Essay: “A Socratic Dialogue with Artificial Intelligence”

For many years, it was taken for granted that one could not truly prepare for a high-school essay. Success was believed to depend primarily on personal talent and momentary inspiration. Preparation usually meant reading the required works and searching for the occasional analysis on internet forums—analyses that were often superficial and, in Slovenia at least, long since read by most students.

With the advent of artificial intelligence, this perception has radically changed. Today, we can converse with AI about literary works in a manner reminiscent of a Socratic dialogue—through questions, doubts, counterarguments, and the pursuit of clarity. AI is no longer just a passive source of information but becomes an active interlocutor that enables a deeper understanding of texts and the development of independent thought.

When it comes to lesser-known works, especially Slovenian ones, it is sensible first to provide AI with the full text. This allows it to respond directly on the basis of the original material and avoid confusing characters, events, or meanings. For world classics, this is usually unnecessary, as such works are already well represented in the models’ knowledge.

AI can efficiently summarize content, analyze characters and their relationships, identify the core ideas and values of a work, and help unpack its literary and philosophical layers. Yet its greatest strength lies elsewhere: it does not impose uniform interpretations but instead stimulates reflection, poses additional questions, and helps students shape their own interpretations—expressing them more clearly and convincingly.

Moreover, AI is not limited to standard analyses. It can assist in formulating answers to more open-ended and creative questions often found in essay prompts: how literary characters might behave in the modern world, how a contemporary reader perceives a certain work, or what relevance its message holds today. In this way, AI is not merely a tool for analysis but becomes a partner in developing argumentation and understanding context.

Preparing for an essay with artificial intelligence is therefore no longer about passively reading summaries but about engaging in a dynamic and creative dialogue, where the student develops personal viewpoints, tests them critically, and refines them. If in the past success relied largely on talent, today that talent can be enhanced through the well-directed use of AI—an instrument that opens new pathways to understanding literature and the world.

Writing Entire Essays with AI

In addition to preparation, many students also use artificial intelligence in a more direct—yet riskier—way: by simply entering the essay prompt and letting the system generate the entire piece. The result is often surprisingly polished and seemingly convincing work, something that could be handed in to a teacher with minimal edits, saving precious time. Yet in practice, this strategy rarely proves successful.

Essays written entirely by AI often differ stylistically and structurally from what is expected in Slovenian schools. They may include subheadings, rely on generic quotations, or present overly generalized claims that are insufficiently grounded in the text itself. More importantly, such essays frequently fail to follow the teacher’s specific instructions—a key criterion in assessment. As a result, teachers often quickly suspect that the essay was not written by the student but by artificial intelligence.

Such essays are often graded lower than those written by students themselves. The reason lies not only in inappropriate form but also in the absence of personal reflection, originality, and an authentic voice. An AI-written essay may be linguistically flawless, but it often lacks what a good teacher notices immediately: the student’s genuine intellectual engagement.

For this reason, most students use AI in this way mainly for assignments that are not directly graded—for instance, within compulsory elective activities. In the second year, we were required to submit an essay on a film watched in class. Pressured by numerous other obligations, some students chose to have AI write it for them. Yet the teachers rejected these submissions, rightly suspecting they were not original. The students had to rewrite their essays under supervision at the end of the year, which meant extra work, stress, and wasted time.

Such use of AI is therefore ineffective in the long run. Not only does it often fail to deliver the desired results, it also contributes nothing to knowledge, intellectual development, or writing skills. Still, it must be acknowledged that many students turn to this option not out of laziness but because of overload: constant assessments, tight deadlines, and numerous extracurricular commitments. In such moments, artificial intelligence primarily represents a convenient solution to a lack of time and a way to secure a bit more peace or sleep.

This is precisely why it is all the more important not to treat AI as a substitute for one’s own work, but as a tool that supports and enhances our abilities. Writing an essay is not merely a school assignment but an exercise in thinking, understanding, and expression—and it is in this role that AI can become an ally, rather than an escape from learning.

Solving and Explaining Science and Math Problems

Beyond literary analysis and essay writing, artificial intelligence can also provide significant support in subjects such as mathematics, physics, and chemistry. This is often where difficulties arise: we may not know how to start a problem, which formulas to use, or how to proceed once we get stuck midway. Sometimes we even reach the wrong result without realizing where we went wrong—and these small, hard-to-detect mistakes are the most frequent source of frustration.

Since a teacher or classmate is not always available—especially not in the afternoon at home or right before a test—AI can step in as a patient explainer and tutor. If we provide it with the text or a photo of a problem, it can not only solve it but also carefully explain the process: step by step, with a clear rationale at each stage. This way, we don’t just get the correct answer—we also understand the path that leads to it.

This feedback loop is immensely valuable: it allows us to quickly spot where we made mistakes and why. Detecting errors in our own work is time-consuming and often unfeasible, especially when we don’t know what we’re actually looking for. In this regard, AI can save us a great deal of time while deepening our understanding, enabling us to learn from our mistakes instead of merely stressing over them.

It is also highly useful for assignments where only the final result is provided but not the full solution process. In such cases, AI can reconstruct the reasoning from beginning to end, allowing us to compare its approach with our own. This makes it possible to verify our work, correct missteps, and strengthen our comprehension.

For more complex problems involving branching procedures, multi-step reasoning, or a degree of creativity, so-called “thinking models” of AI are especially valuable. These advanced approaches deliberately extend the processing time and carry out multiple layers of reasoning in the background, often through a technique known as chain-of-thought prompting. Instead of simply presenting an answer, the system “thinks out loud”—laying out each step while continually checking the consistency of its reasoning. This approach may take longer, sometimes several minutes, but the reward is a carefully considered and logically structured response that reveals not only the final result but also the entire line of reasoning.

Such answers are extremely useful not only for solving individual tasks but also as learning aids when preparing for tests, final exams, or oral defenses. By explaining its reasoning, the model allows the student to follow the thought process, learn from it, and simultaneously check their own understanding. In this way, AI becomes a tool for strengthening analytical thinking rather than merely a supplier of quick solutions.

All of this demonstrates that in science and mathematics, AI does not function merely as a “calculating machine,” but as an instrument for developing understanding—which is, after all, what matters most in learning.

Learning Foreign Languages

Learning a foreign language often requires mastering a vast amount of new vocabulary. Simply reading or passively listening usually isn’t enough, as words are quickly forgotten without active use. This is why flashcards have long been a popular tool: they promote memorization through repetition and recall. Yet preparing such cards takes considerable time—and this is precisely where artificial intelligence can help effectively.

With AI, we can quickly generate exercises tailored specifically to our needs: we provide a list of words we want to learn and ask the system to present them with explanations, translations, or usage examples in random order. We then try to recall the correct word or translation. This method encourages active retrieval, which is proven to enhance long-term memory.

AI is not limited to basic flashcards. If we show it examples of exercises from past tests or workbooks, it can create new tasks in the same format but with fresh vocabulary that we are currently studying. Such individualized practice allows us to learn in a targeted and efficient way rather than mechanically.

Interestingly, artificial intelligence is no longer used for language learning only by students but also by some teachers. Occasionally, a test problem will even include the note: “created with ChatGPT.” This shows that AI is not just a tool for reinforcing knowledge but is becoming part of the broader learning environment—an instrument used by both teachers and students.

Writing is also a crucial part of language learning. When writing in a foreign language, we often repeat the same mistakes, which are difficult to spot on our own—especially when writing by hand. Since teachers, due to time constraints, typically review only a few written assignments, AI can step in as an additional language mentor. If we provide it with a photo of a handwritten text, it can transcribe, correct, and explain the errors. In this way, we don’t just receive an improved version of the text but also gain an understanding of why something was wrong and how to improve it in the future.

With regular use of artificial intelligence for writing and vocabulary practice, we can deepen our knowledge of a foreign language more quickly and more permanently. This means that AI does not replace learning—it complements and strengthens it.

Collecting Information and Organizing Notes

In some subjects—especially biology, geography, history, or social sciences—students must master large amounts of material. Information is often scattered across different sources: textbooks, worksheets, websites, or densely written paragraphs without clear structure. Such material is difficult to digest, requiring considerable time just to understand the basics, let alone to memorize effectively.

In these cases, artificial intelligence proves extremely useful. We can provide it with fragmented notes or raw texts, and it can turn them into neatly organized, logically structured summaries or tables. It sorts data into meaningful categories, organizes them by comparisons, time periods, functions, classes, or features—in short, it reshapes information so that it becomes comprehensible and easier to remember.

For example, in biology we had to compare different organ systems across animals, from sponges and cnidarians to amphibians and mammals. Manually collecting and writing these details into comparative tables would have taken a great deal of time. AI, however, can generate such a table in seconds from the sources we provide (or those already in its training data), showing the differences in digestion, respiration, circulation, excretion, and nervous systems across animal groups.

This leaves the student with more energy and time for understanding and learning. AI-prepared material can be easily exported into Word, PDF, or Excel, and then supplemented, edited, printed, or shared with classmates. This not only supports more effective studying but also facilitates collaboration and knowledge exchange. In this way, AI becomes a digital assistant that doesn’t replace the student’s work but simplifies it where tasks are mechanical and time-consuming, freeing attention for what truly matters: comprehension, connections, and independent thinking.

Just before this article went to press, OpenAI introduced a new feature called Study Mode, designed for targeted learning with artificial intelligence. The system first asks the user about the topic they want to understand, as well as their age or educational level. Based on this, it begins a conversation in the form of questions and feedback. The user first attempts to answer the questions on their own, after which the system offers clarifications, corrections, and prompts for further reflection. What makes this approach distinctive is its use of the Socratic method, which fosters deeper understanding, self-reflection, and active engagement through a series of guiding questions.

Learning with this mode becomes more interactive and in-depth, since detailed instructions are no longer required—the AI itself takes on the role of a tutor, adapting in real time to the learner’s knowledge and pace. The feature was developed in collaboration with several educational institutions and is grounded in the latest didactic insights. It is available to all ChatGPT users regardless of subscription tier and will soon be integrated into the educational version, ChatGPT Edu. This approach not only facilitates comprehension of complex content but also strengthens independent thinking, nurtures curiosity, and encourages the responsible use of AI in education.

Systemic Challenges, Measurable Effects, and Potential Risks

The impact of artificial intelligence on education is not merely theoretical; its effects are already concrete, measurable, and often multidimensional. In this chapter, we illuminate two sides of the same coin. On one hand, there are promising data showing improved learning outcomes, increased motivation, and more personalized support for individual learners. On the other hand, serious questions arise: does excessive reliance on AI foster superficial learning? Do we risk losing independent thought, perseverance, and critical autonomy? It is precisely this tension between opportunities and pitfalls that provides a crucial starting point for deeper reflection on AI’s role in the future of education.

Meta-Analyses and Research on the Pedagogical Effects of AI

The first large-scale studies on the impact of generative artificial intelligence (AI) on learning show predominantly encouraging results. The most reliable insights come from the method of meta-analysis, which aggregates findings from numerous individual studies to assess the overall effect of a given method or technology on learning outcomes.

A meta-analysis that included 65 independent studies (Sun & Zhou 2024) estimated that the use of generative AI improves student performance by about half a standard deviation on average (Hedges’s g ≈ 0.53). This is a moderate yet meaningful effect, comparable to raising an average student from the 50th to the 69th percentile in achievement. The strongest effects appeared in tasks involving text generation (e.g., essay or report writing) and in self-directed learning, where AI acts as a tutor providing real-time feedback.

Another meta-analysis (Zhang, Jantakoon & Laoha 2025) systematically examined the effectiveness of AI technologies in education, drawing on 13 empirical studies from eight countries. The research revealed a significant and large overall positive effect (g ≈ 0.86), highlighting the considerable advantages of integrating AI into educational processes. The most pronounced impact was observed in chatbots and generative AI (g ≈ 1.02), while online learning environments and virtual reality showed moderate effects (g ≈ 0.79). The authors note substantial variability among the analyzed studies but emphasize the robustness of positive outcomes and the importance of considering contextual factors when implementing AI solutions.

Other systematic studies, focused specifically on conversational agents such as ChatGPT, distinguish between cognitive and non-cognitive effects. Analyses of 27 empirical studies (Suo, Yin, Wang, et al. 2025 [preprint]), published between 2022 and 2025, confirm that these agents have a significant positive influence on learning. On the cognitive level (measured by grades, task completion time, etc.), they report moderate positive effects (g ≈ 0.36). Even greater effects (g ≈ 0.52) appear in non-cognitive outcomes such as motivation, perseverance, and self-efficacy—that is, the belief that one is capable of successfully completing tasks. These findings suggest that students often perceive generative AI as a safe, non-judgmental environment that encourages them to engage more with learning material and reduces fear of making mistakes.

Among individual skills, writing stands out as a particularly promising area. Randomized controlled trials (RCTs), considered the gold standard for measuring effectiveness, show that students who receive feedback from AI while writing essays achieve better structure, cohesion, argumentation, and content development compared to those who receive only traditional feedback (e.g., from teachers or peers). Beyond improvements in text quality, studies also highlight higher levels of motivation and engagement among students using AI systems, though they also report mixed emotional responses (Zhang 2025; Lo, Wong & Chan 2025).

In the context of foreign language learning, a large meta-analysis of 31 comparative studies (Lyu, Lai & Guo 2025) confirms that chatbots have a moderate positive effect (g ≈ 0.61) on cognitive skills such as writing and vocabulary acquisition, as well as on affective factors such as motivation and interest. The effects are strongest when interaction with AI is guided and reflective, enabling learners not only to receive immediate feedback on errors but also to understand their causes. Such a deeper approach is made possible primarily by advanced chatbots using generative AI, accessible via mobile devices—identified by the meta-analysis as key factors in boosting effectiveness.

The overall conclusion of these studies is clear: the effects of generative AI use are on average positive and moderate, with the greatest benefits appearing when the technology is not employed passively but as a tool to support comprehension, writing, and active learning. The teacher’s role in guiding this use remains crucial for achieving optimal results and avoiding merely superficial knowledge.

Mechanisms of Effectiveness and Limits of AI Use in Education

At first glance, findings on the impact of artificial intelligence (AI) in education seem riddled with contradictions, creating considerable confusion. On the one hand, large-scale meta-analyses that synthesize results from hundreds of individual studies generally report moderate but statistically significant positive effects on learning.

On the other hand, high-profile field experiments such as the Wharton study (Bastani et al. 2024) paint a far more complex and troubling picture. In this particular study, researchers found that while access to a general model such as GPT-4 temporarily improved students’ performance on assignments, the same group performed worse on a final exam—taken without AI assistance—than students who had never used the tool. This suggests that uncritical use of generative AI, which primarily facilitates access to final answers, may in the long run undermine the development of independent knowledge and problem-solving skills.

The apparent contradiction between these findings is therefore not necessarily a paradox but underscores a key insight: what matters is not simply whether AI is used, but how it is designed and integrated into the learning process. While carefully developed pedagogical tools—with built-in safeguards that encourage critical thinking—can genuinely support learning, unrestricted use of general-purpose systems may lead students toward passive shortcuts, hampering their cognitive growth.

The positive effects of generative AI in education are not accidental; they stem from the technology’s ability to strengthen and automate already established didactic principles. Successful implementations foster active learning, provide immediate feedback, allow for personalized learning paths, and reinforce self-regulation skills.

A study by Iqbal et al. (2025), involving 465 pre-service teachers in China, examined in more detail the impact of these tools on learning outcomes. The researchers found that generative artificial intelligence enhances achievement primarily through two key mechanisms. The first is cognitive offloading, where the tool takes over routine tasks such as information retrieval or syntactic checking, enabling learners to direct their mental resources toward more demanding processes such as critical analysis and the synthesis of ideas. The second mechanism is shared metacognition, in which the technology acts as a catalyst for collaborative learning by encouraging students to reflect together, evaluate one another’s ideas, and coordinate problem-solving strategies. In this way, generative AI functions not merely as an individual tool but as a platform that strengthens collective intelligence and deepens understanding within a learning group.

The key question is not whether a student uses AI, but which part of the cognitive process is delegated to it. Beneficial offloading occurs when AI takes over routine, low-level tasks—for example, text formatting, synonym searching, summarizing a source, or carrying out a familiar calculation. In such cases, AI expands the learner’s cognitive capacity, allowing greater focus on higher-order thinking such as planning, argumentation, synthesis, and conceptual integration. Harmful offloading, however, occurs when AI assumes responsibility for the core intellectual work—such as understanding a problem, formulating a research question, developing a thesis, or constructing an argument. In these cases, AI replaces thinking rather than supporting it. The student does not practice fundamental cognitive skills, and genuine learning simply does not take place. The tool becomes a cognitive crutch.

In addition, generative tools are prone to the phenomenon known as hallucination, in which they produce information with great confidence that is in fact false, fabricated, or lacking any factual basis. This may involve citing non-existent sources, misinterpreting facts, or generating entirely invented data. Consequently, when these tools are used for academic purposes such as essays and research assignments, rigorous verification of all citations, references, and factual claims is essential. This necessity underscores the importance of advanced information and data literacy, enabling users to critically evaluate and validate content.

Generative AI as a Learning Partner: Conditions for Effective Use

The central message of this article has been confirmed across examples and analyses alike: generative artificial intelligence contributes most to education when it shifts from being a passive tool to becoming a partner in thought. It does not write instead of the student but engages with them in dialogue, asking questions, checking understanding, offering counterarguments, and helping to refine initial ideas into clear, original expression. In such a relationship, AI is no longer merely a “task performer” but a co-creator of the learning process—while the student retains the central role of deciding, judging, and taking responsibility.

This mode of use rests on three interlinked didactic principles that together create a learning environment in which the student actively participates in the cognitive process. The first principle is learning dialogue, where the student and AI engage in a sequence of questions, explanations, counterarguments, and revisions. This process promotes active recall, reduces the risk of superficial understanding, and counters the illusion of certainty that may arise from fluent yet shallow answers. The second principle is traceability of the thought process. Drafts, revisions, and explained changes provide insight into the path of reasoning, creating conditions for learning from mistakes, fostering reflection, and enabling the gradual construction of knowledge. The third principle is dual verification. A student first formulates a claim with the help of AI, but must then be able to articulate and defend the same claim independently. Such repetition under identical standards ensures that knowledge is genuinely internalized and transferable to new contexts.

When this logic is transferred into the classroom, the focus shifts from products to processes—from simply having “solved tasks” to explaining methods, weighing alternatives, and applying knowledge in new situations. Assessment practices must adapt accordingly: drafts, ongoing feedback, reflections, and oral defenses gain greater importance, while AI-free tests remain essential to show how much knowledge persists in independent use. This reduces the gap between quick execution with AI and genuine understanding without it.

In such use, another often-overlooked dimension comes to the fore: the emotional literacy of models and the importance of human connection. Communication that is encouraging, clear, and respectful can strengthen a student’s perseverance and confidence. By contrast, a patronizing, sarcastic, or overly self-assured tone can quickly lead to passivity and intellectual withdrawal. Well-designed use of AI should therefore take over routine tasks—such as language corrections, basic error marking, or literature gathering—and free time for teachers and students to focus on what no model can replace: conversation, mentorship, and shared thinking.

Yet such use of AI requires not only sound pedagogical practice but also ethical reflection. It demands careful handling of personal data, transparency of processes, and equitable access to high-quality tools. Early phases of new technologies often favor those with better devices or access to paid model versions. Long-term fairness does not happen on its own; it requires school-level licenses, reliable infrastructure, and didactic support that make new learning patterns accessible to all, not just the best equipped and most motivated. For Slovenian students, solid language support for Slovene in AI models is an advantage, but linguistic accessibility in commercial systems cannot replace the need for systemic regulation in this field.

Although this article provides a clear didactic framework, the discussion does not end here—it truly only begins. Key challenges remain that demand thoughtful answers: Where does assistance in thinking end and replacement of thinking begin? How should assessments be designed to balance evaluation of the learning process (where AI is a partner) with the demonstration of independent knowledge? How can developers design interfaces that encourage deeper understanding rather than merely generating quick answers? And ultimately, how can we guide this collaboration with AI so that it strengthens students’ intellectual autonomy and their capacity for dialogue—both with people and with machines?

References

• Bastani, Hamsa, Osbert Bastani, Ahmet Sungu, Hao Ge, Özge Kabakcı, and Rick Mariman. Generative AI Can Harm Learning. SSRN Working Paper, The Wharton School, 2024.

• Iqbal, Junaid, Zohaib Farooq Hashmi, Muhammad Zubair Asghar, and Muhammad Nauman Abid. “Generative AI Tool Use Enhances Academic Achievement in Sustainable Education through Shared Metacognition and Cognitive Offloading among Preservice Teachers.” Scientific Reports 15 (2025): 16610.

• Lo, N. P. K., Andy Wong, and Samuel Chan. “The Impact of Generative AI on Essay Revisions and Student Engagement.” Computers & Education: Open 100249 (2025).

• Lyu, Bin, Chun Lai, and Jingjing Guo. “Effectiveness of Chatbots in Improving Language Learning: A Meta-Analysis of Comparative Studies.” International Journal of Applied Linguistics (2024).

• Sun, Lin, and Lijun Zhou. “Does Generative Artificial Intelligence Improve the Academic Achievement of College Students? A Meta-Analysis.” Journal of Educational Computing Research (2024).

• Suo, Xiaochen, Baoyuan Yin, Wanqing Wang, et al. “Exploring the Impact of Generative AI-Powered Conversational Agents on Student Learning: A Systematic Review and Meta-Analysis Grounded in Activity Theory.” Authorea, July 1, 2025.

• Zhang, Jincheng, Thada Jantakoon, and Rukthin Laoha. “Meta-Analysis of Artificial Intelligence in Education.” Higher Education Studies 15 (2) (2025): 189–203.

• Zhang, Kai. “Enhancing Critical Writing through AI Feedback: A Randomized Control Study.” Behavioral Sciences 15 (5) (2025): 600.

Translated from the Slovene original, available here: Umetna inteligenca v šoli.