Imagine two students. Both have the same mathematical problem in front of them, and both are using artificial intelligence. The first one writes: “Solve this problem for me,” and receives the answer in seconds. The second one asks: “Don’t tell me the answer. Ask me questions and guide me to the solution.” From the outside, we see the same technology being used in both cases. From a learning perspective, however, we might be dealing with two completely different processes.<\/p>\n
This is exactly where the main paradox of AI’s entry into education lies. Humanity has created a tool that can explain a complex topic to almost any student in a language they understand, find their mistakes, generate additional exercises, and, in a way, fulfill the role of a personal tutor… Yet, the very same tool can write an essay, solve a problem, summarize a book, and perform in seconds the work through which the student was supposed to learn. Therefore, the main question is probably no longer whether we should allow AI into education. It is already there… A more interesting question is: where is the line between assisting and outsourcing thought?<\/p>\n
A noteworthy phenomenon is occurring in education: a good result does not always mean that learning is actually taking place. If a student has to write a complex essay and creates an excellent text with the help of AI, the final product might indeed be better. And yet, this doesn’t tell us whether the student themselves learned better. This distinction is especially crucial now, when generative AI can not only find information but also perform a significant portion of the work.<\/p>\n
The OECD’s 2026 Digital Education Outlook<\/i> highlights this exact problem. Summarizing existing research, it notes that while general-purpose generative AI can often improve the quality of a student’s current assignment, this does not automatically translate into an increase in knowledge and skills. In some cases, the advantage disappears when the student no longer has access to AI. Take a simple example: If I don’t know an English word and look it up in a dictionary, the tool helps me find information. If I write a paragraph and ask AI to show me my mistakes, it provides feedback. But if I say, “Write this paragraph instead of me,” I am outsourcing the very cognitive work that might have been the objective of the learning process. That is why the question “Is AI good or bad for learning?” probably cannot have a single answer… It depends on what it is doing.<\/p>\nAI Can Actually Be a Great Teacher<\/h3>\n
Painting the picture solely in negative tones would be a mistake. In 2025, Scientific Reports<\/i> published a randomized controlled trial in which university physics students were divided into two groups. One group learned through an active learning format, while the other used a specially designed AI tutor. The AI did not simply hand out answers. The system was built on pedagogical principles: it asked the student relevant questions, provided feedback, and helped them step-by-step through the learning process. As a result, the AI-tutor group learned more in less time. The students also reported higher engagement and motivation.<\/p>\n
In another experiment published in PLOS ONE<\/i> in 2024, 274 participants solved mathematical problems. For assistance, they were given hints prepared either by humans or generated by ChatGPT. The AI assistance improved learning outcomes, and no statistically significant difference was found compared to the human-prepared help. This is an important distinction. AI can give the student the answer, or it can help them reach the answer themselves. In the first case, technology might be replacing thinking. In the second case, much like a teacher, it enhances the thinking process.<\/p>\nWhat Happens When We Outsource Complex Work to AI?<\/h3>\n
Here we move to the other side of the spectrum. In 2025, researchers at the MIT Media Lab conducted a small experiment that attracted significant attention. Participants were divided into three groups to write an essay: one group could use a large language model (LLM), another a search engine, and the third had to write the text without any digital assistance. The researchers also monitored brain electrical activity using EEG. In the LLM user group, functional connectivity metrics in the brain were weaker than in the group working independently. They also had more difficulty recalling quotes from their own text and reported a lower “sense of ownership” over the work.<\/p>\n
It is too early to draw far-reaching conclusions from this study. The number of participants was small, the paper was initially released as a preprint, and critical comments on its methodology have been published. Furthermore, an EEG difference does not directly mean that AI “damages the brain” or makes a person less smart. The study is interesting for another reason: it forces us to consider what happens when we reduce the mental effort that is itself a part of learning.<\/p>\n
Along the same lines, in another 2025 randomized study involving 120 university students, participants had to learn new material either with the help of ChatGPT or using traditional methods. In a surprise test 45 days later, the AI group scored an average of 57.5% correct answers, while the traditional group scored 68.5%. This too is a single study, not a final verdict, but the idea these results remind us of is not new to educational psychology at all. Learning requires effort. Sometimes, the very difficulty we want to eliminate is an essential part of the process.<\/p>\n
It is often said that “the brain is like a muscle and needs exercise.” Biologically, this comparison is inaccurate, but it correctly captures an important intuition about learning. When we try to remember something, solve a problem, construct an argument, or struggle with phrasing a text, we are not just creating a final answer. In the process, we are exercising the skill itself. Therefore, simplifying a task for a child and doing the task for them are not the same thing.<\/p>\n
This can be especially important regarding age. A child who has not yet developed the skills of writing, arguing, or problem-solving is in a different position from an adult who already possesses these skills and uses AI to speed up the process. Today, we do not yet have enough long-term neuroscientific data to say exactly how generative AI will alter children’s developing brains. Therefore, claims suggesting we already know ChatGPT makes children’s brains “lazy” are getting ahead of the science. Instead, we have grounds for a more cautious question: which cognitive tasks must the child absolutely perform on their own?<\/p>\n
The simple solution here might seem to be a ban. The problem is that students will use AI outside of school anyway. UNESCO’s approach is increasingly less focused solely on banning. One of the main directions in the organization’s AI competency framework for students is precisely the critical, responsible, and mindful use of artificial intelligence. This suggests a certain shift in educational philosophy as well. If we didn’t cancel teaching mathematics after the invention of the calculator, perhaps the existence of AI shouldn’t cancel writing and thinking, but rather change how<\/i> we teach them.<\/p>\n
For example, a student might be asked to evaluate an AI-written answer: find mistakes, compare different responses, and justify why they disagree with the model. Perhaps they should write their own text first and then use AI as a critic. Or maybe they could fact-check the AI’s response against sources. In this scenario, the technology is no longer an automated answer-dispenser. It becomes an object for the student to think about critically.<\/p>\n
If we give a student a homework assignment to “write an 800-word essay on climate change,” today we can no longer be sure what we are actually testing. Their knowledge? Their writing skills? Their ability to use AI? Or simply how well they can pass off a generated text as their own work? Consequently, generative AI is asking uncomfortable questions of the traditional assessment system.<\/p>\n
It might become more important to focus not just on the final product, but on the process: how the student arrived at the answer, why they chose a specific argument, how they verified the information, and whether they can defend their conclusion. Oral discussions, classroom work, projects, comparing different versions of a text, and explaining one’s own decisions could become much more significant. In this sense, AI is not just a problem. It is a magnifying glass for the problems already existing within the education system. Perhaps our education system rightly lacked more interaction all along.<\/p>\n
A few years ago, one of the main functions of education was delivering information. Today, almost any student might have a system in their pocket that can explain a black hole in seconds, write software code, translate text, or suggest arguments for an essay. In this world, acquiring information is becoming increasingly easy. Evaluating it, however, might become harder. Who wrote it? How do we know? Is it reliable? What is it leaving out? Why should I believe it? Could the answer sound convincing and still be wrong?<\/p>\n
Therefore, in the future, one of the most important skills might not be answering quickly, but rather asking good questions, verifying sources, evaluating arguments, and defending one’s own opinion. AI will probably not disappear from education, nor should we expect a world where children simply won’t use it. Thus, we don’t have a choice between “AI or learning.” The real choice lies elsewhere. We can give a child a system that thinks for them… or we can teach them how to use that same system to think better themselves. At first glance, the difference is small. But for education, this might turn out to be the ultimate difference.<\/p>\n","protected":false},"excerpt":{"rendered":"
Imagine two students. Both have the same mathematical problem in front of them, and both are using artificial intelligence. The first one writes: “Solve this problem for me,” and receives the answer in seconds. The second one asks: “Don’t tell me the answer. Ask me questions and guide me to the solution.” From the outside, […]<\/p>\n","protected":false},"author":30,"featured_media":25785,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[1651],"tags":[1900],"class_list":["post-25786","post","type-post","status-publish","format-standard","has-post-thumbnail","category-insight","tag-artificial-intelligence"],"acf":[],"_links":{"self":[{"href":"https:\/\/medscriptum.org\/en\/wp-json\/wp\/v2\/posts\/25786","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/medscriptum.org\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/medscriptum.org\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/medscriptum.org\/en\/wp-json\/wp\/v2\/users\/30"}],"replies":[{"embeddable":true,"href":"https:\/\/medscriptum.org\/en\/wp-json\/wp\/v2\/comments?post=25786"}],"version-history":[{"count":1,"href":"https:\/\/medscriptum.org\/en\/wp-json\/wp\/v2\/posts\/25786\/revisions"}],"predecessor-version":[{"id":25790,"href":"https:\/\/medscriptum.org\/en\/wp-json\/wp\/v2\/posts\/25786\/revisions\/25790"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/medscriptum.org\/en\/wp-json\/wp\/v2\/media\/25785"}],"wp:attachment":[{"href":"https:\/\/medscriptum.org\/en\/wp-json\/wp\/v2\/media?parent=25786"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/medscriptum.org\/en\/wp-json\/wp\/v2\/categories?post=25786"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/medscriptum.org\/en\/wp-json\/wp\/v2\/tags?post=25786"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}