Experiment: Using AI assistants makes the brain "lazy" and therefore interferes with learning

Contents:

• How the experiment with essays and ChatGPT was conducted
• What the experiment participants could later remember about their essays
• What essays did students produce under different conditions
• How the brain worked according to EEG data
• What conclusions did the study authors draw?
• Who else thinks it is dangerous to delegate intellectual tasks to AI assistants?
• What should be done about all this in education?

In June 2025, an article titled "Your Brain on ChatGPT: Accumulating Cognitive Debt When Using an AI Assistant for Essay Writing" was posted on the preprint publishing platform ArXiv. In this study, scientists from the United States, primarily from the Massachusetts Institute of Technology (MIT), argue that people who frequently turn to AI assistants for educational tasks experience a decline in learning ability. This underscores the importance of consciously using artificial intelligence technologies in the educational process. The study highlights the potential risks of AI addiction and the need to develop critical thinking in students.

It should be noted that this work is one of many studies examining the impact of artificial intelligence, particularly generative neural networks, on human thinking and brain function. However, it has attracted particular attention due to its shocking findings. The experiment found that participants who collaborated with an AI assistant to write essays exhibited significant difficulty memorizing their essays. Furthermore, their brain activity was significantly lower while working with the AI ​​compared to other participants in the experiment. These results indicate that using generative neural networks not only reduces mental activity while working, but may also contribute to the development of a habit of laziness in the thinking process, which may negatively impact cognitive abilities in the long term.

The study's authors, on their website, strongly request that terms such as "stupider," "brain decay," or "brain damage" not be used when describing their findings. Their experiment focused on just one learning task and lasted only four months, with small participant groups. The question of how the use of artificial intelligence in other contexts affects the brain and the stability of the identified effects remains open and requires further research.

The study provides several key findings for the field of education. Let's take a closer look at the aspects examined, what exactly the authors discovered, and what questions remained unanswered.

How the Essay and ChatGPT Experiment Was Conducted

The study, led by Natalia Kosmina of the MIT Media Lab, followed a clearly structured plan. Modern methods and technologies were used to obtain reliable results. The research team analyzed the data to identify key trends and patterns, which will deepen our understanding of the topic and contribute to further research in this area.

Fifty-four people aged 18 to 39 took part in the study. 60 participants initially responded, but six did not complete the study. Of the participants, 35 were undergraduate students, and the remainder were older, including employees of MIT and other universities and colleges. Participants were randomly assigned to three groups, each of which was asked to write an essay under different conditions.

• with access to a chatbot based on a large language model (OpenAI's GPT-4o was used), but no other websites;
• without a chatbot, but with simple internet access (with websites with tools based on generative neural networks blocked);
• without access to any online resources at all, that is, independently.

As part of the experiment, participants wrote three essays in three identical sessions. For this, nine different topics from the SAT (Scholastic Aptitude Test - an academic assessment test for high school students in the USA, similar to our Unified State Exam) were selected. This approach allowed us to assess each participant's writing and argumentation skills and identify trends in their approaches to various topics. The SAT is an important tool for assessing academic achievement, and using its topics in this study helps better understand how high school graduates form and express their thoughts in writing.

Participants in the experiment were asked to answer important questions, such as the need for unconditional support for true loyalty, the possibility of achieving happiness without contributing to the lives of others, and the influence of art on life change. During each of the three sessions, participants independently chose one of three proposed topics, ensuring a unique approach for each. They were allotted 20 minutes to write the text.

The authors argue that essay writing illustrates both the advantages and disadvantages of using artificial intelligence in the educational process. On the one hand, the use of AI significantly speeds up the work process, allowing for a quick draft. On the other hand, the ability to produce a coherent text without much effort may lead students to not delve into the topic. This can reduce critical thinking and independent learning. It is important to find a balance between the use of technology and the need to develop students' analytical skills.

During all three sessions, participants underwent electroencephalography (EEG), which allowed them to record their brain's electrical activity. This also meant that participants were restricted in their movements while writing essays, as excessive activity could negatively impact EEG results. It is important to consider the impact of such factors on the accuracy of the data obtained to ensure high-quality research in neuroscience and psychology.

After completing the work, the researchers interviewed each participant, and the completed essays were assessed twice: once by a professional teacher and once by artificial intelligence. During the interviews, participants were asked to what extent they perceive themselves as the authors of their essays, their ability to cite their work, and their ability to summarize the key ideas of the text. This study aims to understand the relationship between the writing process and perceptions of authorship, which has implications for further research in the field of education and the use of technology in learning.

Participants who used neural networks and the Internet were asked about their work process: whether they started independently or immediately turned to a chatbot, which websites they used, and whether they edited the received content.

The composition of the groups remained unchanged throughout the three sessions: participants initially assigned to the group working with neural networks wrote essays using a chatbot throughout all three meetings. The researchers subsequently held a fourth additional session, to which they invited some of the participants who had the time and desire to join.

The article mentions that all 18 participants in the final session had previously written essays using GPT-4o or independently. However, some tables with data for this session present information about group members who used the internet for the first three times. This creates uncertainty regarding the exact composition of the participants in the final session. More detailed explanations may be provided in the final publication.

The fourth stage of the experiment was devoted to observing the participants' work under new conditions. Participants with experience working with neural networks were asked to write the essay independently. Meanwhile, those who had previously had no internet access now completed the task using GPT-4o. This allowed them to assess the impact of modern technologies on the writing process and identify differences in approaches to text creation.

Each participant was asked to choose one of three essay topics on which they had already written. There were six more unused topics for each. Participants were not informed that this was the assignment. The interview after the fourth session became more in-depth: participants were asked questions about how they compared their approaches to writing a new and previous essay on the same topic, and were also asked which result they liked more. The entire experiment, covering all sessions, was completed within four months.

What the experiment participants could later remember about their essays

The results of the interviews with the participants, conducted after writing the first essay, showed significant differences between the group using neural networks and the other two groups. Participants using neural networks noted an improvement in the quality of their essays, a more structured approach to expressing their thoughts, and an increase in creativity. At the same time, the groups that did not use neural networks experienced difficulties in organizing the material and formulating ideas. These differences highlight the impact of technology on the writing process and creativity. The study confirms that the use of neural networks can significantly improve writing efficiency and improve learning outcomes.

• 83.3% of its participants were unable to cite the essay they had just written (in the other groups, this was much lower - 11.1% each).
• None of the participants with access to AI were able to summarize the content of their essay (in the other groups, a total of five out of 36 people had problems with this).
• Three out of 18 participants in this group reported not feeling like the authors of the essay at all. There were no such responses in the other groups.

Differences in citation difficulties did not persist throughout the experiment. For example, in the second session, the group using neural networks practically did not encounter this problem. This may have been due to the fact that by the second session, participants were already aware of the questions they would be asked in the interview after writing their essays. In the first session, the request to cite a quote came as a surprise to everyone. Reciting the main idea of ​​the text was difficult, especially for those using artificial intelligence to write the essays. However, in the third session, most participants successfully completed this task, having prepared for the relevant questions in advance. Preparation played a key role in their success.

Interview results after the fourth session proved more troubling. In this session, participants switched places, and only three of the nine who had previously used neural networks were able to recall the suggested topics. Most of them did not recall having already written essays on these issues. Meanwhile, participants who had no access to the internet or the chatbot in the first three sessions successfully recognized all the topics. These findings highlight the importance of technology's impact on memory and information retrieval.

The study found that among the nine former users of neural networks, only seven were able to recall and cite their new essays written without the assistance of AI. Meanwhile, among participants in the other group who had not used neural networks, only one person experienced similar difficulties. Analysis of the main idea also confirmed this trend: only one of those who had previously relied on AI was able to successfully summarize the main ideas of his essay, while seven participants in the group with no experience using neural networks completed this task. These results highlight the impact of neural networks on self-expression and critical thinking.

All participants in the session noted that the achieved result was more satisfactory compared to the previous experiment. Users interacting with artificial intelligence for the first time highly rated its ability to help structure their thoughts. At the same time, participants who wrote essays without assistance noticed that they invested more effort and creativity in creating their works. This underscores the importance of technology support in the creative and learning process, and also reveals the value of independent work for skill development.

What essays did students produce under different conditions?

The experimenters did not set strict rules for participants regarding the completion of tasks. Participants with access to artificial intelligence could use it to develop the structure of the essay and select information, and then write the text independently. They also had the option of completely entrusting the text generation to AI. In turn, participants who used the internet could simply copy and paste ready-made materials on the topic without editing. This created a variety of approaches to completing the tasks and demonstrated the influence of technology on the writing process.

The analysis shows that none of the participants submitted completely copied essays. However, the largest number of requests to the chatbot, amounting to 38% of the total, came from those who had access to it and concerned text generation. This suggests that many users preferred to use a simple method to avoid the effort of writing themselves.

In the fourth session, when participants who had previously written essays on their own received access to GPT-4o, an interesting trend in requests was observed. The most common requests were for information search, which accounted for 33%. Requests asking "write an essay" accounted for 21%. These statistics highlight the growing interest in using artificial intelligence for researching topics and finding sources of information, and not just for text generation. This suggests that users seek a deeper understanding of the topic, rather than simply obtaining ready-made solutions.

Text analysis revealed that the unaided group produced more diverse essays. Their works drew primarily on personal experience and, despite using similar keywords, addressed diverse issues. Essays by authors who used the internet or artificial intelligence showed noticeable search trends in the selection of subtopics and examples. For example, participants who used internet searches often focused on homelessness in texts dedicated to charity. This writing approach may indicate the influence of popular queries on the theme and content of the works.

The instructors who graded the essays noted the presence of clearly generated works. These texts were well-structured and highly literate, but suffered from poor and formulaic content. Overall, the essays were satisfactory and received high marks for many criteria, but the instructors considered them to be insufficiently thoughtful and unique. This highlights the importance of originality and depth of thought in academic writing.

How the Brain Worked According to EEG Data

Brain activity patterns varied significantly across participants in the different groups. Those who wrote independently demonstrated more active brain activity, with neural networks in multiple areas simultaneously and consistently activated. Compared to participants using an AI assistant, they showed greater activation of areas responsible for working memory, planning, self-control, and concentration, as well as for retrieving necessary words from long-term memory. This emphasizes the importance of independent task completion for deeper engagement of cognitive processes and development of nonlinear thinking.

Research shows that participants who wrote essays independently demonstrated greater activity in the areas of the brain responsible for memory and recall, compared to those who used online resources to obtain information. Using the Internet reduces the need for active use of one's own memory, which leads to less active work of the corresponding neural networks. This suggests that access to information in the online space can negatively affect the cognitive processes associated with memorizing and analyzing information.

Users of neural networks showed a decrease in the activity of neural networks in the brain compared to participants in other groups. Research has shown that these differences persist even in the fourth session, when participants who had previously used neural networks during the first three sessions write essays independently. This indicates the possible influence of neural networks on cognitive processes and brain activity.

The authors note that the brains of these participants did not use all available resources to analyze information and generate text. This may be due to the fact that they are accustomed to relying on artificial intelligence in these areas. Thus, decreased cognitive activity could lead to less efficient perception and processing of information.

What conclusions did the study's authors draw?

The study's authors conclude that using an AI assistant significantly speeds up the process of writing essays and completing other tasks, allowing students to achieve satisfactory results with minimal effort. Thus, generative neural networks significantly increase productivity in various areas of mental work, which explains their growing popularity as a work tool.

Researchers express concern about the use of AI assistants for educational purposes. Essay preparation using artificial intelligence does not provide the same educational outcomes as independent writing. Students who use neural networks do not immerse themselves in the process as deeply as when writing a text independently. They do not analyze the information they find and may not fully understand the arguments they include in their work. As a result, interaction with the material occurs at a superficial level, and the knowledge is not assimilated as their own. This highlights the importance of active engagement in the learning process to achieve deep understanding and develop critical thinking.

Please note the following materials:

The development of artificial intelligence (AI) poses both a threat and an opportunity for educators around the world. On the one hand, AI may threaten traditional teaching methods and the role of teachers by automating processes and reducing the demand for human labor. This raises concerns about job losses and changes in educational standards.

On the other hand, AI opens up new horizons for improving the educational process. It can personalize learning, adapting to the needs of each student and providing access to resources that were previously unavailable. Teachers can use AI to analyze student performance, improve curricula, and create a more interactive educational environment.

Thus, teachers are faced with the need to adapt to the changes brought by AI. They must embrace new technologies and integrate them into their practices to ensure high-quality education in the future. It is important to find a balance between the use of technology and maintaining human interaction in the educational process to maximize the benefits of AI and minimize its potential risks.

The facts speak for themselves: as the labor required to write essays decreases, so does student engagement in the learning process. However, it's worth considering whether this is a positive or negative aspect. Perhaps the purpose of an essay assignment is not only to ensure that students memorize all the facts and concepts on a topic, but also to develop the ability to quickly and clearly articulate their thoughts in writing. Teachers can develop a deep understanding of the subject through other learning formats where the use of AI assistants is not yet relevant. It is important that the educational process remains multifaceted, combining traditional methods and modern technologies, allowing students to develop both analytical and creative skills. Modern education is focused on the active use of artificial intelligence for routine tasks such as big data analysis, information retrieval, and content generation. At the same time, students can focus on developing important communication, self-presentation, and critical thinking skills in small groups under the guidance of teachers. This approach to learning promotes a deeper understanding of the material and the development of competencies needed in a rapidly changing world. The combination of technology and traditional teaching methods opens up new horizons for education, enabling the efficient use of time and resources.

The authors of this study question the possibility of separating cognitive processes. As a preliminary conclusion, they presented the cognitive debt hypothesis. This hypothesis posits that constant reliance on external tools hinders the active use of one's own mental resources necessary for independent thinking. If a person gets used to repeating cues without critically analyzing them, this leads to a superficial understanding of the topic and the formation of biased opinions. As a result, they lose the ability to independently analyze problems and interest in this process.

The authors of the study emphasize that the concept of cognitive debt is preliminary and requires additional data for confirmation. It is important to note that the study's findings are based on a limited sample of participants; there were only 18 people in each group. This limitation may influence the results due to the individual characteristics of the participants. Thus, to more accurately understand and assess cognitive debt, larger and more diverse studies are needed.

Participants who interacted with chatbots may have had a higher proportion of people with unstable memories, which may have hindered their ability to discuss the content of their essays. While not all of the experiment's results can be explained by random factors, their influence cannot be completely ignored.

Who else thinks it's dangerous to delegate intellectual tasks to AI assistants?

Alongside the publication of the MIT authors' article, another preprint on the impact of artificial intelligence tools on the brain and memory was released on the ArXiv platform. This paper is part of an upcoming book on AI, the lead author of which is linguist and neuroscience popularizer Barbara Oakley. Co-authors include Terrence Sejnowski, a specialist in computational neuroscience and artificial neural networks. He is known for his involvement in the development of Boltzmann machine algorithms, a seminal work in the field of machine learning that was awarded half of the 2024 Nobel Prize in Physics. This work highlights the importance of studying the impact of AI on human cognitive function and memory, which is relevant in the context of rapidly advancing technology.

Barbara Oakley and Terry Sejnowski, renowned for their research in learning methods, previously co-authored a book based on their successful Coursera course. In their new publication, the authors raise the important question of the need for changes in the educational system in light of the rapid development of artificial intelligence. They analyze how modern technologies can influence approaches to teaching and training specialists, as well as what new skills will be in demand in the future. Considering these aspects can help educational institutions adapt to changes and ensure high-quality learning in the AI ​​era.

Unlike Natalia Kosmina and other MIT researchers, Barbara Oakley, Terry Sejnowski, and their colleagues from the US, New Zealand, and Taiwan are not limited to analyzing a single experiment. They seek broader generalizations. The main idea of ​​their work is not that using AI to delegate learning tasks hinders students' intelligence. Instead, the researchers emphasize the importance of a balance between technology and traditional teaching methods, emphasizing how innovative approaches can complement, rather than replace, students' personal efforts in the knowledge acquisition process.

The authors delve into the debate, challenging the widespread belief that the primary focus of learning should be on the ability to learn, not specific knowledge. They argue that the dichotomy between knowledge and skills is misguided. They highlight the outsourcing of cognitive tasks to external tools, such as search engines, calculators, and automatic spell checkers, as a problem. They view the emergence of artificial intelligence as the most striking and extreme manifestation of this trend.

The problem is that knowing how to find information and possessing the information itself are completely different levels of understanding. As the researchers emphasize, when a person with knowledge of photosynthesis answers a question on the topic, neural connections in their brain are activated with much greater intensity than in someone who simply knows which neural network to access for information on photosynthesis. This emphasizes the importance of a deep understanding of the topic, rather than a superficial knowledge of how to obtain information.

Both participants will likely answer a simple question (provided that the second has access to the chatbot), so the difference may not be apparent on the surface. However, without one's own understanding of the topic, it is impossible to identify errors that artificial intelligence may make. The second person will not develop a complete and deep cognitive schema of knowledge on the issue in question. This is precisely the essence of education—developing the ability to analyze, critically evaluate information, and form one's own opinion.

Reworked text:

Read also:

In the modern era of the Internet and artificial intelligence, developing subject knowledge in students remains an important task. Despite the availability of information, critical thinking and a deep understanding of a subject are essential for academic and professional success. Subject knowledge forms the basis for analyzing, synthesizing, and applying information, which is especially important in an environment of constant change and innovation in various fields.

The internet provides instant access to a vast amount of data, but without a deep understanding of the topic, students may face difficulty filtering and interpreting information. Developing subject knowledge fosters self-confidence and skills, enabling students not only to acquire information but also to apply it in practice.

Furthermore, subject knowledge develops problem-solving skills that are relevant in any field. It is important to remember that AI assistants can perform a variety of tasks, but true understanding of a subject and the ability to apply knowledge to new situations remain uniquely human qualities. Thus, developing students' subject-matter knowledge is key to successfully adapting to a rapidly changing world.

The authors emphasize that the richness of cognitive schemas, including details and factual knowledge acquired automatically, plays a key role in their depth and functionality. A cognitive schema containing only a general understanding of a subject has limited practical value. To effectively solve new problems using analogies with familiar ones, a cognitive schema must include a significant amount of factual information. This not only allows for a better understanding of the subject but also for adapting to different situations, which contributes to the development of skills and abilities.

If information is stored exclusively externally, for example, on the internet or in neural networks, then the person understands where exactly it can be found and what exactly to look for. However, details remain inaccessible, and when solving complex problems, familiar elements of information must be perceived as new. It's important to have basic facts at hand, as they may be absent from memory.

The process can be compared to trying to assemble a puzzle, comparing each piece to the image on the box, instead of focusing on how the pieces fit together on the table. This approach emphasizes the importance of working carefully with each piece, which can significantly impact the final result. Comparing pieces to a template helps to better understand the big picture, but it is also important to consider the interactions between the pieces to successfully complete the task.

According to the authors, the same applies to skill learning. For example, the ability to do mental arithmetic has become less important since the advent of calculators. However, if a student needs a calculator for every simple arithmetic operation, it becomes difficult for them to master more complex problems. They are forced to think about steps that could otherwise be performed automatically. This can hinder the development of mathematical skills and reduce their confidence in solving problems.

If the multiplication table is memorized, the process of reducing fractions becomes simple and intuitive. Relationships between multiples are perceived naturally, and finding them does not require much effort. Therefore, it is impossible to develop more complex mathematical skills without a preliminary stage of memorizing the multiplication tables. This emphasizes the importance of foundations in mathematics education.

The authors argue that without the creation of deep cognitive schemas, mental abilities cannot develop. They believe that the modern educational system has experienced the negative consequences of an approach focused on the ability to learn instead of transmitting knowledge. They attribute this opinion to the disappearance of the Flynn effect in developed countries. Since the introduction of IQ tests, the intelligence level of new generations has regularly increased, but in recent decades this trend has stopped.

What to do with all this in education?

Researchers from MIT presented their findings regarding the educational process in the field of teaching writing. They recommend using hybrid strategies in which AI assistants will be available to students selectively, not at every stage and not for all tasks. These approaches can foster a deeper understanding of the writing process, improving students' skills and developing their creative thinking. Integrating AI into instruction allows for a balance between technological support and independent student work, ultimately leading to more effective learning and the development of critical thinking.

At the initial stages of instruction, it's best not to use artificial intelligence. This allows the student to fully immerse themselves in the task, which promotes the development of their neural networks and necessary skills. Once the main task has been mastered, AI can be used selectively. However, students should be responsible for generating ideas, structuring the text, and final revision. This will ensure a deeper understanding of the material and strengthen their abilities.

Barbara Oakley, Terry Sejnowski, and their colleagues recommend emphasizing traditional methods in the educational process, such as memorizing key knowledge until it becomes automatic and developing skills without external assistance. They emphasize the importance of limiting the use of neural networks to generate answers and draw attention to the difference between information about where facts can be found and actual knowledge of those facts. This approach helps students gain a deeper understanding of the material and develop critical thinking.

Recommendations of the authors' groups require Urgent and serious decisions are needed from educators and educational systems. The ideas of Natalia Kosmina and her colleagues from MIT appear more practical and feasible.

Regulation of artificial intelligence in education is in its early stages, but some countries, such as China, are already taking steps to limit the age at which students can access generative neural networks. As legislation evolves, it is expected that regulations will be introduced requiring students to demonstrate a certain level of skill development in order to access this technology. This could facilitate a more responsible and effective integration of AI into the educational process, ensuring safety and compliance with educational standards.

The ideas of Barbara Oakley and Terrence Sejnowski can be supported by such regulations, but they aim for a more complex task: reconsidering the division between knowledge and skills, and restoring a positive attitude toward memorization and routine practice in the educational system. Many professionals will likely support these changes, as they could significantly improve the quality of education and the training of specialists.

How can schools and universities convince students and their parents of the importance of memorizing dates, definitions, and formulas when, in everyday life, they can simply use a chatbot to obtain the necessary information? How can traditional educational processes, which require time to master basic knowledge, be reconciled with the expectations of employers who need ready-made specialists with artificial intelligence skills? And how can the level of assimilation of factual knowledge and routine skills be assessed in an environment where generative AI is becoming an accessible tool?

A return to traditional teaching methods is unlikely. However, the work of Barbara Oakley and Terrence Sejnowski provides valuable insights on how to implement effective neuroscience-based techniques into modern education. Their research highlights the importance of using evidence-based approaches to improve the learning process and enhance learning.

Revised text:

Also explore:

• How can teachers adapt assignments so that students complete them themselves, rather than entrust them to neural networks?
• Does ChatGPT help students learn better or hinder them? What do the studies say?
• What should students learn for life in an AI world: an American professor's opinion?
• Can AI replace a university teacher? Interview about the results of a series of experiments