Artificial Intelligence in Pre-University Education

AI adoption is quasi-universal among students, reaching 98%. This level of adoption does not vary notably between subgroups. In practice, in a context in which there is no defined educational policy framework, the phenomenon we find is that of vernacular adoption, an uncoordinated, bottom-up process of using artificial intelligence. Attention therefore shifts not to the specific groups that use the technology, but rather to the frequency and object of use.

In terms of intensity, use is substantial and regular. Of the students who have ever used artificial intelligence, 25.1% use it daily or almost daily for school, 44.5% several times a week and 14.6% about once a week, which means that 84.2% turn to it at least weekly. At the other end of the distribution, 4.7% use it one to three times a month, 7% less often, 2.7% have tried to interact with an AI tool only once or twice, and 1.4% have never used it for school.

Use does not stop at the school gate. Asked about their use of AI over the last three months, 26.5% of students report they have used artificial intelligence daily in their free time and 35.9% several times a week. The dominant activities that AI facilitates for them in their free time are searching for information or learning something new, at 62.6%, and practical advice or planning a course of action, at 52.6%. This finding is relevant because recent literature signals the risk that the erosion of skills also occurs outside the classroom, through the use of commercial AI tools, where there are no protective barriers5.

The intensity of use is, moreover, remarkably uniform across tracks, from 85.1% at least weekly use in the theoretical track with a sciences profile to 82.6% in the vocational track, differences that fall within the margin of error. The landscape of applications, however, is highly concentrated: of the students who use AI for school, 87.1% use ChatGPT, 51.9% Google Gemini and 28.4% the artificial intelligence functions integrated into Google search, while the other applications remain below 16%. A student uses, on average, 2.4 different applications, which also shows the wide range of commercial options they access in the absence of structured, educational alternatives.

As for the activities for which students turn to AI, these can be grouped into three types. A first type concerns access to information and its comprehension: searching for information or ideas at 74%, explaining difficult concepts at 55% and preparing for tests or exams at 50%. A second refers to interventions on already existing texts, from summarising them, at 47%, to translation, at 44%, and correcting one’s own text, at 39%. A third is the substitution of the task, namely doing the homework, at 50%, and the first draft of an essay or paper, at 22%. It is worth mentioning that, from the list of suggested possible activities, respondents selected on average 4.3 options, which indicates that these types of activities are not mutually exclusive.

Substitution of the task, therefore, is not a marginal behaviour; about 1 in 2 students turn to AI to complete school tasks in their place. It is essential, however, to observe what exactly is being substituted: doing a specific homework assignment is reported by 50% of students, while writing the first draft of a large piece of work, such as an essay or paper, remains at 22%, and the battery on the mode of use shows that taking the answer almost verbatim is frequent for 29.8% of students, while use for understanding the subject matter is frequent for 52%. The picture, therefore, is neither that of a generation that copies systematically, nor that of one that uses the technology exclusively for learning, but that of a mixed practice, in which the same student alternates between the two regimes depending on the stakes of the task.

The same logic is found in the perception of what is acceptable. On a four-point scale, use for searching for information about a new subject is considered acceptable by 85.4% of students, and solving a problem in a subject by 69.4%. By contrast, writing an essay submitted as one’s own remains acceptable for 40.3%, and doing an entire project in the student’s place for 34.9%. Students therefore operate with a clear moral distinction between support and substitution, even if they may breach this distinction in their own activity.

As regards group norms, 73.6% of students consider that using artificial intelligence for school is something normal among classmates, 57.9% state they talk openly about it, but 46.2% say that many classmates hide the fact that they have used it, and 42.6% that whoever uses it too much is seen as a “cheater”. The coexistence of normalisation with stigma explains why use remains, to a large extent, undeclared.

The strongest contrast appears between the use declared by students and the use they attribute to their classmates. Asked how many out of 10 classmates have at least once submitted a homework assignment done entirely by AI, students answer, on average, 8.2. For the similar question about essays or papers, the average is 7.6. For both questions, the most frequent option is 10, that is, the maximum value of the scale, chosen by 53.1% of students for the first question and by 44% for the second. In practice, while 50% of students report they have used AI to do their homework, the perception of classmates reaches 82%, a difference of about 32 percentage points. A first explanation has to do with the way students declare their own use: they may tend to underreport it, in order to fit widely accepted patterns of school behaviour, but they are more open when they talk about their classmates.

The second possible explanation is the misperception of the norm. Respondents may tend, under the influence of perceptions that are also widespread in the public sphere, to overestimate their classmates’ norm of using AI tools to substitute for schoolwork. The survey shows that students who report they have used the technology to do their homework estimate higher values for their classmates, 8.8 out of 10, compared with 7.6 for those who have not, but the difference between the two categories remains modest, and the overestimation persists even among students who have never resorted to substitution, who attribute the behaviour to three quarters of their classmates. In other words, projecting one’s own behaviour onto classmates explains only a small part of the gap: the overestimation of the norm also appears among those who do not use AI for homework. We are therefore talking about a possible coexistence of these two explanations, namely the underreporting of one’s own use, accompanied by an overestimation of use by classmates, mechanisms that cannot be separated methodologically with this instrument, but which, together, account for the distance of 32 percentage points. The consequence for public policy is that the institutional response must be twofold: redesigning the task and the assessment, for the underreporting component, and communicating the real norm in each school, for the overestimation component.

Asked how well they could explain to a classmate what artificial intelligence is, 44% of students answer “well” and 16.7% “very well”, so 60.7% declare themselves capable. At the same time, 34% choose the option “so-so” and only 5.3% admit that they would explain little or not at all. Confidence in one’s own competence manifests itself fairly strongly. The questionnaire, however, allows this self-assessment to be confronted directly with a more objective form of measurement. In practice, students were presented with six statements about how these applications work, with the options “True”, “False” or “I don’t know”. The average score is 4.2 correct answers out of 6, and the distribution shows that 23.6% answer all of them correctly, 27.1% five, and 27.8% at most three, that is, half or fewer.

The errors students make in their knowledge of AI technology are nuanced. For instance, respondents understand the premises of AI well: 84.9% recognise that these systems can give wrong answers that seem correct, 81.6% that a chatbot cannot have real feelings towards them, and 80.2% that the technology can generate fake images or videos of real people. The gaps appear exactly where an understanding of how the system works is needed: only 62.8% know that applications can invent non-existent sources, quotes or statistics, 60.5% that the same question does not always produce the same answer, and 54.5% that the answers are not checked by people before they reach the user. On this last item, 36.2% of students choose the option “I don’t know”. This is also where the structural problem begins that makes AI literacy imperative.

The relationship between declared confidence and measured competence is monotonic, but poorly calibrated. Students who declare themselves very capable score on average 4.75 points, those who declare themselves capable 4.30, and those who answer “so-so” 3.99. Self-assessment therefore has real informative value. The problem, however, is that not even maximum confidence guarantees competence: 15.3% of students who consider themselves very capable of explaining what artificial intelligence is correctly answer at most half of the items.

At the same time, in their school practice, students lack a few essential elements of “hygiene” in the use of artificial intelligence. The pillar of the discussion is the verification of information generated by AI. 45.5% of students who use AI for school state they use information received from these applications without verifying it, and 20.7% report they have actually submitted a piece of work containing invented information or sources. The reflex of verification is thus missing from the student’s school profile. The differences between tracks are significant, although the intensity of use is almost the same: the average literacy score is 4.70 in the theoretical track with a sciences profile, 4.41 in the humanities profile, 4.26 in the vocational track and 3.84 in the technological track. This also gives rise to possible discrepancies that it is important for public policy solutions to address.

The regulatory vacuum and the absence of coherent standards contribute to the gap between students’ use of AI and the objective of a responsible and sustainable approach. For example, only 29.3% of students state their school has clear rules on the use of artificial intelligence, while 29.9% state the rules are unclear or only partial. At the same time, 20.9% of students state there are no regulations, 5.3% state use is completely banned, and 14.6% are not aware of the situation (the corresponding answer being the option “I don’t know”). In other words, almost two thirds of students use AI in the absence of a coherent school framework, and one in seven students does not even know whether any framework exists.

The gap is felt even more sharply in the spectrum of AI literacy and competence building. 27.4% of students report they have not learned about artificial intelligence in any context, and the most frequent source of learning is the circle of family, friends and classmates, at 32%. In educational contexts, only 20.8% have come across information and exercises about artificial intelligence as part of another subject (through curricular infusion), a sign that this practice is rather limited. Only 15.7% have had AI learning experiences in a media education or critical thinking activity, 15.6% in a dedicated class, and 15.5% have received direct guidance from a teacher on responsible use. A dedicated optional course was taken by 7.9% of respondents. It is worth mentioning that most respondents have teachers who have completed the media education programmes of the Center for Independent Journalism and are more inclined to discuss AI with students in class. Even in our survey the percentage of those who have not encountered content on AI in class is high; in reality, however, it could be even more pronounced.

One more contrast underlines the complexity of the issue. Where students report the existence of clear rules, 43.8% report they have used AI to do their homework, compared with 53.3% where there are no rules and 54.7% where the rules are partial or unclear. The same trend appears in the verification of information, where non-verification rates drop from 50.4% in the absence of rules to 42.2% where they are clear, and in the perceived acceptability of substitution (the essay submitted as one’s own), which falls from 45.1% to 37.4%. Although the differences are not spectacular, they are consistent and indicate fertile ground for possible progress.

Even under these conditions, another set of data points to the premise that any educational policy effort must embrace in order to produce concrete and pertinent results: a total ban on use does not produce better effects than clear rules, accompanied by guidance adapted to students. In schools where students report use is banned, 47.5% still resort to homework substitution, almost four points above the level in schools where clear rules are set, and the share of non-verification of AI-produced information remains at 46.2%. The data therefore do not support a policy grounded in restriction, but rather a policy that encourages responsible, competent and well-targeted use.

The use of AI, however, also comes with frequent negative experiences. Only 28.7% of students report that they have not encountered negative experiences related to artificial intelligence at school. The most frequent incident reported is the unfounded accusation of using AI for educational tasks, at 27.2%, while 16.3% have been sanctioned or warned by a teacher for using AI. At the same time, 22.5% of students state they have come to depend too much on artificial intelligence, that is, more than one in five respondents. 20.7% say they have submitted work with invented sources, 18.8% indicate a weaker understanding of the subject matter as a result of using AI, and 18.4% say they received a lower grade than they expected after using AI.

In this way, a considerable background of anxiety is created in the use of AI technology. To respond to this anxiety and to ensure an educational experience that is as beneficial as possible for students, multiple solutions are called for. Asked what possible solutions would help them, students largely ask for guidance and mentoring or other forms of practical, applied support, not strict rules. For instance, 30.1% state they would be helped by concrete examples of using AI in the various subjects they study, 27.9% indicate the need for access to paid versions, with additional and more powerful features, 23.2% suggest that a tool approved and provided by the school would be of real help to them, and 21.9% stress the wish for more coherent training of teachers. Only 19.2% state that clear rules on what is allowed would solve their problems and dilemmas. By contrast, 16.5% consider that they do not need any additional support, and 9.9% show a lack of interest in this topic. The public policy conclusion becomes evident: students do not ask first and foremost for rules, although these are necessary, but very correctly identify the need for forms of specialised support and express it in considerable proportions.

Both the specialist literature and European and international public policy perspectives treat equity as a fundamental pillar in supporting a healthy integration of artificial intelligence into the reality of pre-university education. For example, the Council of the European Union concluded in May 2026 that member states must address structurally the issue of unequal access to digital resources in educational contexts6. In our study, three dimensions concern the question of inequalities in the use of digital/AI tools, namely: access to paid versions, which creates a distinction between users with access to cutting-edge capabilities and those who use exclusively free versions; differences according to the family’s declared financial situation; and access to guidance, which is a form of institutional capital distributed just as unequally as access to tools.

The question of access to subscriptions and, implicitly, to advanced artificial intelligence features is a first level of the discussion. Of the students who use AI for school, 90.2% use exclusively free versions of LLMs, while only 4% have access to a version paid for by themselves or their family, and 4.6% use both free and paid versions. Another statistic, however, is striking for its particularly small proportion: only 0.5% of students use a version paid for by the school. In practice, schools, quasi-universally, do not provide students with access to AI tools dedicated to the educational process, and the explanations on this front are multiple: the administrative question, the budgetary problem and legal considerations. Nevertheless, the idea that the World Bank report also confirms7, namely that the role of the school in ensuring access is practically non-existent, creates an important window for public policy intervention.

Another issue concerns the distribution of access to the various artificial intelligence tools. Among students from families who can afford a comfortable living, 11.8% have access to a paid version, compared with 5.7% in families who are doing well and 4.3% in the families of students who state they have to be careful with spending. Although modest in amplitude, the discrepancy exists, and it can affect school performance, particularly in a context in which commercial LLMs tend to limit the features of their free versions.

Finally, the most striking contrasts are revealed by the issue of literacy in relation to the socio-economic status of the student’s family. The share of students who have not learned about artificial intelligence in any context rises with the family’s degree of financial difficulty, from 23.3% in comfortable households to 28.8% in those who are doing well and 33.3% in those who have to be careful with spending or struggle to get by (the categories “we struggle to get by” and “we struggle a lot” were grouped with this one, being too small for a separate estimate). In other words, a student from a family with limited resources is over 40% more likely to remain completely untrained in a technology they use anyway.

An additional layer that completes this landscape of inequalities consists of the literacy gaps between tracks. Students in the technological track use AI as intensively as those in the theoretical track with a sciences profile, 83.6% compared with 85.1% weekly, resort more often to homework substitution, 53.6% compared with 46%, and obtain a significantly lower literacy score, 3.84 compared with 4.70. The technological track also concentrates students from families with more modest resources: in our sample, only 35.4% of students in the technological track say their family lives comfortably, compared with 55.3% in the theoretical track with a sciences profile. This gap thus amplifies already existing inequalities. We reiterate the working hypothesis of this study: the relevant gap does not manifest itself at the level of adoption, where saturation is almost complete, but at the level of the quality of use and, above all, of access to guidance.

The data show that students display real anxiety about their future prospects. On a scale from 1 to 5, the fear that part of what is learned now will become useless because of AI reaches the highest average, at 3.54, followed by the concern that the technology will reduce employment opportunities in the field of interest, at 3.41, and by the belief that these skills will be needed in their future occupation, at 3.31. Concern about dependence and the loss of certain skills stands at 3.08, the feeling of being prepared at 3.02, and general optimism at 3.01.

Respondents do not reject the technology itself and do not underestimate its potential or relevance, but they perceive its progress as inevitable and are modest in assessing their degree of preparedness for these transformations. The distance between the fear about the usefulness of what they learn, at 3.54, and the feeling of preparedness, at 3.02, may reflect deficient training, not a particular attitude. Put differently, 54.5% of students fear that part of what they learn now will become useless, while only 36.6% feel prepared for a future in which technology is used intensively in the workplace.

This anxiety has tangible consequences. 36.7% of students report they are considering placing an emphasis on skills that are hard for technology to replace, such as creativity, communication and critical thinking, and 32.7% that they are turning towards a field they consider more resistant to automation. A quarter, 25%, intend to learn explicitly how to use these tools. It is worrying, for instance, that 14.4% are reconsidering whether it is still worth pursuing higher education, 12.5% are thinking of changing their field of study, and 13% of postponing important decisions until the “waters” clear. Only 8.6% have discussed these decisions with their family, a teacher or a specialised school counsellor.

Finally, the questionnaire also measured a dimension that flies under the radar of the classic discussion, that of personal use. 23% of students who use AI outside school report they have used these applications to talk about their emotions or personal problems and 12.1% for free conversations with an artificial character or “friend”. On the attitude scale, 21.3% feel understood when they talk to a chatbot, 20.8% consider that it is easier for them to talk this way than with a person, and 18.7% that it helps them when they are anxious. The majority, however, remains clearly anchored in human relationships: 73.4% prefer to talk to someone close when something troubles them, and only 11.7% consider that the application knows them as well as their friends or family. The result does not indicate a substitution of social relationships, but it describes a segment of about one in five students for whom the warnings of the Brookings report on anthropomorphism8 become directly relevant.

Legal obligations: The EU Artificial Intelligence Act

Competence frameworks and guidelines at European and international level

Empirical evidence on the effect on learning

Reference perspectives from the critical literature

PISA tests and AI literacy

The teacher training deficit

The institutional diagnosis

The regulatory vacuum

Where does the absence of policies lead?

The applicable national legal framework

The National Digital Competence Framework for Students was published in 2024 by the Ministry of Education and Research, starting from the DigComp 2.2 model, but supplementing it with a sixth area of competences - the foundational ones. Romania therefore has experience in adapting European educational policy regulations and materials, as well as in supplementing them. For this reason, Romania can take the following steps to adapt the competence framework at national level, and also to raise this issue on the agenda of the European Union:

At the same time, Romania can contribute actively, at European level, to the development of the artificial intelligence component of DigComp 3.0, so that it is accompanied by descriptors and tools directly applicable in pre-university education. In this way, Romania can also position itself on the European stage as a pioneer in advancing an educational framework adapted to current technological developments. By also linking such an effort to the negotiations on the future Multiannual Financial Framework 2028-2034, Romania can also campaign for the implementation of a relevant programme dedicated to education in the field of new technologies at European level. Likewise, through the country’s representative in the College of EU Commissioners, Roxana Mînzatu, Executive Vice-President responsible for social rights and skills, quality jobs and preparedness, Romania can ensure that this cause is well represented at the level of the European Union. The Vice-President’s mandate, moreover, covers exactly the instruments into which such a revision would naturally fit, instruments aimed at accelerating basic literacy across all relevant dimensions, creating a strategic vision for STEM education, as well as other efforts meant to lead to the consolidation of a “Union of Skills”.

In parallel, at curricular level, additional reforms are planned for this school year (2026-2027). Complementing the recent update of the framework curricula, new adaptation efforts dedicated to upper secondary education will take place. Likewise, for schools seeking greater educational autonomy, pilot programmes dedicated to curricular reform will be run. In this context, a new opportunity that is taking shape is the launch of a Plan for the curricular infusion of knowledge, skills and attitudes in the field of artificial intelligence and new technologies (PlanIA). Two courses of action can be pursued here:

Some clear targets for the infusion of AI competences into the current architecture of teaching and learning can be set to support this effort; for instance, integrating tools or information about artificial intelligence into at least 2 class hours in the case of social sciences and humanities subjects and of those in vocational and pedagogical education, and 5 class hours in the case of science subjects and those in technological education. It is also possible to experiment by developing a new specialisation within the Theoretical track - Sciences profile, namely the Mathematics-computer science and new technologies specialisation, a specialisation designed precisely to create generations of new specialists capable of managing developments in the sector from a technical perspective as well.

Recently, within the RECRED project, the National Centre for Curriculum and Assessment (CNCE) of the Ministry of Education and Research developed the “National assessment standards for primary and lower secondary education”. Through a ministerial order adopted in August 2026 (OMEC no. 4615/2026)54, they entered into force from the current school year. This year, the norms for upper secondary education are to be developed in turn. For each subject, the assessment standards were drawn up in the logic of three levels of competence, each with certain descriptors - the basic level, the consolidated level and the advanced level. These standards will apply both to classroom assessment and to standardised national exams such as the intermediate assessments for grades 2, 4 and 6, the National Assessment and the Baccalaureate exam.

The need for a centralised framework regulating the parameters of assessment was substantial; for this reason, this initiative is notable. Nevertheless, the standards, in their current form, do not integrate the various dimensions of the challenge that AI represents for the assessment of students’ competences. For this reason, a revision of the standards, by annexing a dimension on assessment in the context of artificial intelligence, is desirable and even imperative. Because AI has become an inevitable factor in the educational landscape, these courses of action can guide the process of updating the standards:

At the same time, gradually, standardised assessments too, especially the Baccalaureate exam, must take account of developments in AI in order to adapt to the current context. For this reason, a beneficial measure could be to broaden the component of the exam that assesses digital competences by integrating artificial intelligence competences into it. Last but not least, in order to be able to develop these updates to the assessment standards at national level, we recommend setting up, alongside the CNCE, a Special Commission for the assessment and curricular infusion of artificial intelligence, a commission made up of specialists in pedagogy, leaders of the representative organisations of educational actors and representatives of the relevant academic community.

The measures already presented add to the responsibility of teachers. An increase in responsibility, however, cannot come without specialised support. As we showed earlier, teachers already use AI to a large extent, almost 1 in 2 turning to this technology in 2024. The initiatives to increase support for teachers through the “Digital Pedagogy” programme, for instance, started from sound premises; the reported impact, however, is rather limited. The explanation is as simple as it is complex: teachers do not need training alone in order to understand new technologies, given that these are constantly changing. The major need is a broader one, namely permanent support, through tools that help them manage these processes. For this reason, the following measures can create a beneficial framework for action:

Although defined as adjacent support measures, the initiatives presented below constitute equally important pillars in managing the impact of AI on the educational process. A first measure we propose is that the Ministry of Education and Research develop a single national LLM, accessible to all pre-university schools, specially tailored to meet the requirements of the educational process and inspired by the Estonian model. This LLM would be intended for use by students and teachers in teaching and learning, as well as in assessment. In addition, the following measures can represent important steps in accelerating AI literacy:

Likewise, the inclusion component is imperative. All the measures above, taken together, would ensure greater inclusion. Yet, as AI evolves, so do the risks that this technology will deepen the already massive inequalities in the educational process. Multiple measures can be adopted:

Finally, to address the issue of students’ career guidance and their anxiety in this respect, the following measures can be adopted: