Artificial intelligence in higher education

The use of AI by university students, as with high school students, is quasi-universal, at 97.9%. No significant variation is observed by cycle of study, gender or financial situation; by field, adoption remains above 93% in all fields with at least 50 respondents. Vernacular, bottom-up adoption is thus built on a backdrop of the absence of a coherent policy at the level of the system. In terms of intensity of use, it sits below the level found among high school students. Of those who reported having used AI, 21.7% report daily or almost daily use, 32.2% several times a week for academic activities and 14.6% about once a week, which means that 68.5% turn to it at least weekly, compared with 84% among high school students. At the other end, 11.9% use it one to three times a month, 9.4% less often, 6.4% have tried it only once or twice, and 3.9% have never used it for academic purposes.

One possible explanation is that university students perform academic tasks scheduled at specific points in the academic calendar, the main examination sessions being those in the middle and at the end of the semester, whereas for high school students homework is a daily reality. At the same time, only 3.9% of university students report that, although they have used AI, they have not used the technology for educational purposes, another indication that the use of AI carries over into every segment of the respondents’ lives and activity, including (and perhaps especially) the academic sphere.

Intensity does not differ significantly by level of study: 68% at least weekly use at bachelor’s level and 69% at master’s level; doctoral studies have too few respondents for a separate estimate. The variation by field of study is, by contrast, substantial: from 81.8% in mathematics, computer science and information technology and 74.8% in engineering, down to 48.7% in letters and the humanities and 35.3% in arts and architecture (small numbers). The potential explanation is that, while AI has an immediate and far more natural applicability in the exact sciences, the results produced by the AI system being far more pertinent as well, its use in the social sciences and humanities is more nuanced and does not always produce beneficial contributions9.

The landscape of applications is concentrated, but more diversified than among high school students, with 74.1% using ChatGPT, 64.9% Google Gemini, 30.2% Claude and 29.1% the artificial intelligence functions built into Google search. Use does not stop at the university gate, however. Outside academic uses, 65.5% of university students search for information or learn something new through AI, 58.4% ask for practical advice or support with planning, 28.6% seek to understand current affairs, and 22.3% have used these applications to talk about their emotions or personal problems.

The activities for which university students turn to AI can be grouped into three types. The first concerns access to information and its comprehension: searching for information or ideas reaches 70.6% among respondents, explaining difficult concepts 67.5% and preparing for tests or exams 58.8%. The second concerns interventions on existing texts, from summarising them, at 48.4%, to correcting one’s own text, at 46.8%, and translation, at 31.5%. The third, which is also the least frequent in respondents’ answers, is substitution of the task: doing homework (23.7%) and the first draft of an essay or paper (22.7%). It should be noted that, from the list of possible activities suggested, respondents selected 4.2 options on average, which indicates that these types of activity are not mutually exclusive.

If the pattern holds, the public perception of the university student who turns to AI to replace effort and work is not representative. For this reason, a public policy whose central concern is to identify and penalise the substitution of academic work through AI ignores almost two thirds (64.5%) of the university students who use these tools not to replace work, but to make it more efficient or to enhance it. The rates of use are, moreover, similar to those in the British study presented in the chapter on international models of good practice; there, 95% of students used AI for assessed work, but only 12% took text produced by AI and inserted it directly into their academic work.

Particular attention should be paid to the contrast between the school population and the university population in the use of AI for substituting the academic task. While half of high school students turn to AI for this purpose, the rate among university students falls below a quarter, to 23.7% for simpler tasks and 22.7% for more complex tasks, such as a paper. A pertinent explanation has to do with the fact that tasks, as explained above, are rarer and scheduled in a more segmented way in the academic calendar. Universities also generally have tools for detecting plagiarism and the use of AI, which makes university students more cautious about substituting their own ideas and contributions. Last but not least, many academics do not assess university students on the basis of written tasks or tasks prepared at home, but rather through practical tests or examinations taken in the classroom.

The same logic of contrast is also found, however, in the perception of what is acceptable. On a four-point scale, using AI to search for information about a new topic is considered acceptable by 75% of university students, and solving a problem in a subject by 44.3%. By contrast, writing an essay submitted as one’s own remains acceptable for 16.6%, and doing an entire project in the student’s place for 13.1%. The moral threshold is therefore significantly stricter than among high school students, where the same last two thresholds stood at 40.3% and 34.9%. One explanation may lie in the university experience, in which the student generally takes a course in Ethics and academic integrity, and ethical norms are instilled far more rigorously.

The actual way AI is used confirms this pattern of orientation. On a scale of 1 to 5, 61.4% of university students report that they frequently use the technology to understand the subject matter or to learn on their own, and 41.6% to check whether they have solved something correctly, while only 15.3% take the answer almost as is and 20.2% use it to write a first draft that they then complete themselves. In the same vein, 77.3% state that the technology saves them time and 62.6% that it helps them understand the subject matter better, but only 23.8% admit that they rely on it too much.

The differences by field of study are, however, considerable and highlight an atypical case. Among university students in economics and business, the share of those who have used AI to write the first draft of an essay reaches 40.8%, a rate almost double that of the rest of the sample (20.9%), and failure to check the information received reaches 51.6%, compared with 31% in the other fields, the highest level of all fields. Letters and the humanities do not differ significantly from the rest of the sample (19.7% and 24% respectively), despite the fact that writing text is the dominant academic activity in this field.

Asked how many out of ten classmates have at least once submitted a homework assignment done entirely by artificial intelligence, university students answer 7.2 on average, with a median of 8, and 33.3% of respondents choose the maximum value of the scale. For an essay or paper passed off as one’s own, the average is 6.7, with a median of 7 and 28.4% pointing to all ten classmates. Set against the 23.7 percent that respondents report about themselves, these values produce a gap of about 48 percentage points, substantially larger than the 32 points recorded among high school students.

The first explanation is underreporting. Respondents tend to profile their own academic conduct within the institutionally acceptable limits, but are less restrained when asked to assess the academic behaviour of their classmates, a hypothesis supported by the fact that 39% of university students say that many classmates hide their use. The second possible explanation is the misperception of the norm, namely the tendency to extrapolate to the whole group the behaviours that are visible and memorable, or patterns of conduct widespread in public discourse about students.

The data also allow a third mechanism to be tested, however: projection. University students who report having used the technology to do their assignments estimate higher values for their classmates, 8.4 out of 10, compared with 7 for those who have not, which is compatible with a mechanism of justification by reference to the group. The difference nevertheless remains modest, and the overestimation remains massive even among university students who have never resorted to substitution, who attribute the behaviour to seven out of ten classmates. Moreover, agreement with the statement that many classmates hide their use is higher precisely among those who have not substituted, 3.12 compared with 2.76 on a five-point scale, which shows that the perception is not shaped by one’s own conduct alone and strengthens the explanation based on the misperception of the norm.

The normative context explains why use remains, to a large extent, undeclared. 76.6% of university students consider that using AI for study is frequent and a matter of normality among classmates, and 62.1% say they talk openly about it, but 49.6% say that university students who use AI excessively are labelled cheaters. The coexistence of normalisation and stigma produces exactly the pattern observed.

73.7% of university students consider themselves able to explain to a classmate what AI is: 45.9% consider that they could explain it well, and 27.8% very well. At the same time, 23.2% think they could manage such a task moderately (“so-so”) and only 3% say they could explain it poorly or not at all. In parallel, in the battery of questions dedicated to literacy, university students answer 5.3 out of 6 questions correctly on average; 59.2% of university students answer all of them correctly, 25.7% five questions and only 5.4% at most half of the items. By contrast, high school students obtain an average score of 4.2, and the share of those who answer at most half of the questions correctly reaches 27.8%. An important point deserves to be emphasised: university students are, to a large extent, the best-prepared graduates of the high school cycle, and their familiarisation with AI in an academic context is a far more fluid and frequent process than in pre-university education, which accounts for such a performance.

University students’ answers also indicate a good understanding of the limits of the technology: 96.5% know that these systems can give wrong answers that look correct, 94% understand that a chatbot cannot have real feelings, 93.7% are aware that the technology can generate fake images or videos of real people, and 87.8% know that the applications can invent non-existent sources, quotations or statistics. Likewise, 84% of university students know that generative AI answers are not checked by humans, and 78.2% understand that the same question does not always produce the same answer.

It should also be noted that the relationship between the student’s declared confidence in their own competence and their performance in the section dedicated to the level of literacy is relatively flat and compressed: the score ranges from 4.87 among university students who declare themselves only slightly or not at all competent (small numbers) to 5.53 among those who consider themselves very competent. In this sense, the problem is not the one found among high school students, namely a systematic overestimation of one’s own competence, but the fact that the high level of knowledge and awareness of the basic notions of how AI technologies work does not translate into practice.

The behavioural test confirms this dissociation. Although 87.8% of university students know that these applications can invent sources and quotations, 32.9% admit that they use the information received without checking it, and 13.3% report having actually submitted a paper containing invented information or sources. Moreover, university students who know that the applications can invent sources do not check more often than the others (33.2% versus 30.7% not checking), and those who answer all six questions correctly are no different from the rest (33.2% versus 32.4%). The distance between knowledge and its actual application in academic work reflects the operational definition of the deficit of evaluative judgement described in the diagnosis chapter, and at this level it constitutes the central problem.

The differences by field confirm the same pattern. The literacy score varies relatively little in absolute terms, between 4.96 in education sciences and 5.66 in arts and architecture (small numbers), although the differences are statistically significant, whereas the practice of verification varies far more, from 19.3% not checking in arts and architecture to 51.6% in economics and business. Knowledge is therefore distributed relatively evenly; working discipline is not.

In this subchapter, the hypothesis of regulatory fragmentation in the Romanian university space is confirmed. Only 34.9% of university students say that their institution has clear rules on the use of artificial intelligence, while 34.2% describe partial or unclear rules, 16.4% say there is no rule at all, 4.8% that use is banned, and 9.7% do not know. Although the situation is better than in pre-university education, where the share of clear rules stopped at 29.3%, almost two thirds of university students continue to operate without an explicit framework of rules. It is essential to emphasise that, while most university students in Romania pursue their studies at universities that have already adopted norms of their own, generally through decisions of the university senates, the diffusion of these norms into the classrooms of the various faculties and departments can itself be a fragmented process.

The training deficit is, however, a point of similarity between higher and pre-university education. Asked in which contexts they learned about AI, 27.6% of university students say in none, and the most frequent source of learning they report is the circle of family, friends and classmates, at 29%, followed by online courses taken on their own initiative, at 24%. The latter answer, which concerns individual training, is also a good foundation for possible educational policy interventions. Institutional forms remain in the minority, with 22.1% saying they received direct guidance from an academic, 19.3% having encountered the subject as part of another course, 18.7% in a dedicated course and only 7.3% in a specialised elective. In practice, 42.8% of university students have gone through training content and experiences concerning AI over the course of their academic path.

Another aspect, at least as important for the thesis of this study, concerns the regulatory framework the institutional environment opts for. Where university students indicate the existence of clear rules, 19.7% report having used the technology to do their homework, compared with 25.5% where the rules are partial or unclear and 27.8% where there are no rules. The difference is more pronounced with regard to verification: failing to check information drops from 36.4% (partial or unclear rules) and 36.1% (no rules) to 26% where the rules are clearly defined, a difference of about ten percentage points, statistically significant and stable after controlling for cycle of study, field, gender, age, financial situation and frequency of use. The perceived acceptability of an essay submitted as one’s own drops, in turn, from 20.5% to 13% where the rules are clear. The difference concerning essay writing (17.2% versus 22.6%) goes in the same direction, but is statistically more fragile. University students who do not know whether rules exist behave like those with no rules (27% for homework, 36.1% not checking): a rule that is not communicated does not produce the effects of a clear rule.

At least as important is another finding: a total ban does not produce better results than a coherent framework of rules. In the institutions where university students report that use is banned, doing homework with AI remains at the level of clear rules (20.4% versus 19.7%), while assisted essay writing (31%) and failing to check information (41.5%) reach the highest levels of all categories. No indicator shows a reduction in poor use under a ban. The size of this group is small, however (4.8% of university students), so the differences from clear rules are not statistically robust. The data therefore do not support a policy founded on restriction.

Two results deserve particular attention. The first concerns the intersection between the clarity of rules and the ratio between use as support and use as a substitute. The second concerns the share of those wrongly accused, a proportion which, to the extent that automated detection tools spread without regulation, risks intensifying. The reference studies from the United Kingdom show, moreover, that anxiety about false accusations arises even among students who have not used AI at all, which makes this indicator a measure of the institutional climate, not just of conduct10.

The data confirm this concern. Only 33.5% of university students report having had no negative experience with the use of AI in an academic context, and 21.9% say they have been accused of using artificial intelligence although they had not turned to it, a percentage that far exceeds the share of those actually sanctioned, at 9.7%. To these are added 24.4% of respondents who describe excessive dependence, 17.8% who consider that they understood the subject matter less, 13.3% who have submitted papers with invented sources, and 9.3% who received a lower grade as a result of using AI.

A third result deserves separate attention: the degree of compliance with declarations of use, where the institution requires them. International research shows a significant rate of non-compliance, which suggests that the declaration functions as an instrument for forming a norm, not as a verification mechanism11. Asked what would help them most, university students do not ask for rules first and foremost, but for access to subscriptions to AI tools with advanced features and for practical support. Access to paid versions is mentioned by 31.5% of respondents, the usefulness of a tool approved and provided by the institution is invoked by 30.3% of them, concrete examples for their own disciplines by 28.9%, better-trained academics by 26%, and clear rules by just 22.6%. Only 11.1% consider that they need no form of support.

Both in the literature and in the norms proposed by international organisations and European institutions, equity is treated as a pillar of the discussion, a fact also confirmed by the EU Council conclusions of May 2026, which explicitly call on member states to address the issue of unequal access to digital resources and technological tools12. According to the data available through this study, three distinct comparisons can be drawn, concerning the following aspects: access to paid versions, which separates users with access to top-tier capabilities from those using free versions; differences by area of residence and by declared financial situation; and access to guidance, which is a form of institutional capital distributed just as unequally as access to tools.

As regards access to paid versions, the result is more favourable than in pre-university education. 66% of university students use only free versions, while 9.2% turn to a version they pay for themselves, 12.9% use both free and paid versions, and 11.7% a version paid for by the educational institution. By contrast, only 0.5% of respondents in the study dedicated to pre-university education reported access to a tool paid for by the school, which demonstrates a more fluid mobilisation among universities.

The second test nevertheless points to a fragmented landscape of access. Access to a paid version from one’s own resources decreases moderately with the family’s financial difficulty, from 26% in households that live comfortably to 18.5% in those that find it hard to get by, whereas access provided by the institution does not vary significantly with financial situation (between 11.2% and 12.3% in the first three categories; the 17.3% in the group that finds it hard to get by is based on only 74 respondents), and the share of those who have not learned about this technology in any context remains practically constant, between 25.8% and 30.2%. Unlike pre-university education, where socio-economic status influenced the level of familiarity with AI more visibly, in universities it is not associated with access to training, but the data do not show that institutional mechanisms compensate for the initial disadvantage.

A third gap appears, however, between fields of study. University students in economics and business and those in fields such as mathematics and computer science use AI almost as intensely, 73.8% versus 81.8% weekly (a non-significant difference), but resort to assisted essay writing at 40.8% versus 20.1% and fail to check the information received at 51.6% versus 37.7%, against a lower literacy score, 5.20 versus 5.58. Of course, in the specialisations in mathematics and computer science, essay writing is a rarer task; still, disciplinary cultures appear to create an important factor of discrepancy in this equation.

University students are moderate in assessing the impact of AI on their own trajectory. On a scale of 1 to 5, the conviction that these skills will be needed in their future profession reaches the highest average, 3.56, followed closely by the concern that AI will reduce employment opportunities in their field of interest, 3.55, and the fear that part of what is being learned now will become useless, 3.44. Concern about dependence stands at 3.24, the feeling of being prepared at 2.89, and general optimism at 2.78. University students do not reject the technology, do not underestimate it and consider it inevitable, but they declare themselves modestly prepared to manage a future in which its advance intensifies. Asked whether or not they agree with various statements, 59.4% admit that developing AI skills is necessary for them in their future profession, and 57% express anxiety about employment opportunities. Only 37.1% consider themselves prepared for the new transformations and only 32.2% look with optimism at the role of this technology in education and in their own future.

The modest confidence in their own training is coupled with a distrust, rather, of the training they undergo in the university ecosystem. University students’ concern has, moreover, a counterpart in the labour market data.

The World Economic Forum report on the future of jobs places skills related to artificial intelligence and big data processing first among the skills with the fastest-growing demand, and 63% of employers identify the skills shortage as the main obstacle to the transformation of their own activity; estimates derived from recruitment data indicate an increase of about 70% in demand for literacy in the field between 2024 and 2025, while only two in ten business leaders consider that education systems have kept pace13. The OECD analysis of skills in the age of AI nuances this picture, however: in occupations with high exposure to automation, most job advertisements require complementary, management, process and cognitive skills, not technical programming skills14.

Anxiety is already producing measurable behavioural consequences. 51.1% of university students report that they are considering focusing on skills that are hard for technology to replace, such as creativity, communication and critical thinking, 30.4% that they are turning towards a field considered more resistant to automation, and 28.9% that they intend to explicitly learn how to use these tools. Worryingly, 9.8% are reconsidering whether to continue higher education, 11.5% are considering changing their field of study, and 11.1% postponing important decisions. Only 7.9% have discussed these decisions with their family, with an academic or with a counsellor.

Finally, the questionnaire also measured the personal and emotional dimension of use. 22.3% of university students have used these applications to talk about their emotions or personal problems, and on the attitude scale 15.3% feel understood when they talk to a chatbot, 16.9% consider that it is easier for them to talk this way than to a person, and 15.7% that it helps them when they are anxious. The majority nevertheless remains clearly anchored in human relationships, since 74.8% prefer to talk to someone close and only 6.6% consider that the application knows them as well as their friends or family. University students use AI for personal problems as often as high school students (22.3% versus 23%), but their attachment to the chatbot is weaker. Even so, turning to AI for matters concerning mental health is in itself a winding path, with potentially problematic effects.

Legal obligations: the European regulation on artificial intelligence

Empirical evidence on the effect on learning in higher education

The critical literature in studies on higher education

The framework specific to higher education

The starting point: what the cohort entering university brings with it

Higher education: regulatory fragmentation

International data on university students

The applicable national legal framework

Comparative analysis: from high school to higher education, a winding transition

The National Register of Qualifications in Higher Education (RNCIS) currently defines several professional competences dedicated to each specialisation, as well as transversal competences applicable to a broader range of professions. For the latter, a possible territory of action also emerges, through:

Likewise, in the context of the full-scale revision of the European ESG standards now under way, Romania must formulate a pertinent contribution in the field and has the opportunity to place the issue of AI at the epicentre of the process. Such a course of action would also elevate Romania’s standing in the European debate dedicated to higher education.

Because the governance architecture at national level in the field of higher education is atomised owing to the essential criterion of university autonomy, two mechanisms can, as a priority, support the development of common standards dedicated to integrity and ethics in the age of AI, to sound assessment, to the development of specific competences and to the consolidation of partnerships. A first course of action is to turn to the National Council of Rectors, together with the student federations representative at national level, the representative structures of academic and research staff and the representatives of the Ministry of Education and Research, structures which, together, can appoint a task force mandated, for 6 months, to define a national standard in the field.

A second course concerns stimulating the role of national university consortia. One example is the Universitaria Consortium, a consortium that brings together the universities in the country that perform most strongly in the international rankings. In the logic of the British model, a decision by these universities to follow common standards would considerably support their adoption by the universities that are not part of this consortium as well. In addition, universities can decide, in turn, to form consortia in order to deliver double degree programmes or to create incubators or other innovation platforms in the field of artificial intelligence and new technologies.

The Irish model offers the concrete format of this endeavour: a national framework drawn up by the funding authority together with the sector, built on a small set of principles and accompanied by implementation tools, which promotes coherence without imposing a single policy. Transposed to Romania, this means a framework drawn up jointly by the Ministry of Education and Research, ARACIS and the National Council of Rectors, with the participation of the student federations69.

The list below concerns measures that regulation cannot address directly, because they concern the range of specialisations, the student’s individual path and the ecosystem around the higher education institution. These can, however, be built on existing funding instruments or potential new support mechanisms, as well as on institutional infrastructures already in place, without any need for new legislation:

A final note of calibration, formulated by the OECD, makes the point that universities should not train only users of these technologies, but also builders of them, and that preparing future researchers and professionals in various sectors to handle the specialised tools of assisted research is a component distinct from general literacy. For Romania, this means that the measures above must be coupled with a component dedicated to doctoral studies and doctoral schools72.