Institutional Cases of Generative AI in European Education

This report maps the reports, models and practices that the ETHICAI partners named when asked what shapes the use of generative artificial intelligence at the national, institutional and classroom levels in their countries. Six partner institutions from Greece, Cyprus, Portugal, the Netherlands, Italy and Belgium answered a structured call, and the task lead compiled a seventh answer for Anadolu University Open Education Faculty from documented sources. The seven responses were divided into 70 entries, each tied to the sentence in the response or source on which it rests, and coded in two cycles into 30 codes, six categories and five themes. Nearly half of the entries sit at the institutional level, and the three levels do different work. National texts are strategies, laws and advisory frames; binding rules appear at the institutional level and nowhere below it; and fourteen of the eighteen classroom entries concern the capacities of teachers and students rather than control of use. Disclosure of use and the redesign of assessment form the common vocabulary of institutional rules, although half of that vocabulary comes from the Portuguese response. Tools enter classrooms mainly as tutors and pilots whose implementation the respondents themselves describe as incomplete. The middle level is filled unevenly: by national agencies in one response, by twelve institutional documents in another, by the policies of six universities in a third and by nothing in a fourth. The distribution resembles the one found across the ADMIT reports, which suggests that the relative thinness of the classroom level reflects what institutions document rather than what they do.

The call and the responses

Partners received a form asking for reports, models or practices at three levels, each with a link and a short explanation: the macro level of national policy, the meso level of institutional governance and the micro level of classroom and learning design. Six partner institutions answered between 15 and 25 September 2026. One response arrived twice with identical content and is counted once, and one ends mid-sentence in its institutional section, so only the part received was recorded. The Belgian partner, KU Leuven, sent ten separate submissions on 25 September, each naming one source; they are read together as one response. The answers differ in scope as well as in length: the Greek response concerns school education, the Cypriot response covers both school and higher education, and the others concern higher education. The Belgian response also names institutions other than KU Leuven, one of them TU Delft in the Netherlands; these entries stay with the response that named them.

Anadolu University Open Education Faculty, as task lead, did not answer the call in the same way. Its answer was compiled against the same three questions from sources already verified in the ADMIT synthesis, from the faculty's two published guides and programme page, and from two manuscripts by members of the faculty. The answer is marked as compiled throughout, and its entries rest on those sources; it was entered into the form afterwards so that the record of responses is complete.

Unit of analysis

Each item a respondent listed became one entry. An item was recognised by its own title, link or bullet; where a sentence named two instruments, the two became separate entries, and where several publications were listed under one heading, they formed one entry. Every entry carries at least one verbatim quotation from the response or source, and a script searched each quotation in its source before the report was compiled. All 96 quotations were located. An entry keeps the level at which the respondent placed it, even where another reading would be possible, because the placement itself is part of what the responses show.

Coding and theming

Coding followed two cycles (Saldaña, 2021). In the first cycle, descriptive codes recorded what each entry does and which principles it invokes; in the second, pattern coding grouped the 30 codes into six categories: the form of the instrument, capacity building, integrity and accountability, values and protections, provision of tools and infrastructure, and reported absence. The 199 code assignments were then read across levels and responses, and five themes were built from the patterns that held across more than one response (Braun & Clarke, 2006). Each theme is defined by the codes that carry it, so an entry belongs to a theme when it carries one of the theme's codes and may belong to several; the codes of values and protections cut across the themes and define none of them. A code counts once per entry, so a frequency is the number of entries that carry the code.

Word statistics

The word statistics use the partners' own wording: the text of each form cell for the six partners, and the entry summaries for the compiled Anadolu answer. Function words, links, the names of countries and institutions, and the words present in every answer (such as artificial intelligence and generative) were removed, and established terms such as academic integrity were counted as one unit. Reference lists pasted into a response were left out, since they carry authors' names and titles in other languages rather than the respondent's own account. Keywords for each level were identified by log-likelihood against the other two levels (Dunning, 1993), keeping only words more frequent than expected and above the threshold of 3.84. Co-occurring codes were weighed by the Jaccard coefficient, which prevents frequent codes from appearing to accompany everything (Jaccard, 1912).

Limits

Seven responses cannot represent seven systems, and the numbers in this report describe what the respondents chose to name. Most respondents are open and distance teaching institutions, and campus-based universities enter the set mainly through the Belgian response. Response length varies widely, from seven counted words in one response to more than two hundred and fifty in another, and the longest institutional list comes from a single country. Frequencies are therefore read as the weight of a pattern in these answers, and no test of difference between countries is attempted.

The seven responses yield 70 entries: 18 at the macro level, 34 at the meso level and 18 at the micro level. The responses are unequal in size. The Portuguese response alone contributes 16 entries, thirteen of them at the institutional level, which is more than a third of all meso entries; the Belgian response follows with 12, seven of them institutional; the Dutch response contributes three, two of which report that nothing exists. The Greek response is the only one in which the micro level holds the largest share.

What an entry is also differs by level. Binding instruments appear at two levels only: three laws or ministerial decisions at the macro level and fourteen institutional policies or regulations at the meso level. Advisory texts gather in the same two places, four at the macro level and thirteen at the meso level, and two advisory texts reach the micro level. No entry at the classroom level is a binding rule.

The second coding cycle grouped the codes into six categories. Each category is read below through what the respondents wrote, with entries that show the category at its clearest. The quotations are the respondents' own words and appear as they were submitted.

Form of the instrument

The first category records what kind of thing a respondent named. Strategies, laws and advisory frames make up the national layer, while institutional policies make up the layer below: all eight strategies and all three laws are macro entries, and all fourteen institutional policies are meso entries. Advisory guidance is the second most frequent code in the set and appears at all three levels, mostly at the institutional one. The respondents often qualify the status of an instrument themselves, calling it advisory, binding, under consultation or not yet complete, which shows that the form of an instrument matters to them as much as its content.

Capacity building

Capacity building is the category of the classroom. Professional development appears in fourteen entries, five of them Greek and four Belgian, and AI literacy in eleven, spread across students, staff and the public. The programmes differ in depth, from an accredited certification to a single online seminar, but they share an orientation: the teacher is addressed as a designer of material and activity. Curriculum appears twice, once as a degree built around AI-assisted coding and once as a pedagogical model that writes AI into every programme.

Integrity and accountability

This category collects the mechanisms that make use visible. Assessment design appears in nine entries, academic integrity in six and disclosure in five, and all six integrity codes are institutional. The mechanisms differ in how far they reach into practice: a signed declaration records that use occurred, interaction records show how it occurred, and a framework of permitted-use regimes decides in advance where it may occur. The Portuguese texts carry almost half of this category, and their range, from principles to regimes, is itself a finding about how assessment rules mature.

Values and protections

Values are spread across the whole set. Ethical and responsible use is named in twenty-one entries from six countries, critical thinking and verification in ten, data protection in nine and equity in four. The Belgian response alone accounts for six of the ethical entries, from a regional vision on responsible AI to a student council position. The values take a practical form in the classroom entries, where students are asked to cross-check results, keep personal data out of prompts and verify the sources a system gives them. At the national and institutional levels the same values appear as principles, most often transparency named together with human oversight.

Provision of tools and infrastructure

Provision is the least frequent of the substantive categories and the one described with the most qualification. Six entries supply a tool, four concern infrastructure and three describe personalised tutoring, and the Belgian response contributes none of them. Almost every provision entry carries a condition: a pilot, a limited adoption, an implementation not yet complete or a partnership with a commercial provider. National infrastructure appears where language is at stake, in the Pharos-CY antenna for Greek-language AI.

Reported absence

Two entries report that nothing exists, both from the same response and both below the national level. The absence is stated plainly and without qualification. It is kept as a category of its own because an explicit absence is information: it tells the reader that the question was asked and answered, which a missing entry does not.

The most frequent codes

Ethical and responsible use is the most frequent code, carried by 21 entries, followed by advisory guidance with 19. Professional development and institutional policy each appear in 14 entries, AI literacy in 11 and critical thinking and verification in 10. The list is led by the values the texts invoke and the form they take, and the first code on the provision of tools, with six entries, comes after twelve others.

Categories across levels

Each category has a level where it gathers. The form of the instrument dominates the macro level, where 15 of 18 entries carry it. Integrity and accountability gather at the meso level with 15 of 34 entries. Capacity building gathers at the micro level, where 14 of 18 entries carry it, and provision of tools is the only other category that reaches its highest share there. Several codes live at a single level: all eight national strategies are macro, all fourteen institutional policies and all six academic integrity codes are meso.

Codes that travel together

Some codes rarely appear alone. Transparency and human oversight share four entries and have the strongest association among pairs that share at least two entries, with a Jaccard coefficient of 0.44; they are named together as principles in Portuguese and Italian texts. Academic integrity appears with institutional policy in four entries, and critical thinking accompanies data protection in four, mostly in workshop and guidance texts that tell users to protect personal data and cross-check what the tool returns. The most frequent pairing overall is advisory guidance with ethical and responsible use, in nine entries, because guidance is so often written as a statement of responsible use; advisory guidance with critical thinking follows in six, since guidance often asks for verification. Professional development and AI literacy share six entries, mostly in training programmes for staff.

After cleaning, the responses contain 1,050 counted words. The most frequent are teaching, policy, student and research, followed by education, learning and institutional. The cloud shows a vocabulary of education and governance in which tools and technology hold little space: the word tool appears seven times, fewer than recommendation or framework.

The keywords separate the levels more sharply than the frequencies do. The macro level is marked by higher education, national, strategy and advisory, the vocabulary of frames. The meso level is marked by assessment, whose log-likelihood of 16.6 is the highest in the set, and by staff, institutional, responsibility, critical thinking and academic integrity. The micro level is marked by resource, model, material and personalised, the vocabulary of what is placed in front of a learner.

Five themes were built from the categories. The figure shows how often each theme appears at each level. An entry can belong to several themes, and the readings that follow explain what the counts mean.

National frames, institutional rules

National texts set the direction and leave the rule to someone else. Eight of the eighteen macro entries are strategies or plans, and two of them, in Greece and Italy, are explicitly aligned with the EU AI Act. The Belgian entry at this level is regional rather than national: the Flemish Ministry of Education's vision on responsible AI. Where national law does bind, it binds on data, rights and harm, as in the Greek decision on personal data, the Italian law on transparency and oversight and the Cypriot criminal provision, rather than on teaching. The Turkish response states the consequence directly: without a regulation specific to generative AI in education, the detail is left to universities. Binding rules on use appear accordingly at the institutional level, in fourteen policies and regulations, and not in any national entry.

Capacity before control

At the classroom level the responses describe what people learn and say little about prohibition. Fourteen of the eighteen micro entries carry a capacity code, and eight carry professional development. The Greek response consists at this level entirely of teacher programmes, from a certification with AI modules to a short course built on TPACK. The Cypriot workshops, the Anadolu modules on prompt and context engineering and the Belgian examples from the Leuven Learning Lab and the GPT Academy follow the same line. No micro entry is a binding rule. Of the two advisory texts at this level, KU Leuven's five principles state how generative AI should be used, and the Anadolu student guide asks students to verify information against two sources and to decline use where a course does not permit it.

Disclosure and assessment as the institutional grammar

When institutions do write rules, they write them about disclosure and assessment. Integrity and accountability codes appear in 15 of the 34 meso entries: disclosure in four, academic integrity in six and assessment design in seven. The Cypriot binding declaration for students and the Anadolu thesis declaration are the clearest cases, and the Portuguese texts add interaction records and permitted-use regimes for assessment. The weight of this theme rests heavily on one response, since seven of the nine assessment entries are Portuguese. It holds across responses in the narrower form of disclosure, which appears in three countries. The Belgian response enters the theme through research rather than assessment: two of its institutional texts address AI in research, and its only assessment entry is a set of Dutch-language articles on assessment validity.

Tools as tutors and pilots

Tools enter the micro level mainly as tutors and as trials. Five of the six tool entries are micro, and three of them describe personalised tutoring: a tutor in a computer science degree, a dialogic system that asks students to verify its sources, and generative AI inside a learner-centred model. The respondents describe much of this as unfinished. The Portuguese course-specific tutor is not yet complete and adoption remains limited, the Cypriot tools are being piloted, and the Greek classroom tool arrives through a pilot run with a commercial provider. Provision, in these responses, runs ahead of the rules written for it. The Belgian response, built around university policies, has no entry in this theme.

An uneven middle

The meso level holds the most entries and the least agreement about what belongs there. In the Greek response it is filled by national agencies: a policy institute, a bioethics commission and two research centres. In the Portuguese response it is filled by twelve documents from eleven institutions and faculties, and in the Belgian response by the policies and guidance of six universities, one of them in the Netherlands, together with a student council position. In the Dutch response it is empty, with no institutional governance and no classroom guidance reported, although the same university reported in 2024 that 96 per cent of its students and 95 per cent of its academic staff used these tools. The Anadolu case shows a further reading: the thesis declaration was decided by an institute rather than the rectorate, so an analysis looking only for a central text would find none. The middle level is where use is decided most often and recorded least consistently.

Macro · national policy

The macro level holds 18 entries from all seven responses. Strategies are the most common form, followed by laws and advisory frames, and two entries describe instruments still in preparation: the Cypriot strategy to 2032 under consultation and the Portuguese ministerial task force expected to report in autumn 2026. National infrastructure appears where language is at stake, as in the Pharos-CY antenna for Greek-language AI. The Belgian entry is a regional vision, issued by the Flemish Ministry of Education. Rights and data are protected in law: the Greek decision on personal data, the Italian AI law and the Cypriot criminal provision are the three binding entries at this level.

Meso · institutional governance

The meso level holds 34 entries and is where rules become specific. Fourteen entries are institutional policies or regulations, thirteen are advisory, and three describe bodies set up to carry the response: the OUC committee, the UAb task force and the Anadolu research unit. The vocabulary here is that of assessment, responsibility and integrity. The meso level also holds the only boundary drawn by data rather than by tool, the Anadolu rule that keeps personal data outside scope and runs examination item writing on local devices.

Micro · classroom and learning design

The micro level holds 18 entries, as many as the macro level, and the Dutch response reports that none are in place. What appears here is professional development, learning design and tools placed in the learning environment. Eight entries are teacher or staff programmes, seven concern learning design and five provide a tool. The classroom is the level the responses describe with most uncertainty, as pilots, limited adoption and unfinished implementation recur.

Greece · Institute of Educational Policy

The Greek response concerns school education and is built around teachers. Its macro level combines a ministerial decision on personal data with two national strategies, one of them aligned with the EU AI Act. Its meso level is filled by national bodies: the Institute's Greek translation of UNESCO's guidance, an opinion of the bioethics commission and a foresight study. Its micro level holds five teacher programmes, one of them a pilot run with OpenAI, which makes it the response with the strongest orientation towards professional development.

Cyprus · Open University of Cyprus

The Cypriot response spans school and higher education and places the university inside a national picture. The macro level includes a strategy under consultation, advisory guidelines for schools, a criminal provision and a language infrastructure. At the institutional level the university has advisory guidelines, a committee, seminars for staff, students and the public, and a binding declaration of use for students. Its micro entries are pilots in learning design and student workshops on data protection and cross-checking.

Portugal · Universidade Aberta

The Portuguese response is the most extensive and reads as a survey of the national higher education sector. Two national entries, a task force and a recommendation of the National Education Council, frame twelve institutional documents from eleven universities and faculties, which range from binding faculty regulations to institute guidance. The response carries most of the assessment and integrity codes in the set. Its single micro entry is candid about the gap between design and practice: the pedagogical model foresees a course-specific AI tutor that is not yet in place.

Netherlands · Open Universiteit

The Dutch response names the government-wide vision on generative AI and reports no institutional governance and no classroom guidance at the university. It is the shortest response and the only one reporting absence. Read beside the university's own survey of widespread use, the absence describes the state of written governance.

Italy · UNINETTUNO

The Italian response pairs a dense national frame with a university-built learning environment. A national strategy, a national AI law and ministerial policy directions form the macro level. The university's strategic plan integrates AI into its institutional development, and two published models describe generative AI inside its learning design, one of them a dialogic system that asks students to question and verify its answers. The meso answer ends mid-sentence in the submission.

Belgium · KU Leuven

The Belgian response arrived as ten separate submissions on one day, each naming a single source, and is read here as one response. It is built around universities. Its one entry above the institutional level is the Flemish Ministry of Education's policy statement and vision on responsible AI in education. Its institutional level reads as a survey of the sector: policies at KU Leuven, Ghent University, UCLouvain and UHasselt, instructions on AI in research ethics and a student council position at the University of Antwerp, and guidance on safe use from TU Delft in the Netherlands. Its micro entries are the Leuven Learning Lab's examples of responsible use, KU Leuven's five principles, an in-service GPT Academy run by UCLL and five Dutch-language articles on lesson preparation and assessment. Ethical and responsible use runs through six of its twelve entries, and none of them supplies a tool.

Türkiye · Anadolu University Open Education Faculty

The Anadolu response, compiled by the task lead, fills all three levels. National frames are strategic and advisory, and the Council of Higher Education's ethical guide on research leaves binding rules to universities. At the institutional level a graduate school made a declaration of use mandatory for theses, a research unit was established, a boundary was drawn by data rather than tool, and separate guides were published for academics and students, alongside an associate degree built around AI-assisted coding. The faculty's own position looks beyond policy text: requirements such as an audit trail, open protocols and provider-independent routing can be written into procurement, and staff development can move from individual training towards workshops in which colleagues compare their results on the same task.

The second tab of this report synthesises the reports of the ADMIT partnership into a database of 64 practices. The two sets were produced by different instruments: ADMIT by multi-round surveys of students, staff and administrators with institutional returns, the partner cases by a call for named reports and practices. To compare them on the same ground, every case entry was also assigned to one of the twelve domains used in the ADMIT matrix.

Levels: a similar shape

The level distributions are close. ADMIT places 30 of 64 practices at the meso level, 22 at the macro level and 12 at the micro level; the cases place 34 of 70, 18 and 18. The index of dissimilarity between the two distributions is 0.087, meaning that about nine per cent of entries would have to change level for the two to coincide (Duncan & Duncan, 1955). Most of the difference lies in the classroom, which holds a quarter of the case entries against a fifth of the ADMIT practices. The convergence supports a reading already made in the ADMIT synthesis, that the classroom level is under-documented rather than empty: both instruments ask what exists, and what exists in writing gathers at the institution.

Domains: governance steady, emphasis shifted

Across the twelve domains the index of dissimilarity rises to 0.163, so the two sets agree on levels more than on subjects. The governance domain is the steadiest point: it holds 21 case entries against 17 in ADMIT, a similar share of each set, and in both it sits at the macro and meso levels with none in the classroom. Around it the emphasis moves. Ethical frameworks and responsible use account for thirteen case entries against five in ADMIT, mostly through national opinions, institutional principles and the Belgian university policies. AI literacy and professional development are similar in size but sit lower: eight of the eleven case entries are micro, against four of nine in ADMIT. Student support changes level entirely, from four institutional services in ADMIT to three classroom tutors and one student guide in the cases. Curriculum is thinner in the cases, with five entries against seven, and quality assurance, partnership and equity, present though small in ADMIT, do not appear at all.

Models: what the cases confirm and what they add

The ADMIT synthesis describes seven recurring models. Several of them reappear in the cases in other countries, which makes them less dependent on a single partnership. Disclosure made a condition, which ADMIT found in a thesis declaration, appears in the cases as a binding student declaration in Cyprus and as disclosure provisions in two Portuguese institutions. The boundary drawn at the assessment event reappears in a more elaborate form, as regimes of permitted use defined for assessment at Aveiro and Iscte. The sequence of sandbox, unit and licence is visible in its first two stages, in pilots and in the committee, task force and research unit that three responses report; the licence stage appears only as a national pilot. Delegated discretion survives inside guidance: both the Universidade Aberta and the Anadolu student guides defer to what the teacher of the course allows.

The cases also add what the ADMIT instruments did not reach. They name a legal layer, a Greek ministerial decision on personal data, an Italian AI law and a Cypriot criminal provision, where ADMIT recorded mainly advisory national guidance. They bring in national advisory bodies and foresight studies as actors in their own right. And they describe AI tutors built into the learning design of two open universities, a practice the ADMIT rounds recorded mostly as experimentation. The observation that closes the ADMIT synthesis holds here as well: almost every entry is reported as existing, without an evaluation, and the pilots are described without their findings.

Institutions in both sets

Five institutions or systems appear in both sets, and reading them side by side shows where the record moved. At the Open University of Cyprus the two agree: the committee, the pilots in learning design and IT support and the seminars recorded in ADMIT appear again in the form, which adds a binding declaration of use for students, while the prohibition of use during examinations recorded in the third ADMIT report is not mentioned. At the Open Universiteit the two diverge. The first ADMIT report records an institution-wide steering committee installed in October 2024, whereas the 2026 response reports that no institutional governance currently exists; the sources do not show whether the committee is not regarded as governance or is no longer active. For Greece, both sets name the AI in Schools pilot with ChatGPT Edu, but ADMIT places it at the national level as provision of a tool and the form places it in the classroom as teacher development. For Italy, both name the national AI law, with different dates: the third ADMIT report gives 17 September 2025 and the form 23 September 2025, linking to the official consolidated text. The Anadolu answer agrees with ADMIT by construction, since it was compiled from the same sources. Portugal and Belgium have no counterpart in the ADMIT reports, so the twelve Portuguese and seven Belgian institutional entries are entirely new to the record.

The partner cases describe a division of labour between levels. National texts give direction and protect rights; institutions write the rules that bind users, chiefly on disclosure and assessment; and classrooms receive training, design support and tools that are often still being tested. No response places a binding rule at the classroom level, and none places the provision of a tool at the national level except through infrastructure or a pilot.

The unevenness of the middle level is the finding with the most practical weight. The same question about institutional governance returned national agencies in one country, twelve documents in another, the policies of six universities in a third and nothing in a fourth. A comparison across the partnership therefore depends on recording, for each response, which body took the decision and at which level, as well as whether an instrument exists. Where the classroom is concerned, the cases suggest collecting accounts of practice directly, since instruments that ask for documents return documents.