Mapping and Comparative Analysis of Best Practices Across Educational Levels
This report is the deliverable of Task 2.2 of Work Package 2 of the EthicAI project, led by Anadolu University with the contribution of all partners. The task is to map the practices through which generative artificial intelligence has entered European higher education, and to compare them across the levels at which they are decided. Practices are recorded at three levels, namely the classroom and learning design, institutional governance and national policy, and are classified across twelve domains of practice, from curriculum and assessment to data protection, equity and procurement. Each record is linked to the source that reports it, to the page on which it appears and to the sentence found there, so that every statement in this report can be checked against the document it rests on. The outcome is presented in two forms: a curated database of practices and the synthesis that reads across it.
The imprint of the report, its coverage and how its figures are to be read.
Each level has its own actors, its own decision rights and its own form of evidence. Opening a card shows how that level is distributed across domains.
The six domains where the records gather most densely, each with the reading it carries.
The three findings the mapping establishes, and the shape of the sixty-four records.
This report presents a mapping and comparative analysis of generative artificial intelligence practice in European higher education across three educational levels and twelve practice domains. The analysis draws on three rounds of partnership reporting on trends, policies and practices, together with the guidance documents and frameworks published alongside them. Sixty-four practices are recorded, each linked to the page of the source on which it rests and to the sentence located there.
The distribution of the records constitutes the first finding. Documented practice is concentrated at the institutional and national levels and within the governance domain, which together account for approximately one quarter of the records. The micro level is comparatively thinly documented, with twelve of the sixty-four records situated inside a single course. This imbalance appears to be a property of the evidence rather than of teaching practice itself. The source reports survey awareness of policies and practices, and only the third round posed a question capable of eliciting an account of how a single course was redesigned. Where that question was posed, teachers described the preparation of materials and the design of learning activities as routine work.
The second finding concerns the binding force of the instruments recorded. National instruments are numerous and, with few exceptions, advisory in character. Partners state this directly, and one account draws the consequence explicitly: no standardised regulation for generative AI in education exists, which leaves universities responsible for their own institutional policies. The instruments that bind are institutional and specific. They include a declaration form signed as a condition of thesis submission, a prohibition on use during examinations, an internal regulation of learning analytics described as mandatory, and a journal rule refusing authorship to a model.
The third finding concerns direction of travel. Read across the three rounds, institutions appear to move from delegated discretion towards written frameworks, and from single tools towards portfolios and shared provision. This movement is uneven. Institutions reporting no change across two rounds are recorded alongside an institution that issued a second version of its AI guidelines and a third version of its recommendations for action within the same year, and the gap between the policy awareness of staff and that of students narrows without closing: students report the largest gain between the two survey years, from 1.38 to 2.46 on the institutional awareness scale, and still report the lowest awareness of any group.
The coverage of the analysis, the definition of the three levels, and the questions the available material can and cannot answer.
The task addressed by this report is the mapping and comparative analysis of generative AI practice across educational levels, with the outcome defined as a curated database accompanied by a synthesis report. The three levels are taken from that definition. The micro level covers the classroom and learning design, the meso level covers institutional governance, and the macro level covers national policy.
A level is assigned according to where the decision constituting the practice is taken, and not according to who is affected by it. A national ethical guide addressed to researchers is therefore recorded as a macro-level instrument, notwithstanding that each of its readers works within an institution. An assessment tool provided centrally but configured course by course is recorded at the micro level, since what the tool communicates to a student is determined by the teacher responsible for that course.
A further rule applies where a single institution issues paired instruments for different audiences. The instrument addressed to students is recorded at the micro level and the instrument addressed to academic staff at the meso level. Two such pairs appear in the material, and in both cases the pairing reflects a deliberate decision by the issuing institution rather than an artefact of the coding procedure.
Figure 1. The diagram sets out the rule by which every record in this report was assigned to a level. A level is determined by where the decision constituting the practice is taken and not by who is affected by it. Accordingly, a national guide addressed to researchers remains at the macro level, while a centrally provided tool configured course by course is assigned to the micro level. Reading the three bands from the bottom upwards also accounts for the density of the matrix, in that the two lower bands operate through instruments that can be inspected and counted, whereas the upper band operates through practices that exist only where a respondent reports them.
The source reports were designed to measure awareness, purpose of use, opportunities and challenges among four stakeholder groups, and to record the policies and practices that partner institutions and countries report about themselves. They are therefore able to answer questions concerning what exists, who is aware of it, and how it is regarded.
They are less well suited to answering questions concerning effect. Evidence of outcome is scarce throughout the material, and where partners report a practice they generally report its existence rather than its evaluation. Claims about effectiveness are consequently absent from this report, and the maturity, scalability and transferability codings recorded against each practice are analytical judgements derived from the stated evidence rather than measurements reported by the sources.
One property of an instrument is reported with sufficient consistency to be read across the material, and it separates the three levels more sharply than volume does, namely whether the instrument binds those to whom it is addressed.
Figure 2. Position on the axis is derived from what the source states about the instrument rather than from an assessment of its apparent importance. Where a partner records that guidance is advisory, or that a regulation is mandatory, the wording determines the position; where binding force is left unstated the instrument is placed conservatively. Read from left to right, the axis represents the central asymmetry identified in this report: the numerous instruments are located on the advisory side, while the few that bind are institutional and narrow in scope, with one national exception that binds evaluators rather than those being evaluated.
The reading of the sources, the extraction and coding of practices, and the procedure by which every claim is linked to a quotation and a page.
Thirteen documents were read in full. Ten are published project deliverables amounting to 435 pages; the remaining three are a working draft of the third-round report and two published guides. A programme description issued by one institution was consulted in addition, as the source for a single record. The principal weight rests on the three rounds of partnership reporting on trends, policies and practices. The remaining documents were reviewed for corroboration and for instruments that they name and the three rounds do not.
The three rounds are not equivalent in design. The first combined questionnaires with 140 semi-structured interviews across four stakeholder groups. The second dispensed with the interviews in favour of a larger questionnaire and introduced temporal comparison against the first. The third continued with questionnaires alone across eleven partner universities. Sample sizes differ substantially between rounds, and the second report characterises its own sample as a convenience sample. Comparisons across rounds are therefore treated in this report as directional rather than as measurements of change.
Figure 3. The three rounds differ markedly in size and composition, which is the reason comparisons between them are treated in this report as directional rather than as measurements of change. Students predominate in the first and third rounds, while the second round is considerably smaller across every group, and the second report characterises its own sample as a convenience sample. Any reading that compares a figure from one round with a figure from another must therefore take that difference into account.
A practice was recorded where a source names a specific instrument, tool, programme or rule attributable to an identifiable actor. Statements of intention, general attitudes and reports of increased use unaccompanied by a named instrument were not recorded as practices, although such material informs the readings of the matrix cells.
Where a source does not name the institution responsible for a practice, no institution is attributed. One record in the database carries an explicit note that the source leaves the institution unstated, rather than inferring it from context.
Each practice is coded on the axes defined by the task: level, domain, delivery model, context, maturity, scalability and transferability. Delivery model is recorded only where the source supports it, which leaves the field empty for the majority of records rather than completing it by inference from the character of the institution.
Maturity, scalability and transferability are analytical judgements, each bounded by what the source states. Where an institution describes an environment as experimental, or a policy as expected in the following year, the coding does not exceed that description.
Every practice record carries at least one anchor, consisting of a verbatim quotation from the source, the page on which it appears, and a statement of the reason that quotation supports the claim being made. Anchors are verified mechanically: the quotation is searched for in the source document itself, and the page recorded is the page on which the search located it. An anchor whose quotation cannot be located is marked as unverified and counted as such on the source trail page. Page numbers refer to the physical page of the source file, counted from its first page, so that a reader may open the document at that page.
Readings of matrix cells are syntheses of the practices recorded beneath them and are supported by the anchors of those practices. Where a cell reading makes a claim about absence, that claim can be checked against the database itself.
Rows are practice domains, columns are the level at which the practice is enacted: micro for classroom and learning design, meso for institutional governance, macro for national policy. Each cell reports how many documented practices were found and opens the underlying entries.
| Domain | MicroClassroom and learning design | MesoInstitutional governance | MacroNational policy |
|---|---|---|---|
| Curriculum and Learning Design | |||
| Teaching Practice | |||
| Assessment and Academic Integrity | |||
| Student Support and Guidance | |||
| AI Literacy and Professional Development | |||
| Governance, Strategy and Policy | |||
| Ethical Frameworks and Responsible Use | |||
| Data Protection, Privacy and Security | |||
| Equity, Inclusion and Accessibility | |||
| Infrastructure, Tools and Procurement | |||
| Quality Assurance, Monitoring and Evaluation | |||
| Partnership, Networks and Transferability |
Table 1. Rows give the domains of practice and columns the level at which a practice is enacted. Opening a cell brings up who applied it, where and how at that intersection, without leaving the page. The count indicates density rather than the finding itself.
Rows are the countries the records come from and columns are the levels. The order follows density rather than the alphabet: seeing at which level a country works says more than seeing how many records it has.
| Country | MicroClassroom and learning design | MesoInstitutional governance | MacroNational policy | Total |
|---|---|---|---|---|
| Türkiye | 1 | 5 | 6 | 12 |
| Multiple countries | 9 | 1 | — | 10 |
| Spain | — | 6 | 3 | 9 |
| Finland | — | 5 | 2 | 7 |
| Germany | 1 | 4 | 2 | 7 |
| United Kingdom | — | 3 | 3 | 6 |
| Switzerland | 1 | 2 | 1 | 4 |
| Cyprus | — | 2 | 1 | 3 |
| European Union | — | — | 2 | 2 |
| Greece | — | — | 1 | 1 |
| Italy | — | — | 1 | 1 |
| Netherlands | — | 1 | — | 1 |
| Not stated in the source | — | 1 | — | 1 |
Table 2. The country of a record is where the practice takes place, not the country of the partner reporting it. Records whose source does not state a country do not enter the table.
Rows are the institutions and actors carrying out the practice. An institution appearing at more than one level addresses the same technology through different audiences; one appearing at a single level has not yet moved beyond that layer.
| Institution | MicroClassroom and learning design | MesoInstitutional governance | MacroNational policy | Total |
|---|---|---|---|---|
| University of Jyväskylä | — | 5 | — | 5 |
| FernUniversität in Hagen | 1 | 4 | — | 5 |
| Universitat Oberta de Catalunya | — | 4 | — | 4 |
| Anadolu University | — | 3 | — | 3 |
| The Open University | — | 3 | — | 3 |
| Anadolu University, Open Education Faculty | 1 | 2 | — | 3 |
| UniDistance Suisse | 1 | 2 | — | 3 |
| European Union | — | — | 2 | 2 |
| Open University of Cyprus | — | 2 | — | 2 |
| Universidad Nacional de Educación a Distancia | — | 2 | — | 2 |
| ADMIT partnership (EADTU) | — | — | 1 | 1 |
| Conference of Spanish Rectors (CRUE) | — | — | 1 | 1 |
| Council of Higher Education (YÖK) | — | — | 1 | 1 |
| Council of Higher Education and Presidency of Defence Industries | — | — | 1 | 1 |
| Digitale Hochschule NRW | — | — | 1 | 1 |
| Federal Ministry for Research, Technology and Space | — | — | 1 | 1 |
| Finnish National Agency for Education and Ministry of Education and Culture | — | — | 1 | 1 |
| Government of Cyprus | — | — | 1 | 1 |
| Government of Greece | — | — | 1 | 1 |
| Government of Spain | — | — | 1 | 1 |
| Government of the United Kingdom | — | — | 1 | 1 |
| Italian ministries | — | — | 1 | 1 |
| Measuring, Selection and Placement Centre (ÖSYM) | — | — | 1 | 1 |
| Ministry of Interior | — | — | 1 | 1 |
| Open Universiteit | — | 1 | — | 1 |
| Quality Assurance Agency for Higher Education | — | — | 1 | 1 |
| Swiss Confederation and State Secretariat for Education, Research and Innovation | — | — | 1 | 1 |
| Turkish national authorities | — | — | 1 | 1 |
| UK Research and Innovation | — | — | 1 | 1 |
Table 3. Records whose source does not name the institution are absent from the table; attaching an unattributed record to an institution would state what the source does not.
Practices enacted by teaching staff and students in the immediate learning situation: course design, classroom activity, materials, feedback and individual assessment decisions.
12 records are documented at this level, spread across 6 domains. The weight lies in AI Literacy and Professional Development and Teaching Practice.
Practices established at faculty, department or institution level: strategy, regulation, support structures, staff development, procurement and quality assurance.
30 records are documented at this level, spread across 9 domains. The weight lies in Governance, Strategy and Policy and Assessment and Academic Integrity.
Practices set by national and supranational actors: legislation, ministerial guidance, funding instruments, sector-wide frameworks and accreditation requirements.
22 records are documented at this level, spread across 8 domains. The weight lies in Governance, Strategy and Policy and Curriculum and Learning Design.
Practices absent in an earlier round and present in a later one, and the distribution of the documented records.
An emerging practice is defined here as one that was absent in an earlier round of reporting and present in a later one. Reading the three rounds in sequence renders several such practices visible.
Sixty-four practices are recorded in the database. The governance domain alone accounts for seventeen of them, distributed between the institutional and national levels. Capacity building follows with nine records, and assessment and curriculum hold seven each. By level, thirty records are situated at the institutional level, twenty-two at the national level and twelve inside the single course.
This distribution reflects the design of the instrument as much as the state of the field. A survey that asks which policies exist will identify policies. It will not identify a revised course unless a teacher volunteers an account of one, and only the third round posed a question capable of eliciting such an account, in that it asked not which rules apply but what had changed in the way the tools are used. The responses obtained are the reason the curriculum row can be read at the micro level at all, and they were obtained one round later than the instruments recorded above that level. The micro level should therefore be understood as under-documented rather than under-populated, and the count recorded there reflects the frequency with which the question was asked as much as the frequency with which the practice occurs.
Figure 4. The values are the means reported by the second-round report for both survey years, measured on a scale in which a higher figure indicates greater awareness of institutional policy. Three of the four groups rise between the years and students rise the most, from 1.38 to 2.46, while the awareness reported by IT and support staff falls slightly, from 3.43 to 3.33. That small decline changes the top of the ranking, since administrative staff report the highest awareness in 2025 where support staff did so in 2024. At the bottom the position is unchanged: students report the lowest awareness in both years and remain below every other group. The difference of significance is therefore not one of attention but one of distribution, in that those governed by these policies are the least likely to be aware of their existence.
Figure 5. Placing the two measures side by side indicates that the decline from institutional to national awareness holds for every group without exception. It is smallest among students, who report the lowest institutional awareness to begin with and therefore have least to lose, and largest among teachers, whose reported awareness falls by 1.30 points on the same scale. The pattern supports a reading advanced throughout the source reports, according to which national frameworks exist but do not reach the questions higher education actually faces. Awareness of a policy is not equivalent to compliance with it, and the figure should therefore be read as a measure of visibility rather than of effect.
Several practices are new in the sense relevant here. Institution-wide access to a commercial model was extended to all staff at one university, teaching and administrative alike, and was described as a transition from isolated use to integration across the institution. A regional portal began to offer generative AI applications to every university in a federal state, open to staff and students on the same terms.
Assessment moved from statements of principle towards instruments. One institution now prohibits use during examinations and permits it in the preparation of assignments where transparency is ensured. Another attached a signed declaration to thesis submission and made the form a required document.
Curriculum moved as well. An open education faculty now offers a two-year associate degree in AI-assisted coding, whose learning outcomes are written around large language models and generative AI applications rather than adding the subject to an existing programme. Capacity building entered the curriculum in the same period: a module on critical AI literacy was offered to students from autumn 2025, which makes participation part of study rather than an optional addition to it. A pair of published guides also appeared, one addressed to students and one to academics, containing rules that can be applied rather than merely assented to.
Governance moved from delegation towards written provision. A university that had previously left the decision to individual teachers published an internal framework covering educational, administrative and support services. Another established a working group whose institution-wide policy was expected in the following year.
An authority issues an ethical guide, a strategy or a recommendation. Partners report the force of such instruments as advisory, and one adds that no standardised regulation exists for generative AI in education.
An ethical guide on generative AI in scientific research and publication
A faculty revises a guide, adds a form, sets a rule for examinations or writes a regulation. The act is narrow, dated and attributable, and it is what converts guidance into a condition.
The thesis writing guide is revised and a declaration form is attached
A student signs, a teacher applies a boundary, an author cannot attribute authorship to a model. At this point the instrument is no longer advice: it has a signature, a deadline or a refusal attached to it.
The form is a required document for submission
Figure 6. The sequence addresses a question the matrix raises but cannot itself display: if almost every national instrument is advisory, what renders any of it binding upon a student or a member of staff. The conversion is effected through an institutional act, and that act is generally small and procedural rather than strategic. One country appears twice in the material at the first stage, once through higher education policy and once through public administration, and neither route binds until an institution performs the second stage.
One absence has closed and one has not. Teachers now describe the preparation of materials and the design of learning activities as routine work, which is the account the first two rounds did not obtain, and the curriculum row can consequently be read at the level of the single course for the first time.
Equity has not moved. One record treats equity as a design requirement rather than as a concern to be noted, and at the national level the only measure named anywhere in this material is a recommendation, delegated to individual universities, that an equal opportunities programme or an access subsidy programme should exist. Students raise unequal access and the cost of premium tools in every round, while the institutional response recorded here is a licensing decision rather than an equity policy. The training gap indicates the same pattern: students report less access to formal provision than staff and learn instead through public video and social media, and that difference did not narrow between the two survey years.
The recurring forms that documented practice takes, described so that they may be recognised in settings the records do not cover.
Several practices recorded in different countries take the same form. Describing that form makes it recognisable in settings the records do not cover.
An institution issues two documents on the same subject, one addressed to students and one to staff. This pairing appears twice in the material: a published student guide alongside a published guide for academics, and a digital assistant piloted for students alongside a writing assistant piloted for the staff who author their modules.
The model is significant because it fixes the level at which each instrument operates. The student-facing instrument works inside the learning situation, while the staff-facing instrument works on the structures surrounding it. An institution adopting this model addresses the same technology twice, and the intended audience determines the level.
A declaration of use is attached to an existing submission process and made a requirement. The instrument is small and procedural. It requires no infrastructure, no licence and no committee, and it converts advisory national guidance into an obligation that reaches the individual student.
Rather than permitting or prohibiting a particular tool, an institution draws the boundary at the moment of assessment. Use is prohibited during examinations and permitted in the preparation of assignments where transparency is ensured. The model separates the question of learning from the question of demonstrating learning, and it remains applicable through changes of tool because it names none.
Institutional capability accumulates in three stages. An experimental environment is opened so that tools may be trialled. A unit or committee is then established to hold the work. Access is subsequently extended to a wider group through a licence or a shared portal. Records from several universities are situated at different points along this sequence, and reading them together reveals a sequence rather than a single decision.
Procurement, access and cost are settled once for many institutions rather than repeatedly for one. A regional portal and a national pilot both take this form. The model reaches institutions that could not have negotiated independently, and it relocates the equity question from what a student can afford to what a system has licensed.
While policy is being drafted, each teacher determines the extent to which the tool may be used within their own course. The model allows teaching to continue through a policy vacuum. Its cost is that students encounter different rules in different courses, and the material records at least one institution moving out of this arrangement once a framework had been published.
Materials are published for staff and kept current as new material appears, rather than issued once and revised on a fixed cycle. The model addresses a difficulty that fixed guidance does not, in that the tools change more rapidly than documents can be approved.
The practices that continue to function as the number of students, courses or institutions increases.
Scalability is treated here as a single question: if the number of students, courses or institutions doubles, does the practice continue to function without a corresponding doubling of the effort behind it.
A rule scales. A declaration form attached to submission carries the same cost whether ten students or ten thousand sign it. A prohibition applying during examinations governs every examination simultaneously. A journal rule refusing authorship to a model governs every submission to that journal. A published guide is written once and read by all who receive it.
These are the practices coded highest for scalability in the database, and they share a common property, in that none of them requires a person to act on each individual case.
A shared portal scales by design, having been built for many institutions from the outset. A licence extended to all staff scales to the size of the licence. A course delivered through the learning management system reaches everyone already using that system.
The cost in these cases is fixed rather than marginal. The portal must be funded, the licence negotiated and the course maintained. This is the reason such practices appear where a funding structure exists to carry them, and the reason the material records them more frequently at system level than at institutional level.
Certain practices scale only with the time of the people who operate them. Formative feedback configured for each task depends on teaching staff writing criteria course by course; the tool serves any number of students, but the criteria are not generated automatically. A portfolio of student-facing systems depends on the data, the platform and the project team behind it. The practice of checking every output before it reaches a student depends on the person performing the check.
These are not lesser practices. They are the practices situated closest to learning, and the material indicates that the closer a practice sits to the learning situation, the less readily it scales without additional staff time.
Comparison across institutions must therefore take account of the enabling condition as well as of the practice. A record reporting that formative feedback is available conveys little unless it also reports who writes the criteria. The database retains that condition alongside each record, which is the reason the scalability coding is presented together with the evidence rather than in isolation.
The practices that transfer to another institution or system, the conditions under which they transfer, and the points at which comparison across the partnership becomes unreliable.
The practices that transfer most readily share a common property, in that they specify a procedure rather than a system. Examples include a declaration form signed at submission, a boundary drawn at the examination, a rule refusing authorship to a model, a requirement to verify content against two sources, and a prohibition on generative tools in peer assessment. Each of these can be adopted by an institution that owns no infrastructure and licenses no tool, which is the reason each is coded highly on transferability.
A form naming the tool used and the support received, signed as a condition of submission.
A submission process that already collects documents from the student.
Use prohibited during examinations and permitted in preparation where transparency is ensured.
Assessment regulations that distinguish preparation from the examination itself.
A rule requiring definitions, figures and references to be checked against at least two reliable sources.
Course instructions specific enough for a student to defer to them.
Immediate feedback on written work against criteria the instructor writes for each task.
Teachers with the time and the authority to write those criteria course by course.
Central provision of generative AI applications to every university in a federal state.
A funding structure able to procure once on behalf of many institutions.
Early-warning systems, assisted feedback and conversational agents held under one project.
The data, the platform and the project team that carry it.
Figure 7. Transferability in this material follows a single distinction, namely whether a practice specifies a procedure or depends on a system. A procedure can be adopted by an institution that owns no infrastructure, which is the reason disclosure rules, examination boundaries and verification requirements are coded highly. A system carries its enabling conditions with it, and those conditions are frequently not visible in the report describing the practice. The right-hand column therefore records the condition alongside the practice, since a practice detached from its condition would read as a recommendation that was never intended.
The practices that transfer least readily are those sustained by local arrangements. A portfolio combining early-warning systems, assisted feedback and conversational agents depends upon the data, the platform and the project that carry it. A regional portal depends upon a federal structure able to fund shared provision. A national programme channelling a field of application into doctoral and master's theses depends upon an authority able to convene universities and an industry body together.
A practice detached from its enabling conditions cannot be read as a recommendation. The clearest instance in this material is the assessment tool that provides immediate formative feedback against criteria set by the instructor for the task in question. The tool transfers readily; the practice transfers only where teaching staff have the time and the authority to write those criteria course by course. Reported without that condition, the same record would suggest that automated feedback is available for adoption, which is not what the source describes.
The same consideration applies to the guidance documents. A student guide whose rules are operational rests upon a course instruction that is itself sufficiently specific to be deferred to. Where course instructions are silent, a rule that defers to them has no object.
Three obstacles limit comparison across the partnership. The first is that partners report about themselves, as the reports themselves state. A partner may hold instruments that were not captured because the representative completing the return did not list them. The second is that the rounds differ in design and in sample, so that a change observed between rounds may reflect who responded rather than what changed. The third is that binding force is reported unevenly. Some partners state whether an instrument binds and others do not, and where such a statement is absent the instrument has been coded conservatively.
These limits bear upon the density readings of the matrix more than upon the individual records. An empty cell indicates that no practice was documented at that intersection in these sources, and it should be read accordingly rather than as a claim about European higher education as a whole.
The findings established by the mapping, and the additions the next round of evidence would need to make.
The mapping establishes three findings. Documented practice within this partnership is concentrated in the governance domain and at the institutional and national levels, while practice inside the single course remains comparatively unrecorded. National instruments are plentiful and, with one clear exception, advisory in character, which displaces operative rule-making to institutions and, within institutions, on occasion to individual teaching staff. The movement visible across the three rounds runs from delegated discretion towards written frameworks and from single tools towards shared provision, proceeding unevenly and without closing the gap between the policy awareness of staff and that of students, which narrows between the two survey years but leaves students the least aware group.
The additions required of the evidence base are narrower than they might appear. Instruments already exist for asking institutions what they hold. What is missing is a question that asks a teacher what changed in a course. The third round provides partial support for this observation: where respondents were asked how their use had changed over the preceding year, teachers volunteered accounts of preparing materials and designing learning activities, and those accounts are the reason the curriculum row can be read at the micro level at all. A single open question addressed to teaching staff concerning a specific revision to a course or an assessment would populate the level that remains thinnest, and would do so with accounts that can be checked against course documentation rather than recalled as impressions.
The second addition concerns effect. Almost every practice in this database is reported as existing rather than as evaluated. Where an institution has conducted a pilot, the findings of that pilot are rarely reported, and where a tool has been in use across two rounds, no measure of its use accompanies the record. A field recording what was observed after adoption would change what a mapping of this kind is able to support, moving it from a description of what exists towards a comparison of what holds.
The curated records of Task 2.2. The account view reads them grouped by level and domain; the database view opens the same records for filtering, sorting and downloading.
Teachers report a transition from occasional assistance with summaries and ideas towards regular use in preparing materials and designing learning activities. The account is given as a description of current routine rather than as a statement of intention, and it situates a design decision, and not merely a decision about use, within a single course.
A teacher reports assigning a voice to the model, creating several agents specialised in different areas of expertise, and selecting models that reflect a personal teaching style. The tool is not adopted as delivered but configured, and the configuration expresses a pedagogical preference rather than a technical one.
Teachers describe subjecting every AI output to a check for academic precision, pedagogical alignment and ethical standards before use, and describe the tool as a partner for refining ideas rather than an author. The framing that dominates the third round is augmentation rather than replacement, and it recurs regardless of the purpose of use.
Students describe using the tool to obtain direction, to get help with parts they do not understand and to identify the strengths and weaknesses of work they have already done, rather than to have the work done for them. The practice is located within the learning situation and depends on nothing other than the student's own decision as to what to ask.
A tool gives students immediate AI-generated formative feedback on text assignments against criteria the instructor defines. Teachers add contextual material and set expectations for each task, and these elements are passed to the model through quality-secured prompts, which keeps the criteria in the teacher's hands while the response time falls to the student's.
Students and teachers report declining to trust generated text, avoiding direct copying, and verifying academic material in particular, with citations identified as the element verified most frequently. The practice addresses the failure mode that respondents in every group identify as the principal challenge.
A published guide addressed to students positions generative AI as a producer of drafts rather than an authority. It teaches prompt construction from a minimal three-part form up to a five-part template, maps use onto the levels of Bloom's taxonomy, and closes with four practices: withhold identifying data, verify critical claims against at least two sources, treat the course instruction as binding, and state the purpose for which the tool was used.
A module on critical AI literacy for learning was offered to students from autumn 2025. Placing the subject inside a module rather than a workshop makes participation part of study rather than an optional addition to it.
Respondents in every group attribute improved results to their own learning: prompts have become more specific and goal-directed, extended exchanges are sustained more deliberately, and use has moved from a single tool to a set of tools selected for particular purposes. Explaining a task to the model is described as constituting a learning experience in itself.
Where formal training is absent or limited, competence is acquired through colleagues, peer networks, self-paced material and practical use. The channels differ by role: staff describe internal knowledge sharing and consultation with colleagues, while for students public video and social media are the most common route. The report finds that informal learning acts as a complement to formal provision rather than as a substitute for it.
Respondents across all groups report stating where and for what purpose generative tools were used, citing them where the work is academic, and reviewing generated content thoroughly before it reaches anyone else. Disclosure is expressed as a description of purpose rather than as an acknowledgement of a tool.
Respondents in every stakeholder group report withholding personal, sensitive and confidential material from prompts, preferring platforms the institution has licensed to publicly available ones, and using temporary chat modes so that no record is retained. The report characterises the pattern as a habit rather than as compliance with a rule.
An open education faculty offers a two-year associate degree in AI-assisted coding, carrying 120 ECTS. The programme is not a module added to an existing curriculum but a qualification whose outcomes are written around large language models, agent-based systems and generative AI applications, alongside algorithmic thinking, cloud technologies and AI-supported software testing.
The creation of personalised AI-aided learning appears in the plan for the 2025 academic year of an open university. Placing the intention in an annual plan is what separates it from an experiment: a plan carries a date, an owner and a review.
A teacher describes artificial intelligence being applied as part of study modules while modules about artificial intelligence are offered to teachers and students alike, often using the technology itself as the starting point. The account records curriculum change from two directions at once: the technology enters existing modules and becomes the subject of new ones.
The Graduate Education Institute revised its thesis writing guide in 2024 so that every student signs a declaration form, naming the generative AI programme used and the kind of support received. The form is a required document rather than a recommendation, which moves disclosure from good practice to a condition of submission.
A policy addressed to the whole university community aims at transparency on all sides while setting limits on where use is permitted. Use is prohibited during examinations and may be permitted for preparing assignments or activities provided transparency is ensured, which draws the line at the assessment event rather than at the tool.
A funded pilot examines the challenges and opportunities that generative AI poses for assessment in further and higher education. Placing the question inside a funded research project rather than a policy circular is what allows assessment redesign to be tested before it is prescribed.
Guidelines addressed to teachers link the use of ChatGPT directly to remote assessment. The pairing matters in a distance institution, where assessment is conducted at a distance by default and the tool question cannot be separated from the examination question.
A chatbot answers student questions on enrolment, examinations and graduation. The institution is explicit that the deployed assistant is rule-based rather than generative, while research on integrating large language models into academic and non-academic support runs separately at the Student Affairs Department of the Faculty of Open Education.
Operational practices introduced and expanded since October 2024 include AI-based early-warning systems for students, AI-assisted feedback, empathic conversational agents, pilots for automated assessment and personalised learning support. The range shows an institution moving from single tools to a portfolio held under one project.
Pilot projects at an institution the source does not name trial two institutional tools: a digital assistant intended to support students as a study companion, and a module writing assistant intended to support staff in creating and reworking content. The pairing repeats the pattern seen elsewhere, where the same institution builds one instrument for learners and another for the staff who design their learning.
A chatbot named Minerva has been in use on the university web pages since 2023. Participants in the third round still name it as the institutional practice they recognise, which makes it one of the few assistants in the partnership that respondents identify by name across rounds.
The staff counterpart of the student guide addresses academics and researchers. On the teaching side it covers learning outcomes, operations inside the learning management system, assessment and rubrics, accessibility and inclusion. On the research side it covers preparation for literature review, structured reading notes, translation, method and reporting checklists, and reference verification. Data minimisation, disclosure and institutional policy compliance are treated as obligations rather than recommendations.
Resources developed for educators and learners include a course on AI in teaching hosted in the learning management system, an experimentation environment, and a regional hub whose materials teachers can integrate into their own courses. Delivering staff development inside the same system staff already teach in removes one barrier to participation.
A workspace of materials on the basics of artificial intelligence in university work was published for teachers, researchers and support staff in 2025 and is updated as new material appears. A living workspace answers a difficulty that fixed guidance does not, since the tools change faster than documents are revised.
The vice-rectorate for educational innovation runs a training and awareness programme built on several pillars, accompanied by guides for generalists, teachers and students and by courses for teaching, administrative and service staff. Locating the programme in a vice-rectorate gives it a budget line and a permanent owner.
The university maintains a set of instruments that are revised regularly: an AI guide, recommendations for action, a service agreement on the use of AI systems, and a register of processing activities. The combination places advisory guidance and data protection documentation in the same institutional package.
A policy issued by the rector in April 2024 is accompanied by instructions for using AI-based applications in studies. Issuing the instrument at rectoral level gives it institution-wide reach rather than leaving adoption to individual faculties.
A framework introduced in 2025 provides internal guidelines for integrating AI across educational, administrative and support services, with the stated aim of turning the university into an AI-enabled institution. It answers the gap left a year earlier, when policy was still under discussion and the decision was delegated to individual teachers.
An AI committee was established alongside pilot tests of tools in learning design and IT support and seminars for staff, students and the general public. Extending the seminars beyond the university is unusual in the partnership and treats AI literacy as a public rather than an internal concern.
While institution-wide policy was still under discussion, the university adopted an interim rule under which each teacher decides the level at which AI may be applied in learning activities and may rule it out entirely. The arrangement keeps courses operating during a policy vacuum at the cost of leaving students facing different rules in different courses.
A steering committee covering the whole institution was installed in October 2024, at a point when the university reported no institutional practices in place. The sequence constitutes a finding in itself, in that the governance structure was established before any application had been deployed.
A working group was set up to develop an institution-wide AI policy and guidelines, with availability announced for 2026. Recording the group rather than the policy is itself the finding: in the second round the institution had teacher guidelines but no institutional instrument.
Students may use generative AI to support their learning and, unless instructed otherwise, in assessment preparation, provided the use is acknowledged. The policy pairs permission with three prohibitions covering confidential information, personal data and copyrighted assessment content, and warns that over-reliance may hinder skill development.
A library guide published in May 2025 proposes structures for making the use of AI tools transparent: which tools were used, for what purpose and to what extent. A related journal policy states that authorship cannot be attributed to AI, that its use is confined to improving language or readability, that it must be cited, and that authors remain responsible for the validity of what it generates.
Information security guidelines issued in December 2023 direct users to work primarily with Microsoft Copilot on a university device while logged in with university credentials. The instrument governs data exposure by prescribing where the interaction takes place rather than by listing prohibited content.
The institution updated its AI guidelines in April 2025 and its recommendations for action in July 2025, both advisory in character, while an overarching AI strategy and guidelines for AI in examinations were under development. Alongside them the university maintains a mandatory internal regulation of learning analytics to ensure data privacy, which is the one binding instrument in the set.
New assessment approaches aim to keep tasks robust to the use of AI while maintaining academic integrity and providing differentiated support for diverse needs. Staff capacity is developed through training on responsible use, ethics and inclusivity, with the institution reporting a shift from high-level guidance to concrete application.
A dedicated research unit was established in 2023 to study developments in data analytics and artificial intelligence in higher education and to run projects, with training planned at several educational levels. The institution reports that the infrastructure of the unit was still being set up at the time of the survey.
The university runs experimental environments that bring a feedback centre, automated grading, personalised support and an immersive collaboration hub together with access to various models for testing. The arrangement gives staff a bounded space in which tools can be trialled before any decision about institutional deployment.
An administrator describes access to a commercial model being extended to all staff, both teaching and administrative, and frames it as a move from isolated or experimental use to integration across the university. Universal access changes the equity question from who can afford a subscription to what the institution has licensed.
A student reports that every university in the country now offers courses related to AI, and a teacher in the same round describes AI-related content being incorporated into curricula alongside the establishment of dedicated research centres. Two independent accounts describe the same national movement from different positions.
A support staff respondent reports that certificate programmes, associate degrees and undergraduate degrees in the field are available at universities, alongside a large number of scientific events each year. The account records provision at the level of qualifications rather than of individual courses.
A partnership deliverable sets out a framework for learning design with generative AI, organised in stages from analysis and planning through activity design to development and production. It is the only instrument in the material that addresses curriculum design directly rather than regulating the use of tools inside an existing design.
A teacher reports that questions for a national standardised language examination are now generated using artificial intelligence. The account is brief but marks a national assessment body applying generative tools to item production rather than regulating their use by others.
Germany funds specific projects to expand research and knowledge in AI, among them the AI Campus platform. The partnership records it first as a national practice and later as one of four federal initiatives, with a regional hub bundling transfer and networking in North Rhine-Westphalia.
A government action plan published in January 2025 sets out a roadmap for maximising the potential of AI across the country. Recorded as a change in nationwide practice rather than in policy, it marks the point at which the national conversation shifts from regulation towards deployment.
The conference of Spanish rectors published recommendations covering opportunities, risks, ethics and evaluation. The document is not law but guides the university system, and it explicitly delegates implementation to individual universities while proposing shared principles, competence definitions, internal discussion processes, equity of access and an assessment of ecological cost.
A draft law approved by the government in March 2025 aligns national provision with the EU AI Act. Its appearance in the second round is one of the clearest instances of European regulation propagating into national legislation during the observation period.
Material published in April 2025 by the national agency for education and the ministry gathers legislation and recommendations for 2025, prepared with practitioners, researchers and other stakeholders. The partner records the policies it contains as advisory rather than binding.
A national task force was established and seminars and workshops on generative AI and large language models in education were held. The pairing of a committee with public events is the shape national activity takes where no education-specific framework yet exists.
Guidelines for the confederation were published on 1 July 2025 and a report on AI in education followed a day later. In a partnership where the previous round recorded no national policy at all for this country, the appearance of two instruments within two days marks a distinct change of pace.
Respondents describe a framework in which the European AI Act operates together with recent national legislation of September 2025, ministerial guidelines for schools and a national AI strategy covering 2024 to 2026. The layering of European and national instruments is what respondents themselves report as governing use.
An administrator describes a national framework combining the national AI strategy, an AI in education policy document and action plan for 2025 to 2029, and generative AI guidelines from the research council. The same account states that no fully standardised regulation for generative AI in education exists, leaving universities responsible for their own institutional policies.
The regulation is the single most frequently named instrument across countries in the third round, cited by respondents regardless of where they work. It functions as the common reference layer beneath national frameworks that otherwise differ widely in form and force.
A support staff respondent reports that the European regulation takes effect nationally on 2 August 2026 while its prohibitions already apply. The distinction between staged application and full entry into force is what institutions have to plan against.
The Council of Higher Education published an ethical guide covering the use of generative AI in scientific research and publication in higher education institutions in May 2024. The partner reports that the guidance is advisory, which leaves each university to convert it into binding institutional rules.
Principles established in September 2024 were circulated to all public institutions, universities included. The instrument reaches higher education not through education policy but through public administration law, which is a distinct route into the sector.
A central portal offers generative AI applications to all universities in North Rhine-Westphalia and is accessible to both staff and students. The arrangement answers procurement, access and cost at the level of a state system rather than leaving each university to negotiate separately.
A national pilot introduces an educational version of a commercial assistant with teacher training and phased student access from late 2025. It is one of the few national instruments in the partnership that provides a tool rather than a rule.
The quality assurance agency curates resources aimed at helping the education sector use generative AI while maintaining academic standards, and has published supplementary advice on the tools. The instrument works through the quality assurance route rather than through legislation.
Assessors, reviewers and panellists are explicitly prohibited from using generative AI in their assessment activities, while applicants are advised to exercise caution over bias and data protection. The asymmetry between a prohibition for evaluators and advice for applicants is itself a policy design choice.
A programme launched in October 2024 in collaboration between the Council of Higher Education and the Presidency of Defence Industries channels AI in the defence industry into undergraduate senior projects and graduate theses. It is a national instrument that works through the curriculum rather than through regulation.
The documents the synthesis rests on. The weight lies with three partnership reports and two institutional guides, while the rest were reviewed for corroboration. How each document was read and what it contributed is set out on its own page.
ADMIT (Erasmus+, EADTU), 2024
ADMIT (Erasmus+, EADTU), 2025
ADMIT (Erasmus+, EADTU), 2024
The third round of the ADMIT survey, gathered from eleven partner universities between March and July 2026 with 1,570 respondents across four stakeholder groups. Data were collected through questionnaires alone, and the qualitative material was analysed inductively through content analysis with verbatim quotations retained under pseudonyms.
ADMIT (Erasmus+, EADTU), 2024
ADMIT (Erasmus+, EADTU), 2025
ADMIT (Erasmus+, EADTU), 2025
ADMIT (Erasmus+, EADTU), 2025
ADMIT (Erasmus+, EADTU), 2025
ADMIT (Erasmus+, EADTU), 2025
ADMIT (Erasmus+, EADTU), 2025
A student-facing guide issued by the Open Education Faculty of Anadolu University. It positions generative AI as a draft-producing assistant rather than an authority, teaches prompt construction, maps use to the levels of Bloom's taxonomy, and sets out four binding practices: withholding personal data, verifying critical claims against at least two sources, complying with course instructions, and declaring use.
Anadolu University, Open Education Faculty, 2026
The published programme page of a two-year associate degree built around generative AI, used here as a primary institutional source rather than a partnership report.
Every claim in the synthesis is tied to its source, the page it rests on and the sentence found there. Quotations were located in the source document and compared.
In 2023, the Data Analytics and Artificial Intelligence Research Unit was established. The infrastructure setup of the unit is still ongoing.
The second sentence prevents the unit from being recorded as fully operational; maturity is coded from what the source states rather than from the announcement.
A rule-based AI chatbot is being used for non-academic (administrative) support to answer student questions about enrolment, examinations, graduation, and more.
The sentence carries the distinction that matters for classification: the operational system is rule-based, so it cannot be counted as a deployed generative AI practice.
This is a required document.
The declaration form is the clearest case of a practice that specifies a procedure and therefore transfers without infrastructure.
The Graduate Education Institute has updated its Thesis Writing Guide in 2024, requiring students to sign a 'Generative Artificial Intelligence Usage Declaration' form.
The sentence establishes both the instrument (a declaration form) and its binding character, which is what distinguishes this practice from advisory disclosure guidance.
Experimental environments (for use and research of LLMs, GenAI and Learning Analytics): Feedback Centre,
The label "experimental" in the institution's own wording sets the maturity ceiling for this record.
An institution-wide steering committee has just been installed (oct 2024)
The date and the word "just" fix the practice at its beginning, which is what the maturity coding reflects.
Rector (15.4.2024): Policy on the use of artificial intelligence at the University of Jyväskylä
Naming the rector as the issuing authority is the evidence for institution-wide scope, which the classification depends on.
Students may use Generative AI to support their learning.
The permissive opening is what distinguishes this policy from prohibition-led instruments elsewhere in the partnership.
Primarily use Microsoft's Copilot's AI-based conversation on a university device and while logged in to the Edge browser with university credentials!
The instruction names device, browser and credentials, which is the concrete control mechanism the classification rests on.
Service agreement on the use of AI systems at the
The named service agreement shows that the institution regulates AI use contractually and not only through advisory guidance.
One of the key initiatives is a pilot project titled "Developing robust assessment in the light of Generative AI developments".
The named pilot is the practice itself; the classification as an assessment measure rests on its stated title and purpose.
Personalized learning: Creation of personalized AI-aided learning is mentioned in the plan for the academic year 2025
The wording "is mentioned in the plan" bounds the claim: this is a planned curriculum measure, not a running one.
Virtual assistant: Chatbot called ‘Minerva’ is in use in the university webpages (2023-)
The institution records the assistant as in use and dates it, which establishes the practice.
there is a policy that teachers can decide the level of application of AI in learning activities.
The sentence locates the decision right with the individual teacher, which is what makes this a governance arrangement rather than a teaching practice.
A teacher's guidelines to chatGPT and remote assessments (for teachers)
The title itself couples the tool with remote assessment, which is the basis for classifying the record under assessment rather than teaching practice.
the Vice-rectorate for Educational Innovation has developed a training and awareness programme called ‘Use of Generative Artificial Intelligence’
The named owner inside the university establishes that the programme is an institutional commitment rather than a local initiative.
These guidelines are all advisory.
The position of the national instruments on the left of the axis is taken from this statement rather than from an inference about their weight.
funding of specific projects to expand the level of research and knowledge in
The entry records national funding as the mechanism behind the platform.
artificial intelligence in the defense industry will be addressed as senior projects at the
The mechanism named here, coursework and theses, is what places the programme in the partnership and curriculum domain rather than in governance.
The Turkish Higher Education Council published the 'Ethical Guide on the Use of Generative Artificial Intelligence in Scientific Research and Publication Activities in Higher Education Institutions' in May 2024
The entry establishes the existence, the issuing authority and the date of the instrument.
which has been sent to all public institutions, including universities
Circulation to universities is what makes a public administration instrument relevant to higher education practice.
These resources aim to help the education sector use Generative AI as a positive tool while maintaining academic standards.
The stated purpose ties the instrument to academic standards, which is the basis for classifying it under quality assurance.
Assessors, reviewers, and panellists are explicitly prohibited from using Generative AI tools in their assessment activities.
An explicit prohibition is one of the few binding national measures recorded in the partnership, which is why the record is coded at the highest maturity.
The main aim of this report is to map the awareness, challenges and opportunities of policies and practices among students, teachers, IT support/staff and administrative staff
The stated aim of the source instrument is what bounds this cell. The reports set out to map awareness of policies and practices, so an account of how a course or a programme was redesigned falls outside what the questions ask for, and its absence cannot be read as an absence in the field.
The population was composed by 1642 participants belonging to different categories and specifically: 16 administrative staff, 27 IT/support staff, 1284 students, and 315 teachers
The first round figures in the chart are taken from this sentence rather than from the table, so that each number can be checked against a single statement.
This is a required document.
An institutional act converts guidance into an obligation, which is the meso-level case in the same rule.
This is a required document.
The final step of the sequence is visible in this sentence: an institutional act attaches an obligation to an individual submission.
These guidelines are all advisory.
Advisory status is the reason the record is coded as guidance rather than regulation, and it explains why institutional conversion is required.
Switzerland There aren’t any
The first round entry is the baseline against which the change is read; without it the 2025 instruments would look like continuity rather than onset.
Assessors, reviewers, and panellists are explicitly prohibited from using Generative AI tools in their assessment activities.
This is the only national instrument in the material described as an explicit prohibition, which fixes its position at the binding end of the axis.
There are pilot projects happening to trial institutional tools such as AIDA (Artificial Intelligence Digital Assistant) to support students as a study buddy, and AIMWA (Artificial Intelligence Module Writing Assistant) to support staff
Both tools are named and their audiences distinguished, which is what makes the record usable for the level assignment rule applied to paired instruments.
students consistently reported significantly lower awareness of institutional policies than all other groups
The chart plots the means the report gives for these two years, and this sentence is the report's own reading of the same figures.
students reported significantly lower awareness of national policies than teachers
The national series in the chart reproduces the means the report reports for 2025, and this sentence states the group comparison drawn from them.
All universities in Spain now offer courses related to AI.
A statement about system-wide provision, which is what places the record at the macro level.
Artificial intelligence is applied as part of study modules, and study modules related to artificial intelligence are offered to teachers and students
The sentence records both directions of curriculum change in one clause, which is what the record documents.
Academically, certificate programs, associate degree and undergraduate programs are available at universities
The enumeration of qualification types is what distinguishes this from reports of individual courses.
There is also a mandatory internal Regulation of Learning Analytics to ensure data privacy.
The word "mandatory" separates this instrument from the advisory guidelines listed beside it, which is the point the record turns on.
UOC introduced SofIA, its internal advisory framework and guidelines for integrating AI across educational, administrative, and support services
The framework is named and its scope given, which allows the change from delegated discretion to institutional guidance to be traced across rounds.
A working group has been set up to develop an institution-wide AI policy and guidelines. They will be available in 2026.
The future tense fixes the record as preparation rather than adoption.
It offers a central portal for generative AI applications to all universities in North Rhine-
The sentence establishes the scope of the portal, which is what makes it a system-level rather than an institutional practice.
the university has developed resources such as a Moodle (LMS) course titled “AI in teaching” aimed at teachers and tutors
The named course and its audience establish the practice as staff development rather than a student offering.
The workspace is updated as new material is published.
The commitment to updating is what distinguishes this from a one-off publication and supports the maturity coding.
Establishment of AI committee, pilot tests of AI tools in learning design and IT support, seminars/workshops for staff, students and the general public
The inclusion of the general public among the audiences is the detail that distinguishes this record from internal staff development.
practices have shifted from high-level guidance to concrete applications
The institution's own characterisation of the shift is the evidence for coding the record above the guidance stage.
the COFFEE tool offers personalized AI-driven formative feedback for text assignments using course-specific evaluation criteria and context
The description ties the feedback to criteria set for the individual course, which is what places the practice at the micro level rather than among institutional services.
the COFFEE tool offers personalized AI-driven formative feedback for text assignments using course-specific evaluation criteria and context
The instrument is provided centrally but what it says is set course by course, which is the case the level rule is written to resolve.
authors cannot attribute authorship to AI; AI (or assisted technologies) may only be used to improve language or readability
The journal rule turns a general disclosure principle into an enforceable publication condition, which is what raises this record above advisory guidance.
A module on "Critical AI Literacy for Learning" will be offered to the students from autumn 2025.
The instrument is a module offered to students, which is what distinguishes it from staff-facing training recorded elsewhere.
LIS uses LLMs and Generative AI for AI-based early-warning systems for students, AI-assisted feedback, empathic conversational agents
The enumeration establishes the breadth of the portfolio, which is the basis for reading this as a coordinated institutional practice.
In March 2025, the Spanish government approved a draft law
The word "draft" bounds the maturity claim, and the date fixes the instrument within the observation period.
There are Guidelines on Artificial Intelligence for the Confederation published July 1st, 2025.
The instrument and its date establish the record.
Policies introduced in this material are merely advisory than binding.
The advisory status recorded here supports the wider finding that national instruments in the partnership rarely bind.
Establishment of a National Specialized Committee (task force) on
The establishment of a named body is the practice; the seminars recorded with it show the channel through which it reaches the sector.
Certain recommendations were made, but the implementation of practices was delegated to the universities
Delegation is the mechanism that turns a national recommendation into institutional variation, which is a central finding of the synthesis.
In January 2025 the government published the AI Opportunities Action Plan
The date and the named instrument establish the record; its placement under practices establishes the classification.
Recommend the existence of an equal opportunities program in access to AI, or, alternatively, an access subsidy program for disadvantaged groups.
This is the only sentence in the national documents reviewed here that names a distributive measure, and it names it as a recommendation.
incorporation of AI-related content into curricula, and the establishment of dedicated AI research centers
An independent account naming curriculum incorporation directly, which corroborates the first without repeating it.
There is also a mandatory internal Regulation of Learning Analytics to ensure data privacy.
The word mandatory, used by the institution itself, places this instrument near the binding end while the guidelines beside it stay near the advisory end.
LIS uses LLMs and Generative AI for AI-based early-warning systems for students, AI-assisted feedback, empathic conversational agents
The portfolio is the opposite case: it depends on the data, the platform and the project that carry it.
The third pillar, AI Campus, (https://ki-campus.org/) democratizes access to AI education with its comprehensive online platform
The second round places the same platform among four named federal initiatives, which raises the maturity coding beyond a single funded project.
students relied more heavily on informal learning methods such as peer collaboration and online resources, whereas staff groups had greater access to structured training through workshops, seminars, and institutional programs
The second report measures the difference in provision that the third report describes from the side of the learner.
Participants also emphasized the importance of fact-checking AI-generated content and using it critically rather than copying outputs directly.
The second report records the same practice a round earlier and across all groups, which establishes it as sustained rather than newly reported.
technical and accessibility issues, including unequal access and lack of digital skills, were noted as barriers
The demand side of the same question is documented in every round, which is what makes the absence of an instrument a finding rather than a gap in coverage.
Research Question 2: For what specific educational purposes do the four groups of academic users currently use or plan to use LLMs and GenAI?
The question closest to curriculum asks what respondents use the tools for, not what was changed in a programme. The distinction between purpose of use and revision of design is why this intersection stays empty across three rounds.
The population was composed of 523 participants belonging to different categories and specifically: 24 administrative staff, 27 IT/support staff, 126 teachers, and 346 students
The second round figures come from this sentence, which also shows how much smaller the round is than the two around it.
students relied more heavily on informal learning methods such as peer collaboration and online resources, whereas staff groups had greater access to structured training through workshops, seminars, and institutional programs
The second report measures the asymmetry that the third report describes, which turns a reported habit into a documented difference in provision.
YouTube and social media are the most common informal channel. They play a much larger role for students than for teachers.
The sentence names both the channel and the asymmetry between groups, which is what carries the cell beyond a description of habit.
YouTube and social media are the most common informal channel. They play a much larger role for students than for teachers.
The sentence names both the channel and the asymmetry, which is what allows the record to carry an equity reading rather than only a description of habit.
The most significant change has been that all UOC staff now have access to Gemini Pro.
The account states the scope of access, which is what turns a procurement decision into an equity measure.
EU AI Act is undoubtedly the single most-cited one, regardless of the country
The report characterises the distribution of citations across countries, which supports treating the regulation as the shared macro-level reference.
there is still no fully standardized, detailed regulation specifically for generative AI in education, leaving universities with significant responsibility to develop their own institutional policies
The account states the displacement the diagram describes: the operative rule-making moves from the national level to the institution.
The EU’s AI Regulation will take effect in Finland on August 2, 2026, but the prohibitions have already come into force.
The sentence separates the date of full effect from the prohibitions already applying, which is the planning-relevant distinction.
which combines the European AI Act (EU 2024/1689) with recent national legislation (Law No 132 of 17 September 2025)
The respondent names both layers and their identifiers, which is what allows the record to be verified rather than paraphrased.
there is still no fully standardized, detailed regulation specifically for generative AI in education, leaving universities with significant responsibility to develop their own institutional policies
The stated gap is the finding: national instruments exist, but the operative rule-making is displaced to the institutional level.
I always give information about where and what for I've used GenIA tools. And I never share protected data.
Disclosure is expressed as place and purpose and given together with the data rule, which is how respondents themselves join the two.
No adding of personal data in the prompting that is sensitive for reasons of data protection or that are otherwise confidential.
The rule is stated at the level of the prompt, which is where the micro-level decision is actually taken.
their use is prohibited during exams, while it may be permitted for preparing assignments or activities if transparency is ensured
The sentence places the boundary at the assessment event, which is the design choice the record documents.
I always give information about where and what for I've used GenIA tools. And I never share protected data.
Disclosure is expressed as a statement of place and purpose, and it is given together with the data rule, which is how respondents themselves join the two.
I always double-check academic texts in particular, especially for citations.
The respondent names the object of the check, and the object is the element most often fabricated, which is what makes the practice targeted rather than general.
I always double-check academic texts in particular, especially for citations.
The respondent names what is checked, and it is the element most often fabricated, which makes the practice targeted rather than a general disposition to be careful.
No adding of personal data in the prompting that is sensitive for reasons of data protection or that are otherwise confidential.
The rule is stated at the level of the prompt, which is where the decision belonging to this level is actually taken.
I have assigned a voice to the AI and created several agents specialising in different areas of expertise. I use models that reflect my teaching style.
The account states a configuration rather than a use, which is what distinguishes it from the general pattern of checking outputs and makes it a teaching-practice record.
I've learned how effective prompts work and that the effort involved in explaining to the AI what it should do can sometimes be a learning experience in itself.
The respondent identifies the articulation of a task as the learning, which is the strongest pedagogical claim any participant makes for individual use.
I regularly use them to prepare teaching materials, support literature reviews and design learning activities.
The sentence is the account the earlier rounds lacked, and it is what allows the cell to be read rather than only explained.
I have assigned a voice to the AI and created several agents specialising in different areas of expertise. I use models that reflect my teaching style.
The account states a configuration rather than a use, which is the step the other two records at this intersection do not take.
I regularly use them to prepare teaching materials, support literature reviews and design learning activities.
The sentence names design of learning activities as a routine, which is precisely the account the earlier rounds lacked and the reason the curriculum row can now be read at the micro level.
Instead of having AI do my homework for me, I use it to ask for ideas that can guide me in the right direction
The contrast the student draws is the practice itself: the tool is placed before the work rather than in place of it.
Rather than using the raw output, I always apply a 'human-in-the-loop' approach to ensure the final content meets professional standards.
The quotation states the practice as a routine rather than an aspiration, which is what allows it to be recorded as a micro-level practice.
Questions for standardized exams are now being generated using artificial intelligence
The account places a generative tool inside a national assessment process, which no other record in the partnership does.
Greece has launched pilot initiatives, such as the “AI in Schools” program, introducing ChatGPT Edu with teacher training and phased student access starting in late 2025
The account names the programme, the tool, the training and the access schedule, which is enough to record it as a practice rather than an intention.
informal training acts as a complement not a substitute
The report's own conclusion for the teacher group prevents the record from being read as evidence that formal provision is unnecessary.
Use of Ki-connect.nrw has increased, as has the individual use of various LLMs
Independent corroboration a round later distinguishes continued operation from an announcement.
I always review answers/content generated thoroughly before using it for any materials that will be given to others.
The review is tied to the moment the content leaves the author, which is what makes it a practice rather than an attitude.
temporary chats to avoid storage, use of official AI platforms that are in line with University policies
The two further measures show that the practice is not only a prohibition on content but also a choice of venue and of setting.
The only rule I can identify is to cite the LLMs used and specify their applications for academic work.
A student names the rule where no instrument supplies it, which separates the practice from compliance with a policy.
confidentiality has become a habit, rather than a mere policy
The report's own characterisation is what places the measure here rather than counting it again under the instruments that require it.
I know how to prompt better, I know how to maintain an extended conversation better.
Sustaining an exchange is named as a separate skill from composing a request, which is the distinction the record rests on.
The university uses a general AI chatbot (Minerva).
A respondent naming the same assistant two rounds later corroborates continued operation rather than a one-off announcement.
Coffee (LLM feedback system) and many more LMS plugins that help teachers and learners prepare lessons
An independent account a round later confirms that the tool is in use and not merely described.
the dominant framing is augmentation, not replacement
The report's own characterisation establishes that this is the prevailing pattern rather than an isolated position.
Faster generation of tests and materials. Using the appropriate prompts makes it very useful and quick.
A second teacher describes the same work from the side of what it produces, which establishes the pattern as shared rather than individual.
Faster generation of tests and materials. Using the appropriate prompts makes it very useful and quick.
A second teacher names the products of the same work, which establishes the pattern as shared rather than singular.
The population comprised 1570 participants across different categories: 25 Administratoristrative staff, 26 IT/support staff, 176 teachers, and 1343 students
The third round figures come from this sentence in the working draft, quoted as it stands including its typographical error.
The only rule I can identify is to cite the LLMs used and specify their applications for academic work.
A student states the same rule where no institutional instrument supplies it, which shows the practice is not simply the shadow of a policy.
confidentiality has become a habit, rather than a mere policy
The report's own characterisation is what places the practice at the micro level rather than counting it again under the institutional instruments that require it.
I was mostly using chatgpt earlier. Now, i try different ai tools like gemini, notebooklm etc.
The move from one assistant to several is stated by a student, and the same move is described by staff, which shows the differentiation is not confined to one group.
Analysis and Planning
The first stage named in the framework's own structure, which shows that the instrument organises the design process rather than the use of a tool.
Önemli bilgilerin en az iki güvenilir kaynaktan kontrol edilmesinin
The two-source rule is the operational element that separates this guide from general advice to be careful.
Ders yönergesi izin vermiyorsa beyan yazmak yerine ÜYZ kullanılmaz.
The sentence resolves the case where disclosure and prohibition meet, and resolves it against use rather than against silence.
Temel yaklaşım, teknolojiyi ne “kurtarıcı” ne de “tehdit” olarak görmek
The stated position rules out both advocacy and prohibition, which is what allows the guide to regulate workflows rather than attitudes.
çift yönlü doğrulama, veri minimizasyonu ve kurumsal politika uyumu zorunlu kabul edilmelidir
Naming these three as obligations rather than recommendations is what places the instrument above advisory guidance.
hızla dijitalleşen iş dünyasının ihtiyaç duyduğu, yapay zekâ destekli yazılım geliştirme becerilerine sahip nitelikli insan gücünü yetiştirmeyi amaçlamaktadır
The stated aim is written in terms of AI-assisted development, which is what makes the record a curriculum decision rather than a tool decision.
References follow APA 7 and retain the language of the source.
The report can be downloaded in two formats and two languages, generated from the current content on each request.