Report on trends in policies and practices on the use of LLMs and generative AI in the partnership: Third report (draft)

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.

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.

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.

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.

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 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.

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.

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.