AI Between Trust, Evaluation and Human Judgement

The debate therefore moves away from the simple question of whether to allow or ban AI, as such an answer is too narrow for serious pedagogical practice. For certain tasks, it makes sense for students to use AI, compare it with their own thinking, and learn to verify its answers. For other tasks, the very absence of AI is part of the learning objective, as we wish to see how a student understands the problem, how they read a text, how they construct an argument and how they correct their own mistakes. A growing body of evidence suggests that universities will have to define this distinction much more precisely than we were accustomed to a few years ago.

One of the most talked-about news stories of recent days is the stricter AI usage policy at the University of California, Berkeley School of Law. Under the new policy, Berkeley Law does not permit students to use AI in work submitted for assessment, unless a lecturer explicitly authorises its use. The ban also covers activities that many would otherwise regard as less problematic, such as brainstorming, drafting, editing, translating or correcting grammar. Reuters has also reported on the case. What is interesting about this case is not only that it involves a stricter policy, but above all how clearly its pedagogical rationale is defined. In law studies, it is not enough for a student to submit a fluently written text. They must be able to identify the legal issue themselves, evaluate sources, formulate an argument and take responsibility for their citations. It is precisely for this reason that the case is also relevant outside the field of law. It reminds us that rules on AI cannot be limited to a general statement in the syllabus. Good rules must be linked to the learning objective of each individual assignment.

Another recent debate comes from the field of admissions procedures. Inside Higher Ed reports on the SHAPE AI framework, which recommends that higher education institutions, before using AI in the student admissions process, first ask themselves why they wish to use it at all. This sounds almost self-evident, but in practice it often is not. If AI is used to rank applicants, assess applications or guide students, it is no longer a matter of neutral administrative support. These are processes that can influence a person’s educational journey. This is why a clear purpose, human oversight, bias checks, the documentation of decisions and the very careful handling of data are essential.

Similar warnings emerge from research into automated recruitment. Stanford HAI has presented the findings of an extensive study on AI-based candidate selection tools. The study analysed several million job applications and revealed worrying patterns in the rejection of candidates. Among other things, it points out that overall averages can mask differences between job types; that Black and Asian candidates were, in some cases, less frequently recommended for the next stage of the process; and that the use of the same system by multiple employers can lead to the repeated rejection of the same candidates. This news is also relevant for universities because they prepare students for a labour market in which they will increasingly encounter algorithmic filters. AI literacy is therefore not just a matter of writing seminar papers. Students must also understand how AI affects recruitment, career guidance, skills assessment and access to opportunities.

In the US, graduates’ reactions to speeches about AI at graduation ceremonies have also caused quite a stir. The Guardianreports that some students reacted negatively to speakers who presented artificial intelligence primarily as an opportunity and an almost inevitable part of the future. Such a reaction is not surprising. Graduates are entering a labour market where there is simultaneous talk of increased productivity, the automation of entry-level jobs and the need for new skills. If universities speak of AI solely in terms of enthusiasm, they quickly overlook students’ anxieties. A more convincing approach is one that treats AI with a level head, considering its benefits, limitations and implications for specific career paths.

In the field of research integrity, the decision by the arXiv platform stands out, as it is taking a tougher stance on submissions showing clear signs of unverified use of generative AI. Inside Higher Ed reports that the main problem will be papers containing fabricated references, incorrect citations and text where remnants of the model’s instructions appear in the published work. The journal Nature has also written on the same topic. The measure alone will not solve the problem, as such cases are difficult to detect systematically. Nevertheless, it sends a clear message. The use of AI in research writing is not merely a technical issue, but a matter of trust in the scientific record. For students and researchers, this means that the use of AI does not diminish their responsibility for verifying sources, citations and conclusions.

Western Washington University also provides a good insight into the diversity of university practices. Cascadia Daily Newsdescribes examples of lecturers who respond very differently to generative AI. Some are bringing assignments back to paper, into the classroom and into live discussion, whilst others are teaching students how to create their own AI agents and integrate them meaningfully into their academic work. The university has a few basic rules, including that unauthorised use of generative AI may be treated as a breach of academic integrity; however, decisions on specific uses largely remain with individual courses and lecturers. This example clearly illustrates why general rules are often insufficient. In computer science, marketing, writing or environmental studies, AI can play very different pedagogical roles. For students, however, such inconsistency also creates considerable uncertainty if it is not clearly stated for each course what is permitted, what is not, and the reasoning behind such decisions.

A slightly different perspective is offered by the opinion piece ‘AI Didn’t Break the University. It Revealed What Was Already Broken’, published on the RealClearEducation portal. The author does not view AI merely as a tool that has facilitated cheating, but as a symptom of a deeper problem. The piece points out that students do not always turn to AI simply to avoid work. Sometimes they use it because the academic environment lacks a live interlocutor, immediate feedback, a sense of community or intellectual exchange. This is an uncomfortable but useful reminder for universities. If we want students to use AI thoughtfully, we must simultaneously strengthen those forms of work where thinking takes place with people, through conversation, questioning, objection and feedback. AI can fill a gap, but it cannot replace high-quality pedagogical presence.

Stanford HAI has also published an article on the impact of AI on scientific discovery. Examples range from the design of new antibodies to faster simulations and the analysis of large datasets. However, the researchers emphasise that humans must remain at the centre – defining research questions, evaluating results and understanding the significance of findings. The article ‘How AI Is Transforming Scientific Discovery While Keeping Humans at the Centre’ is also a good starting point for teaching. It is not enough to teach students how to use AI to arrive at an answer more quickly. They must be taught when an answer makes sense at all, how it was arrived at, and what else needs to be checked.

There is another piece of news from Stanford that is more technical but of interest to universities. Researchers have presented an approach to predicting how large language models will behave at a larger scale, drawing on ideas from measurement and educational sciences. The article New Approach to Scaling Laws Could Change How AI Models Are Trained is interesting because it shows that the issue of AI infrastructure is not merely a matter of purchasing tools. Universities wishing to develop or evaluate AI systems will also need to understand measurement, costs, model limitations and the conditions under which comparisons between models are even meaningful.

The educational debate is increasingly turning back to assessment. Times Higher Education points out that generative AI has exposed the weakness of assignments in which we assess only the final product. If a student submits a well-structured essay, this does not necessarily mean that they know how to select a good source, recognise a weak argument or independently link theory to a case study. Consequently, there is increasing talk of process-based assessment. This includes shorter, ongoing written records, oral presentations, comparisons between different sources, explanations of one’s own decisions, classroom work, and assignments where students must demonstrate the path to a solution, not just the solution itself. This does not mean that we should revert solely to closed-book exams and prohibitions. It is more useful to consider the rhythm of the course. In certain phases, students should work without the course materials, as they are developing a basic understanding. At other stages, they can use it for comparison, feedback or to explore alternative interpretations. For more challenging tasks, they can use it as a subject for critical evaluation. Such an approach requires more work in preparing assignments, but it gives students a clearer message about what they are actually learning from the task.

Another highly topical issue is the choice of tools. Times Higher Education advises that teachers should first check which tools have already been approved at institutional level and what the conditions for their use are. With AI tools, it is not enough for them simply to provide useful answers. It is necessary to check where the data is stored, whether the inputs are used to train the model, what the options are for deletion, how access to student data is managed, and whether the tool complies with institutional rules. This is particularly important when working with research data, students’ personal data, unpublished materials and copyright-protected content.

There is a growing realisation across the wider education sector that many teachers are already using AI or encountering it in their pupils’ work, yet they lack sufficiently clear guidance. Axios summarises the findings of a study by Gallup and the Walton Family Foundation, which highlight precisely this disconnect. AI is already present in classrooms, but institutional support often lags behind. This is a valuable lesson for universities too. If the rules and training are not clear enough, the burden of judgement falls on individual educators. They then become not just teachers, but also interpreters of the rules, verifiers of authorship, advisors on tools and guardians of academic integrity.

At European level, the European Commission has launched a consultation on draft guidelines for the classification of high-risk AI systems. The consultation is open until 23 June 2026. The guidelines are intended to help providers, users, public authorities, research institutions and other stakeholders assess whether a particular AI system falls within the scope of high-risk systems under the AI Act. The Commission has also published draft guidelines which explain in more detail the application of Article 6 of the AI Act. For higher education, this is a very direct issue. If a lecturer uses AI to generate ideas for exercise set, the risk is different from that when an institution uses AI to rank candidates, assess performance, monitor progress or decide on support for students. The guidelines will therefore also be important for university departments introducing digital tools into teaching, research and administrative processes. For each such tool, it will be necessary to be able to describe its purpose, the data involved, the users, the potential consequences and the role of humans in decision-making.

An interesting regional example comes from Croatia. Netokracija reports on the BrAIn project, which is developing an AI curriculum, virtual learning assistants and activities to foster critical thinking. The Croatian approach is not based on the assumption that young people can simply be shielded from generative AI. It stems from the far more realistic observation that many are already using it, so it is better to teach them how to understand it, verify it and use it responsibly. This is also of interest to universities in Slovenia because future generations of students will arrive with varying levels of prior knowledge. Some will use AI very skilfully, others mainly intuitively, and others with almost no understanding of what is happening behind the scenes. More information about the project is also available on the official BrAIn website.

News of the establishment of the National Council for Ethics in Artificial Intelligence has been widely reported in Slovenia. The Council will act as an independent expert body on ethical issues and the responsible use of AI. Its role will be to prepare opinions, recommendations and guidelines, and to provide advice on issues concerning the social, democratic, ethical and human rights aspects of AI. This is relevant to the university sector because, in higher education, issues that do not fall within a single discipline very quickly arise. In the case of AI, this involves pedagogy, research integrity, data protection, copyright, accessibility, bias, public trust and organisational accountability. Universities will need precisely this interweaving of expert perspectives when formulating their own rules.

The Ministry of Education has published an episode of the podcast ‘Mission: Quality’ entitled How artificial intelligence is changing learning and teaching. The focus is on specific examples of AI’s use in preparing teaching materials, differentiating instruction and monitoring pupils’ progress, whilst also highlighting concerns about superficial learning, the authenticity of work and dependence on digital tools. Although this relates to the school setting, the issues are very similar to those in higher education. Indeed, even at university, it is all too easy for a well-designed piece of work to mask a poor understanding.

Training on authorship, intellectual property and responsibility in the use of AI is also continuing in the public sector. The Administrative Academy’s programme Authorship, Intellectual Property and Responsibility in the Use of Artificial Intelligence addresses questions such as who is the author when content is generated by AI, who is liable for errors, how to handle copyright-protected works, and how to attribute contributions made by humans and by tools. These are very specific issues that we encounter when preparing materials, reports, online texts, research summaries and student work.

The fifth national conference, Artificial Intelligence: From Vision to Trust, has also been announced. The theme of the conference is the transition from discussions about the potential of AI to its responsible implementation in the economy, the public sector and wider society. This, too, clearly demonstrates that Slovenia is moving from a phase of general enthusiasm or general caution towards a more practical phase. The question is no longer whether AI will be present, but how we will implement it, who will be responsible for it and how we will assess its effects.

Recent developments suggest that universities cannot rely on a single solution when it comes to AI. Bans may be appropriate when the aim is to protect core teaching activities. In other cases, a ban would be a poorer choice, as it would leave students without guided experience of the tools they will encounter in their studies, work and research. The hardest part, therefore, is not deciding for or against AI, but drawing up a precise description of what we are learning in a particular task and what role the tool may play in this.

For educators, this means more careful planning. For institutions, it means clearer rules, support and verification of tools. For students, it means greater responsibility, as the use of AI cannot be an excuse for unverified sources, incorrect citations or a lack of understanding of the work submitted. AI can assist with learning, but it cannot develop a student’s judgement on their behalf. It is precisely this distinction that is likely to be one of the central challenges for university teaching in the coming years.

If you have explored an interesting use of AI in education at your faculty, within a course or as part of your professional work, we warmly invite you to share your example with us via the form for collecting examples of good practice. We will include selected contributions in future issues of the UI newsletter, thereby working together to build an overview of well-considered approaches within the university sector.

Authors

Eva Kern Nanut, Sanja Jedrinović Čufer, Mateja Bevčič, Eva Škraba, Maja Kosmač, University of Ljubljana Centre for the Use of ICT in the Teaching Process