From the classroom to the laboratory: AI takes on new roles

A survey of more than 6,000 students at seven British universities, presented in the article Majority of students deny using AI when forbidden, study finds, found that almost a third of those surveyed do not use AI at all in their studies. Among AI users, 67 per cent stated that they would not use it for an assessed assignment if it were prohibited, whilst 5.2 per cent of regular AI users use it frequently or always, despite the ban. It is used more frequently by international students and students for whom English is not their first language, and the authors emphasise the need for clearer and more harmonised rules.

The study Using AI-based Learning Assistants in Higher Education: A Large-Scale Descriptive Analysis is based on usage logs of the Syntea AI-based learning assistant amongst students undertaking distance learning at the German university IU International University of Applied Sciences. In February 2025, 44,035 out of 76,485 (58 per cent) of the students analysed used it, with usage being highest amongst younger students and slightly lower in certain fields of study involving more practical or visual work.

An analysis of more than 15,000 authentic student interactions with generative AI, which took place during voluntary use in the context of study commitments, showed that students employ a number of characteristic, recurring patterns of use when working with AI, such as explaining concepts, assistance with completing assignments, improving their own writing, or checking answers. These vary across subjects depending on the nature of the academic work; therefore, the authors advocate the development of instructions and guidelines tailored to individual subjects and types of assignments, rather than a one-size-fits-all set of rules for all students.

Less obvious may be the impact of AI on group work. Students use it to summarise the contributions of other group members, clarify ambiguities and refine shared ideas; in doing so, they may omit the discussions in which they would otherwise have to explain, justify and reconcile their views with one another. The author therefore suggests that, in group assignments, greater attention should be paid not only to the final product but also to the manner of collaboration.

The report New research reveals the variability of policies, practices and student experience in the age of AI highlights that the rules and practices governing the use of AI vary not only between universities, but also between programmes, modules and individual lecturers within the same institution. Such inconsistency causes confusion amongst students and staff, affects the perception of fairness in assessment, and fuels debate about what grades and degrees actually certify anymore. The QAA therefore recommends more training, the early involvement of students in the drafting of regulations, and greater harmonisation of the student experience, whilst allowing for reasonable differences between academic disciplines.

One of the proposals for the responsible and systematic introduction of AI at universities is the ABC framework, which links the responsible management of AI, its thoughtful integration into the institution’s operations, and the creative transformation of teaching and research. General principles alone are not sufficient: it must be determined who decides on the use of AI in assessment, student admissions, recruitment, research, the procurement of tools and the publication of institutional content, and how individuals can challenge such decisions. The management of IT therefore cannot remain solely the responsibility of IT departments, but requires the involvement of management, academic councils, teaching staff, specialist services and students.

When using AI in education, it is important whether the tool provides the student with a definitive answer or guides them through the thought process. The Socratic Mind tool, presented in the study Scaling Socratic Dialogue with Generative AI, was designed for an undergraduate computer science course to encourage students, through questions, to explain their own reasoning, challenge misconceptions and monitor their understanding.

An alternative option is offered by data comics created using generative AI. In a study involving 60 university students, these proved more effective than conventional visualisations in data comprehension tasks, and students also rated them as more understandable and engaging. The authors also highlight the potential for misinformation, as well as issues of authorship and ownership of materials created in this way.

The article Inside a University’s ‘AI Kitchen’ presents a somewhat broader approach to developing practical skills for working with AI. At Santa Clara University, students, lecturers, academic staff and guests from industry meet weekly in workshops where, without placing much emphasis on programming, they try out new tools and discuss their application in various fields. The aim of these sessions is not to uncritically promote the use of AI, but to create a space in which participants can collectively explore its possibilities, limitations and implications.

At Stanford, researchers have developed Biomni, a virtual research partner, which integrates language models with bioinformatics tools, databases and software for biomedical analysis. A researcher can upload data and pose a research question, and the system will draw up a research plan, carry out an analysis or design a laboratory protocol, whilst the entire process can be monitored and guided as required. Biomni was used by more than 15,000 scientists in its first nine months, although the developers acknowledge that it still struggles with tasks requiring in-depth biological understanding or clinical judgement.

Large language models could also assist in the design of social science research. In the study Large language models can predict the results of social science experiments, they predicted the results of 70 pre-registered US survey experiments involving more than 119,000 participants. Their predictions were comparable to the collective predictions of a group of researchers, but the models generally overestimated the size of the effects. The authors therefore believe that it would make sense to use them to support the piloting of research and the selection of promising research interventions, rather than as a substitute for empirical research involving real participants.

The development of new tools is also having an impact on scientific writing. *Nature* reports on an academic tool for ‘humanising’ texts generated by AI, which adapts their style and removes linguistic features that might reveal the use of AI in articles or project applications. Whilst some see this as a means of improving expression, others warn that it may blur the line between language editing and concealing the use of AI in the creation of a text.

With faster analyses and the automatic generation of hypotheses, however, the question remains as to what happens to the very experience of research – to curiosity, creativity and the joy of discovery. In the commentary As AI transforms science, don’t lose the joy of discovery, the author points out that scientific discoveries often arise through wrong turns, shifts in perspective and unexpected connections, not merely through the efficient testing of hypotheses. Universities should therefore train students to use AI, whilst at the same time maintaining inquiry-based learning, in which they ask their own questions, critically evaluate findings and cultivate their curiosity.

As part of the PoVeJMo project, the GaMS large language model was developed, designed to improve the understanding and generation of texts in Slovenian. The model will be freely available, and its adaptations already support use in medicine, industry, museums and IT infrastructure. The development of domestic models is important not only for providing better support for the Slovenian language, but also for reducing dependence on systems developed primarily for larger languages and different cultural contexts.

The question of what knowledge actually appears in the responses of large language models was examined by the authors of the article What do LLMs say about development in the Global South?. ChatGPT, Claude, Grok and Copilot mostly explained the development of African countries using the prevailing indicators of economic growth, infrastructure and human capital, whilst colonial exploitation, global inequalities and the contributions of African thinkers remained in the background. When the authors asked the models directly about these thinkers, they produced high-quality responses. This suggests that the knowledge is not necessarily absent, but is not automatically activated in response to general questions, as large language models often reflect the dominant views and development paradigms of the Global North.

The report by the United Nations’ independent scientific panel also highlights the wider implications of such differences. The development of the most powerful models, computing infrastructure and control over data are concentrated in a small number of countries and companies, whilst a large part of the world lacks equal access to the internet, powerful tools and support for local languages. Universities must therefore, in addition to teaching the use of AI, also foster critical thinking: which perspectives are represented in the responses, which are missing, and who is actually involved in developing systems that are increasingly influencing the creation and dissemination of knowledge.

Research summarised in the article AI altering meaning of users’ drafts on issues from abortion to climate, study finds has shown that language models can alter the message of a text when rewriting it. This happened even when the models were instructed to preserve the original meaning, and users might not even notice the changes due to the fluid and persuasive style. When using AI for language editing, it is therefore not enough to check only the grammar and tone, but also whether the text still expresses the same viewpoint.

A similar problem arises with automatic review summaries on Tripadvisor. An investigation by the British consumer organisation Which? revealed that, in some cases, AI summaries had toned down or omitted serious guest complaints and highlighted more positive aspects instead. When a platform uses AI to summarise large amounts of content, it decides not only what to omit, but also which information to prioritise.

The use of AI in facial recognition has even more far-reaching consequences. The Facewatch system, used by British shops, is designed to alert the police within seconds upon detecting a person on a private watchlist. The article Alarm over launch of facial recognition in UK shops that instantly alerts police highlights the possibility of false positives, a disproportionate invasion of privacy, and the question of who decides whether to place an individual on such a list and what recourse the individual has to appeal.

Trust is also becoming crucial in the context of agent-based AI, which no longer merely makes suggestions but can carry out actions on the user’s behalf. The article Work on agent-based AI and shopping presents the first European transactions in which AI agents searched for a product, selected an offer and arranged the purchase, whilst the user explicitly confirmed the payment. Such use raises new questions regarding liability in the event of an incorrect purchase, data protection, the transparency of recommendations, and the limits of the authorisation granted by the user to the agent.

The development of AI is also bringing about significant changes in healthcare. A new blood test, developed for the British healthcare system, uses machine learning to analyse around 30 blood markers and assess the risk of gynaecological cancers. The study, presented in the article Thousands of women could be spared painful cancer exam by new NHS AI blood test, involved more than 16,000 female patients. The test could spare low-risk women certain invasive examinations, whilst directing those at higher risk more quickly towards further diagnostic tests; however, its effectiveness will need to be confirmed through wider practical application.

Another example is adaptive brain stimulation for Parkinson’s disease, in which an AI system adjusts electrical brain stimulation in real time according to the patient’s current condition. In four participants, the researchers observed improvements in gait, stability and a reduction in falls. The results are promising, but this was a very small initial trial, so larger clinical studies will be needed.

At this year’s FIFA World Cup, national teams were given access to the FIFA AI Pro tool, which enables coaches and analysts to carry out natural-language queries, review official data and video footage, and perform three-dimensional analysis of match situations. The article gives an example of how the England national team can now analyse their opponents’ penalty takers – a task that previously took around five days – in just a few hours. However, equal access to the tool does not necessarily mean equal opportunities, as wealthier football associations have more analysts and their own data systems. In football, too, it is not just technology that brings an advantage, but above all the ability to transform data into better decisions.

If you have tried out an interesting use of AI in education at your faculty, as part of a course or within the context of your professional work, we warmly invite you to share your example with us via the form for collecting examples of good practice. We look forward to reviewing the submissions and including selected contributions in future issues of the AI newsletter, so that together we can build an overview of useful and well-considered approaches within the university environment.

Authors: Maja Kosmač, Sanja Jedrinović Čufer, Mateja Bevčič, Eva Kern Nanut, Eva Škraba, University of Ljubljana Center for the use of ICT in pedagogical process