AI can help. But do we know how to use that help properly?

Recent research shows that when it comes to using AI in academic study, what matters most is what students use it for and how its use is integrated into the design of learning activities. This is shown by a study in an undergraduate physics course, in which researchers monitored the use of generative AI across three consecutive assignments, during which the rules of its permitted use varied. For all 36 students, AI usage varied depending on the task design: in the final task, 66.7% of the reported usage patterns were related to obtaining feedback and verification. While AI users achieved slightly higher scores, the differences were not statistically significant. The authors emphasize that designing assignments that require students to demonstrate understanding, critical judgment, and responsibility for the final product is more important than simply allowing the use of the IA.

A University of Maryland study shows that access to a dedicated AI tool does not in itself guarantee better academic performance. A total of 2,379 undergraduate students participated in the study, but only about 15% of those who had access to the virtual study assistant actually used it. Students with access to the assistant achieved slightly lower final grades on average and were less active in the online learning environment; however, the low usage rate and the possibility of using other AI tools make it difficult to interpret the results. Among the recorded ways the assistant was used, only 0.7% were intended to obtain feedback on one’s own work.

It is therefore also important how AI support is integrated into the learning process. In a quasi-experimental study in a biochemistry course, traditional lectures, problem-based learning (without the use of AI), and problem-based learning that combined AI support with elements of cognitive apprenticeship were compared. Students in the AI group achieved better results on assessments of their knowledge of lipid metabolism and clinical reasoning; compared to the problem-based learning group without AI, they also reported lower perceived cognitive load and greater confidence in their own abilities. Since the use of AI and additional pedagogical structuring were not tested separately, the differences cannot be attributed solely to AI.

The way AI is used is not only related to the design of the task but also to how students experience their academic work. The Help-seeking 2.0 study, conducted among 389 undergraduate students in science, engineering, and mathematics programs, distinguished between seeking help from AI to better understand the material and using AI as a shortcut to complete assignments. Students who perceived academic demands as a greater burden in terms of time, effort, or psychological stress were less likely to use digital tools for in-depth learning and more likely to seek quick solutions.

The use of AI in studying also raises the question of what distinguishes high-quality student work when its use is permitted. In a study of a qualitative research methods course, 33 research papers were analyzed that were written with unrestricted access to generative AI. In describing research methods, the students’ papers mostly received high grades; however, significantly greater differences emerged in substantiating claims with evidence and in linking theory to the interpretation of findings. The study thus suggests that, when evaluating student work, it makes sense to pay closer attention to precisely those aspects of the work that require the student’s independent professional judgment.

A study among future teachers also highlights the importance of critically reviewing and improving AI-generated results. In this study, 67 education students used generative AI to develop lesson plans for two different tasks. For higher-quality products, students more often first planned their work, then used AI to create content, critically evaluated it, revised it, and reviewed it again. For lower-quality products, however, they more often repeated requests for additional clarifications and revisions without making clear progress toward the goal. Since the researchers grouped participants based on the quality of their final products, the study does not prove that a specific collaboration method led to better results. However, it suggests that what matters most is how students guide their work with the AI and evaluate its suggestions.

Immediate assistance can also have less obvious consequences. In three experiments involving a total of 1,222 participants, researchers examined what happens when people first solve tasks with the help of an AI and then have to continue without it. As reported by UC Berkeley, AI users were initially more successful when the AI was present, but after it was removed, they gave up sooner on challenging math problems and performed worse on reading comprehension tasks. These findings highlight the importance of tackling challenging tasks independently and raise the question of how to balance the use of AI with opportunities to develop perseverance.

The studies discussed thus emphasize that when integrating AI into learning, it is not only its accessibility that matters, but above all whether it continues to encourage students to think independently and exercise critical judgment. Similarly, the article AI literacy is a habit, not a bolt-on emphasizes that AI literacy must become an integral part of learning tasks and assessment. For example, a student can use AI to prepare a first draft, then verify sources, identify errors, evaluate the result against professional criteria, and improve it. Therefore, assessment should not focus solely on the final product, but also on the student’s ability to verify information, justify decisions, improve the text, and transparently document the use of AI.

The integration of AI into higher education places new demands not only on students but also on teachers. A survey of 486 higher education instructors from five public universities in China showed that higher levels of technostress associated with the use of AI were linked to a weaker sense of professional identity among instructors, partly through reduced confidence in their own teaching abilities. Among teachers who perceived greater support from their institution, the link between technostress and confidence in their teaching abilities was weaker. Since this is a cross-sectional study with a convenience sample, the findings indicate associations, not causation.

A different perspective is offered by a two-year study on the integration of generative AI into the teaching of biostatistics in undergraduate health sciences programs. In addition to teachers’ reflections, the researchers analyzed 4,697 messages exchanged among team members during the planning and implementation of the course and tracked changes in their collaboration. Teachers gradually shifted from the role of knowledge transmitters to that of facilitators of critical thinking and collaboration between humans and AI. The authors therefore emphasize the importance of systematic interdisciplinary collaboration and professional development for teachers when introducing AI.

Alongside these changes in the teacher’s role, the question of human relationships—upon which education is founded—remains crucial. In his commentary In the AI era, HE must still strengthen human connection, Fredrick Muyia Nafukho highlights the African philosophy of Ubuntu, according to which the use of AI should also be evaluated in terms of its impact on the community, mutual care, and human dignity. For example, AI can answer routine questions, but it should not replace conversations with teachers, collaboration among students, or mentoring. The author therefore proposes learning activities such as group research, peer assessment, oral presentations, and community-based projects.

AI capabilities are also increasingly approaching tasks that have traditionally been at the core of research work. OpenAI has published an extensive set of new mathematical results generated by its internal model, along with formalizations of numerous proofs in the Lean language, which enables their computational verification. Such achievements also raise questions about how researchers can understand, verify, and evaluate the results generated by AI, as well as how to ensure fairer access to the most powerful models. OpenAI is collaborating with an independent advisory group on mathematics and AI to shape its approach to publishing results and has announced support for professional conferences dedicated to understanding them.

These developments also raise questions regarding the training of future researchers. Sasha Rakhlin of MIT, in his article What is the best path forward for AI in academia?, points out that calculations, programming, failed experiments, and minor discoveries are not merely routine work but part of the process through which doctoral students develop intuition and professional judgment. If AI takes over these tasks, it is therefore necessary to distinguish between unnecessary effort and activities that are important for the development of research expertise. In his view, students should be able to formulate research problems, verify model results, replicate findings, and justify their decisions.

An example of how AI can complement research work without replacing existing expert knowledge comes from the University of Oxford. Researchers have developed an approach in which known physical laws are embedded directly into the generative model. When simulating the motion of charged particles through a turbulent magnetic field, the model was thus able to focus on the more difficult-to-predict aspects of the system, rather than having to learn already known physical laws from the data. This approach reduced computational requirements and sped up simulations, and the results demonstrate how incorporating physical knowledge can improve the effectiveness of AI models.

AI is no longer limited to research applications; its potential use in decision-making regarding research funding is also being explored. In a study on the evaluation of research proposals, the assessments of six large language models were compared with those of expert reviewers for 2,267 proposals submitted to British research calls. The models’ evaluations showed only limited agreement with the expert evaluations; therefore, the authors do not propose AI as a replacement for human judgment, but rather as a potential aid in the initial screening of proposals, checking for bias, or identifying proposals that require more detailed review.

However, the practical application of such an approach demonstrates why technical feasibility alone is not enough. The British research network CRANE used AI in a call for proposals for cybersecurity research to conduct an initial assessment of 179 applications, with nearly half being rejected without human review. Following criticism from researchers, the network apologized for the lack of transparency in the process and announced a re-evaluation of all rejected applications. This example highlights that the use of AI in funding decisions raises questions of transparency, accountability, and trust.

The increasingly widespread use of AI is also changing the scope of scientific output and increasing pressure on scientific publishing systems. Due to the rapid growth in the number of submissions, the arXiv repository has introduced a limit of two new submissions per person per month. In September 2026, it received 40,363 submissions—more than four times as many as in September 2016—and the number of submissions has nearly doubled over the past two years. arXiv also attributes this change to the increasingly widespread use of AI in the preparation of scientific texts, which is placing a greater burden on its team and volunteer moderators. Some researchers support the limit as an incentive for quality, while others point out that limiting quantity does not necessarily mean recognizing scientific value.

The way an organization communicates the purpose of AI to its employees can significantly influence their willingness to use it. Stanford researchers, in a study of the implementation of generative AI at a law firm, an advertising agency, and an IT services company, found that employees were less inclined to use it when it was presented primarily as a tool for increasing productivity. When managers presented it as an opportunity to reduce routine work and take on more substantive tasks—and gave employees time to experiment with it—they were more willing to adopt it. The authors also point out that performance metrics must be adjusted in line with changes in work, as completing more tasks does not necessarily mean greater value.

A completely different example comes from the British company Co-op Legal Services, where the system records employees’ conversations with clients regarding sensitive legal matters, such as wills and probate proceedings, and analyzes them using AI. The system, which uses a model from OpenAI, evaluates each conversation based on more than 50 criteria, and the results are also used to monitor employee performance. The company presents this practice as a way to improve service quality, but some employees and union representatives point to a sense of constant surveillance, stress, and the question of how much professional judgment should be left to automated systems.

It is precisely these risks that new legislation in California addresses. Among other things, the new laws prohibit employers from making termination decisions based solely on AI, restrict the use of systems designed to assess employees’ emotional states, and limit the collection of data from brain and nerve signals. Companies must notify employees when a termination is related to the use of AI, and AI monitoring in restrooms is specifically prohibited. These measures reflect lawmakers’ growing concern about the impact of AI on privacy, surveillance, and employee rights.

Large language models can provide their answers with greater confidence than their actual accuracy warrants. In a UC Berkeley Haas study, researchers tested 11 leading models from OpenAI, Google, Anthropic, and Meta. On average, the models expressed 88% confidence, but answered correctly in only 79% of cases. For more challenging questions, they overestimated their performance more often, while for easier ones, they were sometimes even too unconfident. Newer models with more advanced reasoning were better at assessing their own uncertainty, but still not entirely reliably.

Caution is also needed when interpreting AI behavior. Researcher Anna Hedström of ETH Zurich points out that we often describe AI behavior as if it had human intentions. For example, when we say that a model is misleading or trying to avoid being shut down, we may create the impression that it is doing so with a specific goal in mind, even though we do not have sufficient evidence for such a conclusion. Researchers therefore emphasize the need to distinguish between what the model actually does, what the consequences of its behavior are, and what mechanisms caused it. Such a distinction allows for a more accurate assessment of risks, without exaggerating or underestimating them.

AI can also help researchers in cases where access to data itself is the biggest obstacle. A team from the University of California, Berkeley, used AI to collect local regulations from more than 9,000 U.S. municipalities and converted nearly 10,000 documents—approximately seven million pages—into a machine-readable format. The result is a publicly accessible database containing millions of local regulations from all 50 U.S. states, which allows researchers, journalists, policymakers, and the general public to search for and compare rules that were previously scattered across numerous sources. The database opens up new possibilities for exploring differences between local regulations and their effects, and in the future, it could also serve as the basis for specialized search engines and chatbots—for example, to facilitate a better understanding of building codes.

From November 3 to 5, 2026, the Rog Center in Ljubljana will host AI WEEK 2026, organized by the AI-D Institute, which is dedicated to discussing the development, use, and regulation of artificial intelligence. The program will bring together researchers and representatives from the public sector, businesses, and other organizations. Participants include the Jožef Stefan Institute, the Slovenian Competence Center for Artificial Intelligence, AKOS, Urban Municipality of Ljubljana, and the European Commission Representation in Slovenia. Each day will be dedicated to strategic and regulatory issues, the business applications of AI, and practical examples and presentations of solutions.

If you have explored an interesting use of AI in education at your college, in a course, or as part of your professional work, we cordially invite you to share your example with us via the form for collecting examples of best practices. We look forward to reviewing the submissions and including selected contributions in future issues of the AI newsletter so that, together, we can build a compendium of useful and thoughtful approaches in the university setting.

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

Generative artificial intelligence (ChatGPT) was used in the preparation of this AI newsletter to assist with searching for and reviewing publicly available sources, verifying publication dates, identifying common themes, and preparing initial drafts of summaries. The final selection of sources, assessment of their relevance, content highlights and adaptations, editorial decisions, and responsibility for the published text remain with the authors.