When I think about artificial intelligence at university, it brings first the work that fills an ordinary academic day to my mind. Preparing a class, reading a scientific article, carrying out an analysis, working through a set of student assignments. My work in language and translation studies in academia and in the private sector gives me several reasons to be interested in AI, and some reasons to be cautious.
Research and teaching at university both involve time-consuming tasks for which assistance would be welcome. They also require judgements that can be difficult to explain, even to ourselves, and skills that take practice to develop. This is why we might wonder how AI might fit into academic work when the effort it can spare us is sometimes also the means through which we learn or arrive at an insight. This question invites us to consider the purpose of a particular task, the contribution we expect from the person doing it, and their ability to assess the help they receive.
There is understandable unease about bringing AI into academic activities. A university is a place where people learn to formulate their own questions and explain how they arrived at an answer. It certainly matters who has done the thinking. Yet I also find it difficult to see why every use of AI should cast doubt on the value of the work. Looking closely at what someone has actually used it for seems a more useful starting point. With that in mind it seems vital for universities to make at least some room for tools that can support academic work, while taking seriously a question that follows us from research into the classroom: where does assistance begin to replace the contribution we want a person to make?
I think the boundary depends on what the task is meant to achieve. Assistance should leave room for the understanding or skill that the task requires a person to develop or demonstrate. Anyone presenting the work as their own should be able to understand it and justify the decisions behind it. That seems to me to be the bare minimum. That understanding is necessary, but it cannot stand in for work the task requires them to carry out themselves. You can’t simply let AI do all your work for you.
Suppose you are a researcher whose research project or didactic activities involves comparing a source text with its translations, looking for particular translation techniques or analysing other aspects of the translations. The texts are long, spread across several files, perhaps available only as scans. There are passages to find, examples to record and earlier observations to check. Missing a passage in which a translator departs from an otherwise recurring pattern could make that pattern seem more consistent than it is. I can therefore see why some help with locating material for closer examination would be welcome in this situation. I imagine AI helping to locate relevant passages or organise material for closer examination or challenge patterns the researcher thought they had identified. The researcher would formulate the question, establish the categories, specify the kinds of passages to be look for and decide which examples should be analyzed. A suggested classification would need to remain connected to its source, so that it could be checked in context. Deciding what to do with it is part of the interpretative work. I would thus welcome assistance that leaves more time for that decision, provided the person presenting the interpretation has made the judgement themselves and has the knowledge to justify it.
In my opinion this is also relevant to a research paper, a set of teaching materials or a student’s essay prepared with AI. Saying that someone used AI tells us very little about their involvement. They may have developed the argument and carefully assessed suggestions about its wording. They may also have accepted an argument they cannot explain. I would want to know what they contributed and how they went about it, what they questioned and what they checked. Putting one’s name on a text entails being able to discuss its ideas and account for the decisions that shaped it. Yet a student could receive a complete argument from AI and then learn to explain it convincingly. That would show understanding of the result, but would not establish that the student had done the work of constructing the argument. If that is what the assignment assesses, explaining it afterwards cannot substitute for the student’s own work of selecting the evidence, weighing alternatives and developing the reasoning themselves. Without that work, they haven’t completed the assignment independently.
The ability to check delegated work also depends on competence in the particular task. This seems crucial when using AI for academic purposes. An experienced translation scholar may recognise a misleading interpretation of a passage while remaining a beginner in statistical analysis, whereas a student with relevant training may be better placed to check the latter. The boundary therefore depends on both what someone is expected to learn and what they are equipped to assess. Academic status alone settles neither question. This makes me cautious about treating every reduction in effort as a gain. Time spent comparing possible interpretations or trying to explain an idea can be a necessary part of coming to understand it.
The difference between completing a task and learning from it becomes especially clear in an interesting experiment in school mathematics conducted by Bastani et al. [1]. Students with access to either of two AI tools did better during practice than those working without them. But when the tools were removed, those who had used an interface that readily supplied solutions performed worse on the subsequent exam than students who had practised without AI. A tutor designed to give guidance largely avoided that harm, without producing a significant exam advantage. The practice scores showed what students could accomplish with assistance, and the exam tested what they could subsequently do on their own. Those are different achievements, and the same distinction matters when we consider language assistance. An assignment written in English may have been developed by someone who thinks and reads in several languages. The effort needed to express an idea clearly in English is an additional part of their work. A student working through an awkward sentence may be practising precisely the skill that the assignment is meant to develop. If that is the case, when AI completes the difficult part, the finished work may look better but it tells us little about what the student has learned and what they actually did on their own. This is one of the reasons I understand the concern among lecturers and some of their skepticism about incorporating AI into teaching and learning.
That brings to mind a different question, which Kestin et al. explored in their experiment: could a carefully designed AI tutor teach university physics more effectively than an active classroom lesson? Across two lessons, students achieved better results on tests taken just after learning with the tutor than after classroom instruction [2]. This offers evidence of learning following a particular form of AI teaching, rather than simply better performance while a tool supplies help. It does not establish how long those learning gains would last. Read together, these studies make me want to ask what an improved score represents: success with assistance, understanding demonstrated after a lesson, or knowledge that remains available later. The answer matters when we decide whether a tool serves the purpose of an assignment.
If acceptable assistance depends on what students are meant to learn, its limits need to be explained for the particular assignment. Their use of AI can seem like an open secret, acknowledged indirectly while the boundaries remain uncertain. A student may know that submitting a generated essay as their own work is unacceptable, yet be unsure about asking for an explanation of a concept or feedback on an outline. Those are questions we should be able to discuss before the work is submitted.
I also wonder how much remains unspoken because people are unsure how an admission of AI use will be received. There is no doubt that any undisclosed use is a dishonest practice. But I can imagine a student or researcher hesitating to mention legitimate assistance because they fear it will cast doubt on the whole work. That possibility deserves attention. Asking people to be transparent will achieve little if they cannot understand what is allowed or trust that an explanation of their process will be considered fairly.
Disclosure could take the form of a straightforward account of how AI was used and how its contribution was checked. Where the development of an argument is being assessed, an outline and successive revisions could provide a basis for discussing how the student selected evidence and changed their reasoning. A conversation about the final text would then be considered alongside that record of the work. The detail should serve a purpose. Thus the effort of describing assistance shouldn’t become so burdensome that it discourages an honest account of it.
For lecturers, AI could help with aspects of teaching such as the repetitive work of checking assignments. For short written responses, I would explore whether AI could flag apparent omissions against criteria supplied by the lecturer. I can also see benefits in using AI to develop teaching materials. It could help lecturers turn their own material and ideas into activities that give students something to work through. A lecturer preparing a class from selected readings and notes might ask for questions that move from explaining a concept to applying it in an unfamiliar case. I find it useful to consider several possible ways into a difficult topic: a comparison, a short problem or a claim students have to challenge. Choosing among them would mean asking what each activity requires students to notice and where they might need further explanation.
Both lecturers and students would need opportunities to learn these practices. Using AI well also includes recognising an unsuitable method, finding an original source and deciding when to work without the tool. I would like to see this taught through examples from people’s own disciplines, where they can discuss why an answer is useful or misleading. Lecturers also need time for such learning. Access to a subscription does not give someone the confidence to use it thoughtfully, and relying entirely on self-teaching assumes that everyone has the same resources and time available.
The place we give AI in academic life reflects a wider understanding of what universities are for developing knowledge, cultivating judgement, and enabling people to participate in intellectual exchange. Assistance with research material, writing in another language or preparing feedback may serve those purposes, but its value cannot be judged from the finished work alone. We also need to consider what people have learned, which decisions they can explain, and whether they have become better able to question the results before them. Universities can support this through clear expectations, opportunities to learn, and time to examine how AI is changing everyday practice. The choices made in individual courses and research projects will gradually shape what an academic education enables people to understand and do for themselves.
Notes
[1] Hamsa Bastani et al., “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS (2025). link.
[2] Greg Kestin et al., “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting,” Scientific Reports (2025). link.