Ground educational AI in trusted content, not open internet

For several years, the debate around remote education has been framed too narrowly. We have asked whether online learning can reproduce the experience of the physical university, when perhaps the more useful question is whether it should try to.

Students increasingly choose remote education for reasons that go well beyond convenience.

It can allow a professional to acquire new skills without leaving a job, a parent to study around family responsibilities, or a student to access a specialised programme that does not exist locally. It can also enable people to learn at different speeds rather than requiring everyone to follow the same timetable.

This is particularly important as the shelf life of professional skills shortens. Higher education is no longer necessarily something that happens once, in a particular location and at a particular stage of life. For many students, learning is becoming a recurring activity alongside employment and other responsibilities.

But flexibility alone does not make remote education effective.

One legacy of the pandemic is that ‘remote learning’ is still sometimes confused with taking a conventional university model and putting lectures on Zoom. High-quality online education requires almost the opposite approach: it needs to be designed for distance from the beginning.

The hidden challenges of studying remotely

The obvious challenges are familiar: unreliable connectivity, unequal access to devices and the risk of social isolation. But there are subtler problems that institutions also need to address.

One is temporal inequality.

A university may be geographically accessible to a student while remaining temporally inaccessible. For example, a live session at a convenient European time may take place in the middle of the night elsewhere; or a working student may technically be able to participate in a programme but be unable to attend activities scheduled during working hours.

Institutions therefore need to think about accessibility not simply in terms of where students are but also when they are able to learn.

There is also the cost of flexibility. Giving students greater control of their schedule is valuable, but it transfers some of the organisational burden traditionally carried by the institution to the learner. Students must decide when to study, what to prioritise and how to remain on track, often while surrounded by the distractions of everyday life.

This can particularly affect students who are neurodivergent, those returning to education after a long absence and those balancing study with unpredictable work or caring responsibilities.

Another challenge is that problems become less visible online. In a physical classroom, an instructor can often see confusion, frustration, or withdrawal. In a remote environment, a student can quietly disengage for days before anyone notices.

And then there is AI.

Generative AI gives remote students something previous generations never had: almost instantaneous assistance at any time of day. Yet the same technology can remove an essential ingredient of learning: the struggle required to understand something for oneself.

The challenge is therefore not simply to give students more technology. It is to design technology that knows when to help, how much to help and sometimes when not to provide the answer.

From content delivery to learning architecture

The first principle for effective remote education should be to design online-first, rather than asynchronous-only learning.

Recorded material, readings and exercises give students flexibility, while synchronous sessions should be reserved for activities where being together genuinely adds value: discussion, problem-solving, debate, mentoring and feedback.

At the OPIT – Open Institute of Technology, for example, asynchronous material is combined with weekly live faculty sessions, continuous tutor support and peer interaction. The objective is not to replicate a lecture theatre online but to use each format for what it does best.

This also suggests that universities should pay more attention to ‘latency in support’: how long does a student remain stuck before meaningful help arrives?

In a campus environment, students have multiple informal opportunities to resolve problems. They can approach a lecturer after class or ask a peer. Those accidental interactions largely disappear online if not addressed properly.

Institutions should combine seven-days-a-week online tutors, peer communities, structured office hours and professors’ ‘office hours’ so that a student rarely reaches a dead end.

AI should create better questions, not simply faster answers

The most interesting development in educational AI is therefore not the generic chatbot. It is the emergence of course-specific systems that understand the curriculum and can scaffold learning. In other words, smart AI tools are designed for specific classes and support students through tough topics instead of just giving them answers.

While, at OPIT, the core of the learning experience provided to students is ‘human-led’ (and this will remain core going forward), a well-designed AI tutor can identify what a student is trying to understand, ask questions, provide hints, change the level of an explanation or direct the student back to relevant material.

Importantly, it should be grounded in trusted course content rather than treating the open internet as an unquestioned source of truth.

There is now encouraging evidence behind this model.

A 2025 randomised controlled trial involving undergraduate physics students at Harvard found substantial learning gains from a carefully designed AI tutor. The important qualification is that the researchers did not simply give students access to a general-purpose chatbot: the AI experience incorporated established principles including scaffolding, active learning, targeted feedback and self-pacing.

That distinction matters.

At OPIT, we have been exploring the same principle through an AI Copilot integrated into the learning environment. Rather than functioning simply as an answer generator, it is grounded in institutional course material and designed to provide contextual support; its behaviour also changes during assessments to protect academic integrity.

This is one example of a wider direction higher education should explore: AI as part of the pedagogical architecture, rather than an external tool sitting beside it.

UNESCO's 2025 survey of institutions associated with its Chairs and UNITWIN networks found that nearly two-thirds had already introduced or were developing guidance on AI use, while confidence in its effective pedagogical application remained uneven.

The next stage therefore needs to be less about whether universities ‘allow AI’ and more about precisely what kinds of learning behaviour they want AI to encourage.

Using data without reducing students to data

Online environments produce signals that physical classrooms often cannot: whether students have accessed material, attempted exercises, returned to difficult concepts, or gradually stopped participating.

The temptation is to use these signals to predict failure. But prediction is valuable only if somebody acts on it.

Recent research on early-warning systems illustrates this point. Machine-learning models were able to identify at-risk students with relatively high accuracy, but the more important element was the intervention that followed: relational, carefully designed communication from instructors prompted students to re-engage with learning activities.

The principle should be human-in-the-loop analytics. Data should create an opportunity for a conversation, not an automated judgement about a student.

Designing for difference

Remote education can also become more inclusive precisely because it does not require every learner to experience education in the same way.

Recorded lectures can have transcripts and summaries. Students can pause, replay or accelerate content. Interfaces can offer alternative representations of material. Translation and captioning tools can reduce linguistic barriers. Virtual laboratories and simulations can increasingly give students access to practical experiences without requiring physical proximity to specialist equipment.

At OPIT, we have found that features such as recordings, transcripts, flexible pacing and one-to-one support are particularly valuable for neurodiverse learners, but these are not ‘special’ features. They frequently improve the experience for everyone.

That is perhaps the larger lesson.

The future of remote education will not be determined by whether universities have better video platforms or more sophisticated AI. It will depend on whether they redesign the learning experience around the reality of their students and on whether they can create a genuinely enriching ‘human learning experience’ in an online setting, end-to-end.

Universities should measure not only attendance and completion but also the time students spend stuck, the speed and quality of feedback, participation across time zones, opportunities for meaningful interaction and whether technology is increasing or reducing student agency.

Remote education is sometimes described as education without proximity. Done properly, it can be something more interesting: education in which proximity, time, pace and support are deliberately redesigned.

The goal should be to use the possibilities of the medium to create a form of higher education that could not have existed before.

Riccardo Ocleppo is the founder and CEO of Docsity and founder and director of OPIT – Open Institute of Technology.

This article is a commentary. Commentary articles are the opinions of the author and do not necessarily reflect the views of University World News.