The debate is already behind student behaviour
Universities are still asking whether AI belongs in education. Students have already answered through the way they study.
AI now sits beside search, lecture notes and video as a routine interface to knowledge. It is available at any hour, responds to follow-up questions and can make a difficult topic easier to enter.
The HEPI Student Generative AI Survey 2026 found that 95% of UK undergraduates use AI in at least one way. Some 94% use generative AI to help with assessed work.
HEPI polled 1,054 UK undergraduates through Savanta in December 2025. Results were weighted by gender, institution type and study year, with a margin of error near 3%.
The 2026 Lumina Foundation-Gallup study adds a US view. It found that 57% of college students use AI in coursework at least weekly, including roughly one in five who use it daily.
The web survey covered 3,801 students pursuing associate or bachelor’s degrees. Among weekly activities, 64% used AI for coursework they did not understand and 60% used it to check answers.
These surveys use different samples and definitions, so their percentages are not interchangeable. They still show the same direction: AI use has moved from experimentation to routine.
The question is no longer whether students will adopt AI. It is whether universities will shape that use towards learning, or leave students to develop their habits alone.
Why AI became the default companion
From my work at the intersection of data, AI and learning, I see the technology’s strongest appeal as its ability to shorten the distance between confusion and a workable first explanation.
A student can ask for another example without embarrassment, test a counterargument or generate practice questions in seconds. That immediacy makes AI feel less like software and more like a study partner.
This does not mean students only want machines to complete their work. Jisc’s 2025 research found AI being used for research, revision, planning, note-taking and understanding difficult material.
The research combined discussions with 173 further- and higher-education students with seven surveys covering 1,274 responses. Students still valued personalised teacher feedback, discussion and collaboration.
That pattern fits a wider change in study behaviour. In Jisc’s 2025 digital experience survey, 15,398 higher-education students across 30 UK providers described how they learn with technology.
Although 72% of teaching was mainly on campus, 48% of students preferred blended or online learning. Some 46% preferred to study mainly online when they were not in class.
Learning is therefore increasingly flexible, digital and self-directed. When a student encounters a gap in understanding outside office hours, the resource that responds immediately will often be the first one consulted.
If explanation is available on demand, a university cannot base its value mainly on delivering information. Its value must come from trusted context, intellectual challenge, community and the development of judgement.
What Docsity AI reveals about the next study interface
Docsity AI makes this shift concrete. Its core idea is not to start with a blank prompt, but with material the student is already using.
The Docsity AI app combines capture and study in one workflow. Students can upload documents, record live audio, scan text with OCR,or add links from YouTube, Wikipedia and other websites.
It then turns those inputs into structured notes, summaries, concept maps and flashcard quizzes. Students can keep private notes and manage material created from shared or downloaded documents.
The web experience adds another layer. Students can ask questions about a document, then generate summaries, maps and quizzes from the same material.
This matters because each format can support a different study action. A summary can provide orientation, a concept map can expose relationships, and a quiz can move the learner from review to retrieval.
Document chat can support follow-up questions when a passage is unclear. The value lies in moving between representations of the same topic, rather than treating one generated answer as the final word.
That workflow reflects how independent study actually unfolds. Students capture material in class, encounter new sources later and need ways to organise, revisit and test what they have collected.
The design also carries risks. A summary can omit a qualification, a concept map can imply a false relationship, and a generated quiz can reward recall without testing deeper understanding.
Docsity AI should therefore be used as a layer over source material, not a substitute for it. Students should compare outputs with the original, challenge omissions and verify any claim that affects their work.
For universities, the wider lesson is to judge educational AI by its learning sequence. Good tools should make sources visible, invite checking and move students from compression towards retrieval, application and reflection.
Attention is now part of the learning infrastructure
Attention is not a side issue in a digital learning environment. Study, communication, entertainment and AI often share the same screen, making interruption part of the learning context.
A 2025 meta-analysis of 27 randomised experiments, involving 2,245 young adults, found that mobile-phone distraction had a medium negative effect on immediate recall.
The issue is not a moral failure among students. It is an environment designed around prompts to switch context, precisely when complex learning requires sustained thought.
AI can worsen that environment when it supplies a polished answer before the learner has made any effort. Fast completion can create the feeling of progress without the mental work that makes knowledge durable.
The same technology can also protect attention. It can divide complex material into stages, create retrieval exercises, adapt an explanation and offer feedback while the learner is still engaged.
The key test is whether a system captures attention only long enough to deliver an answer, or directs attention back towards recall, comparison, explanation, application and reflection.
Shorter formats are not necessarily shallower. A concise explanation can open the door to a demanding task, while a long lecture can still be consumed passively.
Better performance is not the same as learning
A student can produce a better answer with AI and still understand less when the tool is removed. That distinction should guide every university AI strategy.
The OECD Digital Education Outlook 2026 warns that general-purpose AI can improve task performance without creating durable learning when students offload the work needed to acquire a skill.
I have seen the same gap in my own work on AI summary generation. We followed a strong technical benchmark and produced summaries that looked impressive, yet students found them too short for effective study.
Testing different lengths showed that learner fit mattered more than model performance alone. A technically elegant output has little educational value if it does not help the student build a usable mental model.
In a large randomised field experiment published in PNAS, nearly 1,000 secondary-school mathematics students using unrestricted AI performed 48% better during assisted practice.
When the tool was removed, those students performed 17% worse than the control group. A tutor with teacher-informed safeguards largely removed that later penalty, though it did not create a clear unassisted exam advantage.
The experiment involved school mathematics in one setting, not every university subject. Its principle still matters: an answer engine and a system designed to build capability are not the same product.
Educational AI should behave more like a good tutor. It should ask for an attempt, offer a hint, probe the reasoning and gradually withdraw support.
The goal is augmentation, not autopilot. Universities should use AI to remove low value friction and create more room for reasoning, not to remove the reasoning itself.
AI literacy must become a core graduate capability
Students need more than a policy page listing prohibited uses. They need practical instruction in how AI systems work, where they fail and when their use is inappropriate.
That includes checking claims against primary sources, recognising invented citations, testing for bias, protecting confidential data, documenting assistance and retaining responsibility for the final work.
The UNESCO AI Competency Framework for Students centres a human mindset, ethics, techniques, applications and system design. Its emphasis on critical judgement is essential.
The need is visible in the HEPI data. While 68% of surveyed students said generative AI skills are essential to thrive, only 48% felt teaching staff were helping them build those skills for a future career.
Universities should embed AI literacy within disciplines. A historian must test provenance differently from an engineer, while a medical student faces different reliability and privacy risks from a designer.
Course-specific practice will do more for responsible use than a generic workshop. Students need repeated opportunities to examine outputs, trace sources and explain why an AI suggestion should be accepted or rejected.
Guidance must also be precise at assignment level. Students should know whether AI is prohibited, allowed for defined tasks or expected, and what disclosure is required.
Ambiguity does not protect academic integrity. It rewards those most willing to take risks and disadvantages students who are trying to follow the rules.
Assessment must reveal the thinking process
Assessment is already changing. The 2026 HEPI survey found that 65% of respondents believed their institution’s assessment had changed significantly in response to generative AI.
Trying to preserve every pre-AI task unchanged is a losing strategy. If a take-home assignment can be completed convincingly by a general-purpose model, it may no longer show what the student knows.
A 2026 evaluation of two AI-content detectors found important accuracy and fairness limits. Detector scores should support investigation, not decide misconduct on their own.
Assessment should make the process more visible. Drafts, source trails, decision logs, oral defences, live problem-solving and reflections on rejected AI suggestions can reveal how a student reached a conclusion.
Authentic projects can also apply knowledge to unfamiliar, local or current contexts. These formats make judgement more important than fluent reproduction.
This does not mean every assessment must become a supervised exam. Universities need a better mix of practice and validation, with clear reasons for when AI is available and when independent capability must be shown.
Students should learn in three modes: without AI, with education-focused AI and with general-purpose AI. Each mode develops a different capability, and each should be assessed accordingly.
Universities must strengthen the human layer
As AI becomes more capable, human teaching becomes more important, not less. Students still need experts who can spot misconceptions, connect ideas, set standards and sustain motivation.
They also need peers, debate and the productive discomfort of having their views challenged by another person. Learning is not only a transfer of information; it is participation in an intellectual community.
The best use of AI is to extend that human layer. It can handle routine clarification, generate low-stakes practice and help staff identify patterns of difficulty, leaving more time for feedback and mentorship.
Accountability for curriculum, assessment and student welfare must remain human. An institution cannot outsource the decisions that define educational quality.
Access and governance matter too. If effective AI support depends on who can pay for a premium service, existing inequalities will widen.
Universities need equitable access, privacy standards, secure handling of student work and faculty development. They must evaluate tools through retention, transfer, critical thinking and inclusion, not usage alone.
The future belongs to institutions that design for learning
Universities cannot outwait this behavioural shift. AI is already part of how students search, organise, practise and make sense of difficult material.
The right standard is simple to state and demanding to meet: does this use of AI make the student think more deeply, or merely finish more quickly?
Institutions that design around that question can turn an informal study companion into a learning scaffold. Those that do not risk allowing convenience to replace capability.
AI should remain a companion: available, adaptive and useful, but never the owner of the intellectual journey. Higher education must teach students to work with AI, question it, resist it and think beyond it.
