Should Students Be Allowed to Use AI?
Universities and schools should neither ban AI outright nor hand students unrestricted access to it. The evidence supports a bounded, structured adoption: AI tools produce real learning and efficiency gains, but only when students already bring — or are actively taught — critical thinking and academic self-belief. Left unstructured, AI use degrades the very cognitive struggle that builds thinking skill, and it opens the door to undetectable cheating. The right policy question isn't "AI or no AI," but "which tasks, with how much AI, and with what safeguards."
AI Delivers Measurable Learning Gains (Under the Right Conditions)
A Harvard randomized controlled trial found that students using a purpose-built AI tutor scored significantly higher on post-tests than students in an active-learning classroom, while spending less time on task, and also reported higher engagement and motivation. Critically, this tutor wasn't a generic chatbot, it was engineered around pedagogical best practices (active learning, cognitive-load management, growth mindset, accurate step-by-step solutions) and used to introduce material, not to replace higher-order in-class work like problem-solving and critique.
Separately, a peer-reviewed study of 744 Chinese university students found that AI literacy - awareness, skillful use, evaluation, and ethical use of GenAI, was positively associated with academic achievement. This relationship ran through students' academic self-efficacy, successfully using AI tools for academic tasks builds a sense of competence that in turn drives better academic outcomes. In other words, AI helps not just by producing answers, but by giving students an experience of mastery that boosts their confidence to tackle harder work.
Takeaway: AI can genuinely accelerate learning, but the studies showing this used carefully designed, pedagogy-aware tools, not open-ended essay generation.
The Benefit Depends on the Student
This is the pivot point of the whole debate, and it's where the research gets most interesting.
The same 744-student study found that critical thinking moderates how much AI literacy translates into self-efficacy. Students who scored high on critical thinking showed a much stronger link between AI use and academic self-efficacy than students who scored low. Practically, this means:
High critical-thinkers treat AI as a "sparring partner" — they question its outputs, catch its errors, and each catch reinforces their own sense of intellectual competence.
Low critical-thinkers treat AI as an answer machine — they copy outputs uncritically, so any success gets credited to the tool, not to themselves, and their confidence and independent capability stagnate or even erode.
This single finding reframes the entire "merits vs. demerits" debate: AI isn't inherently good or bad for learning, it simply amplifies whatever thinking capacity a student already has. A policy built around unrestricted access implicitly assumes all students are high critical-thinkers.
Unassisted Writing Builds Cognitive Capacities That AI-Assisted Writing Doesn't
Writing researchers describe drafting an essay as one of the most cognitively demanding tasks a student performs, comparable in mental effort to hard physical labor, precisely because it forces sustained use of working memory, planning, and self-critique. That difficulty is not incidental; it's the mechanism by which writing builds thinking.
The concern raised by educators is that when part of this struggle is outsourced, the corresponding cognitive growth doesn't happen. A small MIT study is often cited here: students who wrote with chatbot assistance showed lower brain activity while writing than unassisted students, and struggled more to recall what they'd supposedly written and felt less ownership over it. Newer classroom-facing "Socratic" AI tutors (e.g. ChatGPT Study Mode or Gemini Guided Learning) marketed to schools show a similar pattern in practice, they start by asking guiding questions, but when pushed, they readily produce polished draft paragraphs, effectively doing the thinking work students are supposed to be doing themselves.
Takeaway: For tasks meant to build reasoning and expression (not just deliver content), unassisted struggle helps in developing real cognitive skills.
Unstructured Access Creates a Real, Measurable Cheating Problem
A survey of 498 European university students using a randomized-response technique, designed specifically to counter dishonest self-reporting on sensitive topics, estimated that roughly 16–21% of students had used AI tools in an exam or assignment without disclosing it, roughly double the rate obtained by asking directly. This aligns with similar estimates from other countries (around 24% in a Vietnamese sample). Students also rated their positive experience with AI tools lowest for tasks with clear right-or-wrong answers requiring precision, like generating diagrams or tables, and highest for open-ended tasks like idea generation — exactly the kind of task where instructors have the least ability to verify effort.
Takeaway: Even well-intentioned integrity policies are undermined once AI text becomes functionally indistinguishable from human writing, and detection tools remain unreliable.
Resolving the Tension: A Task-Differentiated Policy
Putting these four pillars together, the disagreement between "AI helps learning" and "AI hurts learning" mostly dissolves once you separate what kind of task AI is applied to and whether critical thinking is being built alongside it:
Task type | AI role | Rationale |
Introducing new content, guided practice | AI tutor encouraged | RCT evidence shows real learning gains here |
Idea generation, outlining, feedback on a draft | AI as tool, disclosed | Builds self-efficacy without replacing core writing struggle |
Drafting original argument, analysis, in-class essays | AI restricted / in-person, unassisted | Preserves the cognitive work writing is meant to build |
High-stakes assessment of independent skill | AI prohibited, verified conditions | Prevents undisclosed misuse found in survey data |
The University of Sydney, for example, has adopted what it calls a "two-lane" model, requiring some assessments in secure, AI-free conditions while permitting AI on others so students can demonstrate they can use it well (Goldstein, The New York Times, Aug. 17, 2026) - reflecting exactly this task-differentiated logic rather than a blanket rule.
Practical Recommendations
The task-differentiated policy above answers when AI belongs in the writing process. The practices below answer how to use it well once it's allowed both for students, and for the instructors designing assignments around it.
For students
Know what the tool can't see. Every model has a knowledge cutoff and no live awareness of events after it. For anything time-sensitive for e.g. statistics, recent research, current events - treat the AI's answer as a starting point, not a citation, and verify it against a primary source before it goes in your paper.
Watch for baked-in bias. Training data carries the biases of whatever it was drawn from, so AI output can default to narrow or stereotyped framings without flagging that it's doing so. If a response feels one-sided, ask it to represent the other side or push back explicitly. This correction usually has to come from you.
Use AI to interrogate your draft, not replace it. Feedback, counterarguments, and clarity checks build your self-efficacy; having AI generate a paragraph does not.
Protect the drafting stage. Outlining and revision are safe to share with AI. The first attempt to state your own argument in your own words is the stage that builds the thinking skill described above - do that part unassisted.
Disclose, don't hide. Cheating data is driven almost entirely by undisclosed use. Most of the risk disappears the moment use is transparent and permitted.
For educators and institutions
Specify AI's role per assignment, not per course. A blanket "AI allowed" or "AI banned" policy ignores that the right answer depends on the task (see the policy table above).
Teach verification as a skill, not a warning. Students need explicit practice fact-checking AI output and noticing bias, the same way they're taught to evaluate any other source.
Design for drift. AI behavior and formatting shift as models update even without a prompt change; instructors relying on AI-generated rubrics or examples should periodically re-check them rather than assuming static output.
Reserve unassisted, in-person work for skills you actually want to assess. If independent argument or original analysis is the learning goal, structure the assessment (blue books, in-class writing) so that goal can't be quietly outsourced.
Bottom Line
The merits of AI in education are real but conditional while the demerits are real but avoidable. The lever that determines which outcome a student gets isn't the technology - it's whether critical thinking is deliberately cultivated alongside AI use, and whether the specific task still requires the student to do the cognitive work that builds durable skill. Policy should be designed around that distinction, not around "for" or "against."
Sources:
Kestin et al., "AI tutoring outperforms in-class active learning" (Scientific Reports, PMC12179260)
Wang, "The effect of generative artificial intelligence literacy on academic achievement" (Frontiers in Psychology, 2026)
Reiter, Jörling, Fuchs & Böhm, "Student (Mis)Use of Generative AI Tools for University-Related Tasks" (IJHCI, 2025)
Goldstein, "What Do Students Lose When They Stop Writing?" (The New York Times, Aug. 17, 2026)
The Economist, "Does AI stop children from learning?" (Aug. 18, 2026)
Comments