AI Reliance and the Future of Critical Thinking

As generative AI has become part of university life, a familiar concern arises: are students becoming too reliant on it? Perhaps the more useful question is what students rely on AI to do. Frequency alone does not imply surrendering judgement.

Two kinds of reliance

We might distinguish between substitutive and developmental reliance. These are proposed distinctions, not established research categories. Substitutive reliance occurs when AI performs the thinking students are expected to practise. A student might ask it to construct an essay’s argument, choose the evidence and write the conclusion. The result may look impressive, but the student has delegated much of the reasoning.

Developmental reliance looks different. The student writes an initial argument, then asks AI to challenge its assumptions or suggest counterarguments. The student checks those suggestions against evidence and decides what to change. Likewise, a business student might compare their own proposed solutions with AI-generated alternatives rather than simply adopt its recommendation. In both cases, the decisive issue is who extends judgement.

The distinction shifts. A student may begin by using AI to explore a problem and later let it write the final answer. Another may accept an AI suggestion, then question it after consulting the course materials. What matters is the sequence of decisions: who frames the question, tests the evidence and accepts responsibility for the conclusion?

What does recent research tell us?

Hou et al. (2025) studied undergraduates’ AI reliance during problem-solving and identified reflective, cautious, thoughtless and collaborative behaviours. Critical thinking was positively associated with several more reflective patterns of use. Importantly, their findings describe relationships, not proof that one behaviour causes another.

Other findings point in different directions. In a randomised study, Barcaui (2025) found lower delayed knowledge-retention scores among students who used ChatGPT as a study aid than among those using traditional methods. Yet Kestin et al. (2025) found that students using a carefully designed AI tutor achieved greater immediate learning gains in less time than peers in an active-learning class. Li et al. (2025) also examined structured, progressive AI prompting in professional training, including its effects on critical thinking and cognitive load. Together, these studies suggest that the design and purpose of AI assistance matter more than a simple yes-or-no judgement about use. 

What should universities do?

The practical implication is not to ban AI or encourage its unrestricted use. Teachers can ask students to show their first reasoning, explain which AI suggestions they rejected, and justify the decisions behind their final work. Such tasks make thinking visible without assuming all AI use is valuable. This also calls for transparency: students should be able to explain not only what AI contributed, but why they accepted or rejected its suggestions.

Nor should we romanticise independent struggle. AI can remove unnecessary barriers, offer practice and provide timely feedback. But assistance becomes questionable when it removes the very reasoning a course aims to develop.

One practical approach is to build moments of independent reasoning into AI-supported assignments. Before consulting AI, students could record a provisional answer and the evidence behind it. Afterwards, they could annotate the suggestions they used, challenged or discarded. In a seminar format session, comparing these decisions with classmates could reveal that a fluent AI response is not always a well-supported one. Such practices assess judgement, not merely the polished final product.

The future of critical thinking may depend less on how often students consult AI than on whether they remain responsible for questioning, evaluating and deciding. The central educational question is therefore not simply how much AI students use, but what intellectual work they still do themselves.