When AI Becomes the Thinker, Students Risk Losing the Learning

New research suggests that the cognitive consequences of AI may depend less on how often students use it than on the role they assign it in their own thinking.

The study, "AI as Cognitive Complement or Replacement? Perceived AI Role and Cognitive Independence Among University Students," was published in the journal Computers. It was authored by Dalma Lilla Dominek, Vanessza Kapusi, Szabolcs Ceglédi and Zoltán Szűts, with affiliations spanning Ludovika University of Public Service, the Mathias Corvinus Collegium Foundation, Eszterházy Károly Catholic University and János Selye University.

Based on survey data from 93 part-time students at a Hungarian university, and statistical analysis of the 73 respondents who used AI for educational purposes, the researchers examined whether students saw AI as a complement to their thinking or a replacement for it. They then tested how that perception related to self-reported cognitive independence, flexibility and augmentation. The study found that students who viewed AI as a cognitive partner reported greater independence, while frequency of use and specific use cases did not significantly predict independent thinking.

The Mindset Around AI Matters More Than the Amount of Use

The study's strongest result concerns perceived AI role. Students who saw AI as supporting rather than replacing their own thinking reported significantly higher cognitive independence. The relationship was moderately strong: perceived AI role correlated with cognitive independence at r = 0.431, with a probability below 0.001. In a regression model, this perception explained 18.6% of the variation in cognitive independence and emerged as a significant positive predictor.

On the other hand, neither AI-use frequency nor common purposes, including information seeking, summarisation, translation, coding, problem-solving and creative-content production, individually predicted cognitive independence. When these variables were tested together, perceived AI role remained the only significant predictor.

This challenges a widespread assumption that heavier AI use automatically weakens intellectual autonomy. Two students may use the same tool equally often yet experience different outcomes. One may ask AI to generate a final answer and accept it with little scrutiny. Another may use the same system to test an argument, identify weaknesses, compare alternatives and refine an original response.

Measuring AI use by frequency alone may reveal little about whether students are learning deeply or outsourcing thought.

Cognitive Offloading Is Not the Same as Cognitive Decline

The research also adds nuance to the debate over cognitive offloading, the use of external tools to reduce mental effort. Offloading is not inherently harmful. Calculators, search engines, databases and note-taking systems all reduce certain cognitive burdens. The critical issue is what gets offloaded and whether the user remains actively engaged in interpretation, verification and decision-making.

Generative AI raises the stakes because it does more than retrieve information. It produces explanations, arguments, code, summaries and creative suggestions that can appear complete and authoritative. This can free students to focus on higher-level reasoning, but it can also encourage automation bias, reduced vigilance and acceptance of plausible but inaccurate outputs.

The study's findings support the view that offloading becomes risky when AI substitutes for executive functions rather than supporting them. Students who regarded AI as complementary reported stronger independence regardless of how frequently they used it. The authors therefore argue that cognitive effects may be shaped more by metacognitive control and active engagement than by exposure itself.

An exploratory finding offers a possible example. AI use for coding and problem-solving was positively associated with perceived cognitive augmentation. These activities often require iterative testing, debugging, evaluation and revision, which may keep the learner intellectually involved.

However, the result is not conclusive. Coding and problem-solving was the only significant predictor inside a regression model that was not statistically significant overall. The authors correctly describe it as hypothesis-generating rather than confirmatory.

Universities Need AI Pedagogy, Not Just AI Policing

Many universities have focused their initial response to generative AI on misconduct detection, assessment security and restrictions. Those concerns remain legitimate, but this study suggests that control-oriented policies are insufficient.

The more important educational question is whether students learn to position AI as a tool for inquiry rather than an authority that ends inquiry. Universities can promote augmentation-oriented use by designing assignments that require students to document their reasoning, compare AI-generated alternatives, identify errors, defend revisions and distinguish between machine suggestions and their own conclusions. Oral examinations, reflective commentaries and process-based assessment can make intellectual ownership more visible.

Teachers should also explicitly discuss when cognitive offloading is useful and when it becomes dependency. Using AI to reorganise notes may save time without weakening understanding. Asking it to produce a complete argument that the student cannot explain is a different cognitive act.

AI literacy must therefore extend beyond prompt writing. Students need skills in source verification, uncertainty assessment, bias detection, model limitations and metacognitive monitoring. They should understand that fluent output is not necessarily accurate, original or appropriate.

This has development-policy relevance beyond Hungary. Universities in developing economies face pressure to adopt AI quickly while managing unequal digital access, limited faculty training and variable institutional capacity. Blanket bans may disadvantage students who need supported access, while uncritical adoption may deepen learning inequalities between those who can evaluate AI and those who cannot.

The Finding Is Promising, but It Does Not Prove AI Builds Independence

The research offers a useful conceptual shift, but its limitations are substantial. First, the analytical sample contained only 73 AI users from a single Hungarian university. Most participants were part-time students, more than half held master's degrees, and nearly two-thirds came from humanities and education-related fields. The findings cannot be assumed to represent younger full-time students, STEM-heavy institutions or universities in different cultural and economic settings.

Second, the study measures perceived cognitive independence rather than objective performance. Students who feel independent may not necessarily perform better on unaided reasoning, originality or problem-solving tasks.

Third, the cross-sectional design cannot establish causality. Viewing AI as a complement may encourage independence, but the reverse is equally plausible: students who already feel intellectually autonomous may be more likely to treat AI as a tool rather than a substitute.

Measurement also requires caution. The central complement-versus-replacement variable relied on one survey item. The cognitive scales were newly developed, produced moderate reliability scores and require validation in larger samples. No formal adjustment was made for multiple statistical comparisons, increasing the need for caution around marginal findings.

Future research should combine surveys with experiments, longitudinal tracking and objective tasks. Researchers could compare students asked to accept AI answers, critique them, revise them or work without assistance. Such designs would clarify whether augmentation-oriented instruction produces measurable gains in reasoning and retention.

The study does not settle whether generative AI strengthens or weakens student cognition. It offers a more productive question: does the student remain the active thinker?