AI is transforming education, but its effects on learning remain complex.
Research reports improvements in academic achievement, engagement, and motivation, while other studies raise concerns about long-term retention, cognitive offloading, and independent thinking.
Continuing my thoughts on the purpose of learning in the age of AI, I’ve been exploring how AI might shape the way we think, learn, and develop knowledge.
This blog post presents both perspectives and argues that AI’s educational value depends on how teachers design learning, how students engage with the technology, and what outcomes educators measure.
Introduction
AI can explain concepts, generate essays, analyze information, and provide personalized feedback. These capabilities challenge educational systems that traditionally measure learning through examinations, assignments, and completed tasks.
Zhao (2026) argues that educators must reconsider what remains worth teaching when AI can perform many activities that schools have historically rewarded.
Similarly, the Massachusetts Institute of Technology’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training (2026) calls for a thoughtful reassessment of teaching, assessment, and the purpose of education.
The report argues that education must adapt to an AI-enabled world without sacrificing the cognitive effort, human relationships, and independent judgment essential to learning. It recommends moving beyond simply restricting AI and instead redesigning teaching, assessment, and curricula around meaningful learning outcomes.
It emphasizes augmentation over automation, protecting productive struggle, strengthening experiential and project-based learning, and teaching responsible AI literacy. It also calls for more meaningful assessment, stronger student–teacher and peer interactions, flexible policies suited to different disciplines, and continuous evaluation of educational practices.
Ultimately, MIT advocates using AI to expand human curiosity, creativity, expertise, and learning — not replace them.
However, the debate extends beyond whether AI belongs in classrooms. Researchers must establish when AI supports learning, when it risks replacing essential cognitive processes, and how educators can distinguish immediate performance from lasting understanding.
The positive perspective
Research provides evidence that generative AI can improve several educational outcomes. Chen and Cheung (2025) synthesized 57 studies involving 97 effect estimates and reported positive effects on language skills, academic achievement, motivation, and higher-order thinking.
However, they found no statistically significant overall effect on metacognition. This distinction matters because performing better on a task doesn’t necessarily mean students have become better at monitoring and regulating their own learning.
Han et al. (2025) examined 68 experimental and quasi-experimental studies and reported a moderate overall positive effect of generative AI on learning outcomes. Their analysis also identified substantial variation across studies, suggesting that the results depend on educational contexts and intervention design.
AI may help learners by providing immediate explanations, adapting examples to their level, offering feedback, and enabling repeated practice.
Students who hesitate to ask questions in class may also benefit from a low-pressure environment in which they can request clarification.
These findings support the use of AI as an educational resource. They don’t establish that every application improves learning or that AI can replace effective teaching. Instead, they show that AI has potential when educators align its use with clear learning objectives.
The cautionary perspective
Other research raises concerns about what students retain after using AI.
Barcaui (2025) conducted a randomized controlled trial involving 120 undergraduate participants studying AI. In a surprise knowledge-retention test conducted 45 days after learning, the AI-assisted group scored 57.5%, compared with 68.5% for the traditional research group.
The author interpreted this finding as evidence that unrestricted ChatGPT assistance may undermine long-term retention in some circumstances.
This result deserves attention, but it shouldn’t be generalized to all AI-supported learning. The research examined a particular intervention, subject, and student population. Other instructional approaches may produce different outcomes.
One possible explanation involves cognitive offloading. Learners can delegate mental tasks to external tools, reducing the effort they invest in retrieving, organizing, and explaining information.
Such offloading can be useful when it frees attention for complex reasoning. However, it may become counterproductive when students delegate the very cognitive activities they need to develop.
Retrieval practice, metacognition, and independent explanation help learners consolidate knowledge and identify gaps in understanding (Roediger & Karpicke, 2006).
Educators should therefore provide opportunities for students to work without AI, retrieve knowledge from memory, and apply what they’ve learned independently.
Why do the research findings differ?
The evidence doesn’t support a simple conclusion that AI either improves or harms learning. Differences in research design, educational level, subject area, intervention duration, and assessment methods help explain why studies reach different conclusions.
For example, Han et al. (2025) found substantial heterogeneity across their included studies, despite a positive average effect. Chen and Cheung (2025) also identified contextual factors that influenced outcomes.
These findings suggest that the effects of AI vary according to how learners use the technology and how educators structure learning activities.
A particularly important contribution comes from Boolzen et al. (2026), who reviewed 85 studies on generative AI in STEM education.
Their meta-analysis found an overall positive effect on cognitive learning outcomes, but the results varied considerably across studies. The prediction interval ranged from a negative to a strongly positive effect, and the authors identified potential publication bias.
Consequently, the average effect shouldn’t be interpreted as a guarantee that AI will benefit learners in every context.
These differences also highlight the importance of measurement. A study might assess immediate test performance, while another measures retention several weeks later. One might examine motivation, while another investigates metacognition. These outcomes are related, but they aren’t interchangeable.
Future research should examine not only whether AI improves performance, but also whether learners retain knowledge, transfer it to unfamiliar problems, develop independent reasoning, and become more capable of regulating their learning.
From automation to augmentation
MIT (2026) recommends adapting education to an AI-enabled environment while keeping human capabilities at the center. Its emphasis on augmentation rather than automation offers a useful principle for instructional design.
AI can support students as they explore unfamiliar subjects, compare interpretations, generate ideas, and receive feedback. However, students should still evaluate the information they receive, verify claims, identify errors, and justify their decisions.
For example, instead of asking AI to write an essay from beginning to end, a student might use it to identify counterarguments, examine weaknesses in an initial draft, and compare alternative explanations.
The student then decides which suggestions are defensible and explains the reasoning behind the final work.
This approach also connects with UNESCO’s human-centered guidance for generative AI in education, which emphasizes appropriate use, inclusion, and the protection of human agency (Miao & Holmes, 2023).
The objective isn’t to eliminate cognitive effort. It’s to direct that effort toward learning, evaluation, and meaningful problem-solving.
Rethinking assessment and student agency
When AI can produce polished assignments, educators need assessment methods that reveal how students think. MIT (2026) recommends reconsidering assessment practices and creating clear policies for AI use.
Oral examinations, portfolios, practical projects, supervised tasks, and reflective discussions can help educators understand students’ reasoning. Students can also critique AI-generated answers, verify sources, identify unsupported claims, and explain how they revised a response.
Zhao (2026) offers a complementary perspective. He argues that education should give students greater ownership of their learning by connecting curiosity with meaningful problems, experimentation, and collaboration.
Rather than requiring students to master every foundation before engaging in practical work, educators can allow authentic problems to motivate the acquisition of new knowledge.
This approach doesn’t make foundational knowledge or teacher guidance unnecessary. Students still need support in evaluating evidence, recognizing gaps, and developing effective strategies. However, it creates opportunities for students to apply knowledge in contexts that matter to them.
A balanced approach to AI integration
Effective AI integration requires educators to consider both its opportunities and limitations.
AI can provide accessible explanations and individualized feedback, but students also need independent practice, human interaction, and opportunities to struggle productively with challenging ideas.
Teachers should communicate clear expectations about when AI is permitted, required, or prohibited. Students should learn to verify information, acknowledge AI contributions, protect personal data, and recognize situations in which independent work is essential.
Educational institutions should also evaluate access, accessibility, and equity. Students don’t all have the same technological resources, prior knowledge, or ability to evaluate AI-generated information.
Most importantly, institutions should assess whether AI-supported activities achieve their intended learning objectives. Small-scale trials, transparent evaluation, and follow-up assessments can help educators identify which practices work, for whom, and under what conditions.
These ideas also highlight the importance of developing true AI literacy, enabling learners to engage with AI critically, responsibly, and thoughtfully.
So, who does the thinking?
From the perspectives of cognitive psychology and neuroscience, the key question isn’t whether students use AI, but how it affects their thinking and learning. Learning involves attention, working memory, retrieval, and the integration of new knowledge. Producing a correct answer doesn’t necessarily mean a student understands or can apply it independently.
Cognitive offloading allows people to delegate mental tasks to external tools (Risko & Gilbert, 2016). While this can be useful, relying on AI to do the reasoning may reduce opportunities to practice essential cognitive skills. Research shows that actively retrieving information strengthens long-term retention (Roediger & Karpicke, 2006).
AI should support thinking, not replace it. Students need opportunities to question AI-generated answers, evaluate evidence, retrieve knowledge, and apply concepts independently.
The goal is to develop learners who can use AI critically while retaining the capacity to reason, reflect, and make informed decisions for themselves.
I’ve been thinking about a deeper question: what happens to human learning when AI reaches superintelligence? We don’t yet know how that future will unfold, but it makes rethinking the purpose of education more important than ever.
Conclusion
AI’s role in education shouldn’t be judged through technological enthusiasm or fear alone. Research identifies real opportunities to improve academic performance and engagement, alongside important questions about retention, metacognition, cognitive effort, and independent learning.
The most defensible position is conditional rather than absolute. AI can support learning when educators use it intentionally, but it can also undermine learning when it replaces the cognitive activities that students need to practice.
These reflections also connect with a broader idea: education must teach wisdom, not just knowledge. As AI becomes more capable of providing answers, education must help learners develop independent judgment, ethical reasoning, creativity, and the ability to use knowledge responsibly.
Education must therefore move beyond measuring what students produce and pay greater attention to what they understand, remember, question, and can apply independently.
AI should expand opportunities for learning without displacing the human judgment, curiosity, relationships, and intellectual effort that make education meaningful.
Education must prepare students not just to navigate the world as it is, but to create a better world for everyone.
