Student agency and AI are often discussed as if the relationship were straightforward. Give students a powerful tool, and they gain more control over their learning. Yet control over a tool is not necessarily control over the thinking, decisions, and knowledge-making involved in a task.
This is the question Jasper Roe and Mike Perkins take up in a new critical review. Their concern is larger than whether generative AI helps students finish their work. They ask what happens to learners’ capacity to choose, judge, act, and take responsibility when part of that work is handed to an AI system.
This question sits at the centre of my broader work on AI and cognition, especially the effects of AI on judgment, thinking, and intellectual control.
Summary of the Study
The authors conducted an exploratory review of research on generative AI and agency in education. They searched four academic databases and analysed 11 relevant publications through the lens of critical digital pedagogy, a perspective influenced by Paulo Freire’s concern with freedom, power, and learners’ ability to understand and act upon their world.
The review distinguishes learner agency, teacher agency, and epistemic agency: the last refers to a person’s capacity to participate in creating and evaluating knowledge.
Three ideas run through the literature. The first concerns control in digital learning spaces: students can decide whether to accept, reject, revise, or question an AI response, but the system also shapes the choices available to them. This is where evaluative judgement becomes central. Access to AI does little for agency if a learner cannot assess the quality of its output or decide when using it would interfere with the purpose of the task.
The second idea is inequality. Paid tools, stronger models, connectivity, language, and digital literacy affect who can use AI effectively. Personalisation may give some learners greater freedom while leaving others with weaker access or less influence over the technologies introduced into their classrooms.
The third idea is that agency itself may be changing as thinking and decision-making become distributed across people and machines. I find this claim worth taking seriously, but I would be careful about describing AI as an agent in the same sense as a student or teacher. A system can influence action without possessing educational intentions or accepting responsibility for the outcome.
The authors are appropriately cautious about the strength of the evidence. The research base is small, recent, largely theoretical, limited to English-language publications, and thin on long-term classroom evidence. Their themes help us name emerging concerns; they cannot establish that AI has already strengthened or diminished student agency across education.
The authors also used Claude to support coding and theme development, with final interpretations retained by the researchers. They acknowledge the tension: a technology under critical examination also participated in the analytical process. That decision does not invalidate the review, though it makes transparency about prompts, model influence, and human disagreement especially important.
Practical Insights
The most useful classroom question raised by this review is simple: who is making the consequential decisions? Teachers can turn that question into an agency audit for any AI-supported activity.
- Before using AI, ask students to identify the decisions they must retain: defining the problem, selecting evidence, judging accuracy, forming a position, or deciding what the final work should say.
- After using it, have students mark what they accepted, revised, rejected, and independently verified. This keeps the focus on judgment and connects with concerns about cognitive offloading.
- Give students a genuine choice about whether to use AI when the learning outcome permits it. Requiring AI can also reduce agency, particularly for learners who have concerns about privacy, authorship, access, or the environmental and social costs of these systems.
- Discuss the tool’s influence on the task. What options did it make visible? What did it leave out? Whose language and assumptions appeared in its response? These questions bring critical AI literacy into ordinary subject teaching.
- Involve students in reviewing classroom AI rules and proposed tools. Agency becomes difficult to defend when learners have no say in technologies that collect their data or shape how their work is produced and judged.
For researchers, the next step is empirical. Studies need to observe agency during actual AI-supported tasks, distinguish it from engagement or satisfaction, and follow learners over time. Comparison across cultures, socioeconomic settings, disciplines, and levels of access is also necessary. Without that work, agency risks becoming another appealing word attached to whichever educational future a writer already prefers.
The review leaves us with a useful warning. Faster completion and personalised assistance can coexist with less intellectual control. If students are still making the decisions that matter, AI may expand what they can do. When the system begins deciding the direction, substance, and standards of the work, the appearance of choice can hide a gradual surrender of agency.
Reference
Roe, J., & Perkins, M. (2026). Agency in the age of generative AI: A critical review of educational implications. Academia AI and Applications, 2(3), Article 8422. https://doi.org/10.20935/AcadAI8422
