Author – Hewei Gao

From Tool to Learning Partner

A student sits down to write an essay, solve a programming exercise or prepare for an exam. Instead of searching through textbooks or waiting for office hours, they open an AI assistant. The system offers explanations, suggests what to try next and responds when the student asks follow-up questions. In many ways, this looks like a breakthrough for education: support is available at any time, at low cost and in a conversational form.

But this also changes the educational relationship. AI is no longer just a tool for retrieving information. It becomes a learning partner, taking part in the decisions that shape how the learner understands and improves their work. This raises an important alignment question for education: when AI becomes a learning partner, how can we preserve learner agency?

Learner agency is not simply the freedom to click different buttons or ask different prompts. In education, it means students remain active participants in their own learning. They can set goals, ask questions, monitor their understanding and learn from the consequences of their choices. Education researchers often describe part of this as self-regulated learning: the ability to plan, monitor and adjust one’s approach when something is not working (Zimmerman, 2000).

Help That Builds Capacity

Preserving agency does not mean leaving students alone. Education has always involved support. Teachers, peers and learning tools can help students reach beyond what they could do independently. The question is not whether AI should assist learners, but what kind of assistance expands rather than narrows their capacity to think, choose and act.

This is where scaffolding matters. Wood et al. (1976) used the term to describe how a tutor can support a learner through a task while gradually handing responsibility back to them. A good scaffold reduces unnecessary difficulty without removing the learner from the problem-solving process. In this sense, AI assistants can be valuable educational tools. They can make feedback more immediate, explanations more adaptive and help-seeking less intimidating. Recent work on hybrid human-AI learning technologies also suggests that AI in education should not be understood simply as replacing teachers or learners. It is better seen as redistributing roles between humans and technical systems (Molenaar, 2022).

If designed well, AI can help learners enter more challenging tasks and build confidence through guided practice. Large language models may also support personalisation and accessibility in education, even though these opportunities come with real risks (Kasneci et al., 2023).

When Assistance Becomes Dependence

Those risks begin with the same features that make AI assistants useful. A chatbot that is always available and speaks with confidence can easily become a source of conversational authority. It does not need to command the student. It only needs to keep suggesting the next step. Over time, the learner may stop doing the work of framing the problem and judging the answer for themselves. The task is still completed, but the student’s own judgement has become less central.

This risk is related to cognitive offloading: the use of external tools to reduce internal cognitive effort (Risko & Gilbert, 2016). Offloading is not always bad. Much of everyday thinking already depends on tools around us. But in education, some cognitive effort is the point. If the AI routinely performs the thinking that students are meant to practise, then assistance can quietly become dependence.

Designing for Empowerment

So how can we think more precisely about agency in AI-assisted learning? One useful lens comes from the concept of empowerment. In AI and information theory, empowerment is a way of asking how much influence an agent’s actions have over future states. Put simply, someone is more empowered when what they do now makes a meaningful difference to what they can experience or achieve later (Klyubin et al., 2005; Salge et al., 2014).

This gives us a helpful way to think about learner agency. In a high-agency learning environment, the learner’s choices genuinely affect how the learning process unfolds. In a low-agency environment, the interaction may look lively on the surface, while the system quietly guides learners towards a pre-shaped outcome regardless of what they do.

The paper Learning to Assist Humans without Inferring Rewards develops this idea in the context of assistive agents. Instead of asking an AI system to infer a human’s true goal, Myers et al. (2024) explore assistance through empowerment: the assistant should increase the human’s ability to influence future outcomes. Earlier work on assistance via empowerment points in a similar direction, framing assistance as a way to increase human control over future possibilities rather than simply optimise a task on the human’s behalf (Du et al., 2020).

For education, this is a powerful shift. The goal of an AI tutor should not simply be to guess what the student wants or make the task easier. It should increase the learner’s ability to make meaningful decisions about the learning process. A useful test is whether the AI leaves the learner with more room to think and choose, not just a cleaner final answer.

This perspective leads to concrete design principles. AI systems should ask learners to articulate their own reasoning before offering solutions. They should make uncertainty visible, give students room to compare different approaches and know when to step back so that students take on more of the thinking themselves.

Teachers remain central in this picture. AI can support explanation, feedback and practice, but teachers are responsible for the broader educational judgement: what students should struggle with, when help becomes too much and which skills are really being developed. Preserving learner agency therefore also requires preserving teacher agency. If AI systems quietly take over the structure of learning, then both students and teachers may lose meaningful control.

So the alignment question is not only whether the AI gives correct or safe answers. It is also whether the system fits into human learning practices without displacing the forms of agency those practices are meant to cultivate. A well-aligned educational AI should not make students dependent on better answers. It should help them become better authors of their own learning trajectories.

References

Du, Y., Tiomkin, S., Kiciman, E., Polani, D., Abbeel, P., & Dragan, A. (2020). AvE: Assistance via empowerment. Advances in Neural Information Processing Systems, 33.

Kasneci, E., Sessler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., Gasser, U., Groh, G., Günnemann, S., Hüllermeier, E., Krusche, S., Kutyniok, G., Michaeli, T., Nerdel, C., Pfeffer, J., Poquet, O., Sailer, M., Schmidt, A., Seidel, T., Stadler, M., Weller, J., Kuhn, J., & Kasneci, G. (2023). ChatGPT for good? On opportunities and challenges of large language models for education. Learning and Individual Differences, 103, Article 102274. https://doi.org/10.1016/j.lindif.2023.102274.

Klyubin, A. S., Polani, D., & Nehaniv, C. L. (2005). Empowerment: A universal agent-centric measure of control. Proceedings of the 2005 IEEE Congress on Evolutionary Computation, 128–135. https://doi.org/10.1109/CEC.2005.1554676.

Molenaar, I. (2022). Towards hybrid human-AI learning technologies. European Journal of Education, 57(4), 632–645. https://doi.org/10.1111/ejed.12527.

Myers, V., Ellis, E., Levine, S., Eysenbach, B., & Dragan, A. (2024). Learning to assist humans without inferring rewards. arXiv. https://arxiv.org/abs/2411.02623.

Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002.

Salge, C., Glackin, C., & Polani, D. (2014). Empowerment: An introduction. In M. Prokopenko (Ed.), Guided self-organization: Inception (pp. 67–114). Springer. https://doi.org/10.1007/978-3-642-53734-9_4.

Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 17(2), 89–100. https://doi.org/10.1111/j.1469-7610.1976.tb00381.x.

Zimmerman, B. J. (2000). Attaining self-regulation: A social cognitive perspective. In M. Boekaerts, P. R. Pintrich, & M. Zeidner (Eds.), Handbook of self-regulation (pp. 13–39). Academic Press.

Further Reading:

Disempowerment patterns in real-world AI usage https://www.anthropic.com/research/disempowerment-patterns.

Kasneci, E., Sessler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., … & Kasneci, G. (2023). ChatGPT for good? On opportunities and challenges of large language models for education. Learning and Individual Differences, 103, 102274. https://doi.org/10.1016/j.lindif.2023.102274.

Molenaar, I. (2022). Towards hybrid human-AI learning technologies. European Journal of Education, 57(4), 632–645. https://doi.org/10.1111/ejed.12527.

Image Attribution

Generated by: ChatGPT 5.5

Date: 02 June 2026

Prompt: “Create a 1920×1080 landscape image as an artistic oil painting in the spirit of early 20th-century expressionist abstraction, loosely inspired by the Blue Rider movement and Kandinsky’s use of dynamic colour, geometric rhythm and symbolic composition. Do not copy any specific existing artwork.

The image should represent the theme of preserving learner agency when AI becomes a learning partner in education. Show a human learner at the centre of the composition, not passively receiving an answer, but actively shaping their own learning path. Around the learner, suggest the presence of an AI assistant as an abstract conversational form: luminous lines, geometric shapes, branching paths or subtle machine-like patterns. The AI should feel supportive but not dominant.

The visual metaphor should be about balance: guidance without control, assistance without dependence, and technology expanding rather than replacing human judgement. Include motifs such as diverging pathways, open books, faint dialogue shapes, interconnected nodes or shifting forms that respond to the learner’s movement. The learner’s own choices should appear to influence the direction of the composition.

Use rich oil-paint texture, visible brushstrokes, layered colour, expressive movement and a thoughtful educational atmosphere. The palette should combine warm human tones with cooler technological blues and greens, with enough contrast to make the central learner visually important. The mood should be reflective, hopeful and slightly uncertain, not utopian or dystopian.

Avoid photorealism, corporate illustration, flat vector style, robots, classrooms full of devices, screens with readable text, logos or overly literal AI imagery. No text in the image. Composition should leave some calmer visual space near the top or side so the image can work as a website header.”