How Shall We Learn and Live with AI?

I do not think the future of learning asks us to choose between using AI and protecting our ability to think. The more useful question is how we can arrange the relationship so that assistance helps us become more capable. We need room for convenience, but also room for understanding, uncertainty, and other people.

Across this series, I have returned to one idea: AI can extend thought and accelerate learning, but the knowledge available through a tool does not automatically become the learner's understanding. Accepting that distinction allows us to welcome what the technology offers without asking it to perform the whole work of education.

Begin with what we are trying to do

The appropriate use of AI depends partly on the purpose. Sometimes we need a result: a clearer draft, a translation to help us begin reading, or a first arrangement of information. Sometimes we need to learn a skill that we will have to use again. Sometimes we need to make a decision for which we remain accountable.

These purposes can overlap, but they should not be confused. If I want to learn how an argument works, receiving a finished argument may be less useful than discussing my own attempt. If I simply need help organising notes, delegating some of the arrangement can free attention for the ideas themselves.

We do not need to master every task that a tool can perform. Nor should we assume that a task stops requiring understanding because it can be automated. The useful boundary is where understanding matters to the purpose, the consequence, and our responsibility to others.

Before asking AI for help, one small question can clarify the arrangement: what do I want to be able to do after this conversation? The answer may be modest. I want to explain one concept, identify the next question, or make a decision with a clearer view of the uncertainty.

Preserve an active part for the learner

Chi and Wylie's ICAP framework offers a useful way to think about engagement. It distinguishes passive reception, active engagement with material, constructive generation of ideas beyond what is supplied, and interactive engagement in which participants build understanding together. The framework proposes that these different forms of engagement have different implications for learning.

This is not a study of generative AI. It does, however, encourage a careful question about how we use it: what is the learner actually doing? Clicking through a conversation is not necessarily constructive learning, and exchanging messages with a chatbot does not automatically create the reciprocal learning described by the framework.

A learner might read an explanation, then produce an example that was not supplied, compare it with a counterexample, and discuss where their reasoning breaks down. AI can help generate material for that process. The learner's contribution is more than choosing which finished answer to keep.

We can build such moments into ordinary study without making the process elaborate. Attempt something, receive support, explain what changed, and try a slightly different situation. This is a possible routine, not a universal prescription; different subjects and learners will need different forms of practice.

Good assistance is something we design

There are promising examples of AI supporting learning when the assistance is carefully structured. A 2025 randomised study published in Scientific Reports compared a research-based AI tutor with in-class active learning in a college setting. In the studied arrangement, the AI tutoring condition produced stronger learning outcomes with less time spent.

That finding should not be turned into a claim that any chatbot can replace any teacher. The tutor was designed around educational principles, and the result belongs to a particular setting and comparison. It is valuable precisely because it draws attention to the design of assistance, rather than to access alone.

Read alongside the mathematics study discussed in Part 1, it suggests a productive direction. The question is not merely whether AI is present. We should examine what it encourages the learner to attempt, when it gives an answer, how feedback is provided, and what the learner can do afterwards.

I find that more encouraging than a contest between human and machine. If design matters, we have work we can do: develop better learning arrangements, evaluate them honestly, and revise the parts that do not help. Neither enthusiasm nor fear needs to be the final answer.

Keep people within reach

Learning is also a social experience. A person may need help interpreting a requirement, recovering confidence after a difficult attempt, or deciding how a newly learned skill connects to a possible future. A generated explanation can be useful without meeting every one of these needs.

Teachers, peers, mentors, and colleagues offer forms of attention that belong to particular relationships. They can notice a change over time, understand a shared circumstance, and take responsibility within an educational or working environment. AI can support those relationships; its availability should not become a reason to withdraw them.

This matters when we discuss accessible education. A learner should not be left alone simply because a tool is available around the clock. Access to patient information and access to dependable human support are compatible ambitions.

We should also leave room for a person to ask a question imperfectly. Sometimes someone cannot yet describe the problem because describing it is part of what they are learning. A good educational environment helps shape the question without making them feel inadequate for arriving with it unfinished.

A pathway from learning to a possible future

These questions are also part of the background to an educational business I have been working on recently. I am interested in what helps people connect learning with a next step, especially when access to information does not yet provide a clear route through an unfamiliar environment.

The concern begins before choosing a tool. What does the learner already understand? What support can they reach? Which opportunities can they realistically move towards, and what would help them take the next step? I do not think these questions have a single solution, but they deserve to shape the work from the beginning.

Education becomes more meaningful when learning is connected to a life someone can imagine living. That connection should be made carefully, without promising outcomes we cannot guarantee. A useful pathway offers orientation, support, and room to make informed choices, rather than asking a learner to trust an impressive description of the future.

AI can contribute to this work by widening explanations, helping people explore possibilities, and making some forms of assistance easier to reach. The educational responsibility is to make those possibilities understandable and usable within the person's circumstances.

Learning alongside AI, without disappearing from the process

A sustainable relationship with AI does not require constant resistance to convenience. It asks for a few continuing habits: clarify the purpose, keep an appropriate part of the thinking, check in proportion to the consequence, and know when another person's help is needed. These habits can develop gradually.

We will still misunderstand things. We will sometimes discover that an answer we trusted was incomplete, or that a method we could follow was not yet one we could use independently. Such moments need not disqualify us from learning. They can show where learning should continue.

What I hope for is an education that makes people more able to participate in their own futures. AI can be a powerful companion in that process. Its most valuable contribution may be helping us reach further while leaving us better able to understand where we are going, why it matters, and whom we can learn with along the way.

Sources: Chi and Wylie, “The ICAP Framework: Linking Cognitive Engagement to Active Learning Outcomes” (2014). This framework is used here as a lens for learning activity, not as evidence about AI itself. https://education.asu.edu/sites/g/files/litvpz656/files/lcl/chiwylie2014icap_2.pdf

“AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting,” Scientific Reports (2025). Its results concern the studied tutor, setting, and comparison, not a universal replacement for teachers. https://www.nature.com/articles/s41598-025-97652-6