Generative AI and human learning

Generative AI (ChatGPT and the like) is taking an increasingly important place in our lives, starting from a very young age. Recent work suggests that a naive usage of these tools can improve immediate performance but have a detrimental effect on learning (Kosmyna et al., preprint, Bastani et al., 2025).

The aim of this post is not to criticize generative AI as such, but rather to think about how it can be used in the context of human learning, in particular for young ones. Here are my thoughts on the subject. As a teacher, these reflections are oriented around college students, but I think they are also valid for younger students in school or adults learning by themselves.

TL;DR

AI should not replace cognitive effort, but only accompany it.
AI should not replace the learning environment (human interaction), but only supplement it.
The teacher should not only deliver knowledge, but also focus on social development (personal and group life).

Problem statement

Here is a student profile I observe quite frequently: seeing the answer key as the holy grail of a tutorial session, even if it means bypassing the prior and necessary cognitive efforts. This profile was already deleterious before the democratization of generative AI, but I fear it is begin reinforced because the answer key becomes accessible with a simple click via prompts such as:

Give me answer keys for this tutorial sheet.

Help me with my homework.

At the time of writing this post, the second prompt appears in an image of ChatGPT webpage for college students.

The “Why?” question

Regarding generative AI, or any tool, the main question is: “Which human skill does tool X replace?”

Throughout history, the answer to this question was generally simple and concerned ungratifying skills (which is less and less true with generative AI). The answer to this question should be more or less universal (for instance GPS replaces the human sense of direction).

That first question is in fact a preliminary question to a second one, which is more personal: “Do I want to preserve skill Y?”

The answer is now deeply personal and depends on the skill in question. Let’s take the example of the sense of direction. Some people want to preserve this ability by limiting the use of GPS. Others, however, may not mind losing this skill since their primary goal is to reach their destination. This second choice reduces cognitive load and allows them to focus on other skills (cognitive offloading).

I think that we should keep this question in mind when using generative AI in general and in the context of learning in particular.

Going back to the student profile I sketched above, it seems that their primary goal is to pass their academic year and they think that the most efficient way to do so is via the answer key of tutorial sheets and previous exams. However, I hope that, for most of them, the main reason for going to college is not just to pass the year, but also to learn, acquire knowledge and be able to apply it. If so, generative AI must not be used in a naive way, but rather used with preservation of cognitive effort in mind.

Summary: When using generative AI, one should always ask oneself: “Which human skill does this tool replace?” and “Do I want to preserve this skill?”. Hence, no human skill should be replaced without a conscious choice.

The “Where and when?” question

Unlike human beings (teachers, peers, etc.), generative AI is available everywhere and all the time. I think that this could drive students to isolate themselves and must be fought against seriously.

On the other hand, college and tutorial sessions provide a work environment with social interactions. I believe it is essential to human learning, development and well-being in general. This human-centred perspective is also consistent with UNESCO’s Guidance for generative AI in education and research. Therefore, students’ questions and discussions should be addressed to teachers and peers preferably.

When confronted with an issue, the first step is to identify it and formulate it clearly. Then, we should keep the reflex of asking other human beings for help on that well-formulated issue. If time or human resources are limited, generative AI can be used as a last resort.

Summary: When human help is available, use it. When it is not, prefer to reformulate the problem and ask human later rather than asking generative AI immediately.

The “How?” question

Assume that you want to learn a concept you can remember or a skill that you can apply (and not just a way to make AI reproduce that skill). The natural next question is: “How can I use generative AI in a way that preserves cognitive effort?”

Here is a typical learning workflow (even before the age of generative AI):

Search for material (books, online resources, etc.) and gather it.
Read the material and try to understand it.
Identify what is understood and what is not. If necessary, go back to step 1 or 2 to clarify what is not understood.

Then, when confronted with an exercise, the workflow continues as:
Try to solve the exercise.
Identify what is done and what remains to be done. If necessary, go back to step 3 to understand how to do it.
Verify the proposed solution. If necessary, go back to step 3 to understand how to correct it.

To preserve cognitive effort, and in turn learning, generative AI must not be used in steps 2 and 4 (hints and questions belong to the identification steps 3 and 5). On the contrary, I believe that it can be used as a last resort in steps 1 (as a search engine on steroids) and 6 (as a critical reviewer of the proposed solution, but beware of hallucinations). Finally, steps 3 and 5 are related to the skill of introspection and critical thinking. They may not be directly related to the skill you want to learn but I think that they are very important for any human being, hence I would recommend to preserve them from generative AI.

Here are some guidelines regarding prompts:

For step 1, the prompt should specify the subject of study and the current level of knowledge. For instance,

I want to study metaprogramming in Julia. I already know the basics of Julia and programming in general. Give me a list of references to study this subject and a reading guide.
For step 6, the prompt should specify the solution to be reviewed and some guardrails. For instance,

Here is an exercise: …

Here is a solution: …

Do not rewrite the solution. Do not provide an alternative solution. Make a critical review of the solution. Point out the steps that are correctly justified, the errors, the missing hypotheses and the points that need to be checked.
When stuck between steps 5 and 6, the prompt should specify the current state of the solution and ask for hints or clarifications. For instance,

Here is an exercise: …

Here is my current solution: …

I know that it remains to: …

Do not give me the complete solution. Ask me a question or give me a hint to help me move forward. Wait for my attempt before helping me further.

Summary: AI should be mainly used at step 1 (reference gathering) and step 6 (critical review) with guardrails to prevent it from polluting the other steps.

The teacher’s role

The teacher’s role is to drive students through the 3 questions (“Why?”, “Where and When?”, “How?”). Since AI usage will become extensive in the coming years, I think that we must provide guardrails to students.

To conclude, I believe that we are living through times where the main role of colleges and teachers is shifting from knowledge delivery to providing a learning environment. To sum up, I think that we should focus on enhancing the social interactions between students (still providing a source of truth and knowledge).

Generative AI disclosure

Generative AI (Copilot) was used as an autocomplete tool to help write this post.

Comment about AI agent skills

The guidelines provided above may be included in agent skills. By the way, I believe that the /grill-me skill can be used prior to step 1 (reference gathering) in order to align the AI agent with the current level of knowledge.

AI & learning