How Should You Learn AI? What the Evidence Says About Effective Learning

Many people imagine that learning AI begins with mountains of mathematics or with memorising ready-made prompts. The available evidence says something simpler and more precise: what counts is what remains in your mind when the tool is absent.

The tool that answers for you may teach you less

In a study published in the Proceedings of the National Academy of Sciences (PNAS) in 2025, Hamsa Bastani and colleagues ran a field experiment with nearly a thousand secondary-school mathematics students. Some were given an assistant resembling ChatGPT, built on GPT-4; others a tutor built on the same model but with instructions that protect learning.

The result: practice performance rose by 48% with the ordinary assistant and by 127% with the designed tutor, but when the ordinary assistant was withdrawn its users' grades fell 17% below those of students who never had it. The designed version largely protected learning. The researchers attribute the harm to students using the assistant as a "crutch" that spares them the effort that fixes a skill in place. For more on education and AI, see the article "The End of the Age of Rote Instruction?".

What makes knowledge stick?

John Dunlosky and colleagues reviewed ten learning techniques in a 2013 study in Psychological Science in the Public Interest. Practice testing and distributed practice, spreading study over time, were described as highly useful, while underlining and rereading were described as of low utility.

In practice: do not merely reread the lesson. Close the book and try to recall it, and spread your sessions over several days instead of cramming them into one night. After a lesson on image classification, for example, write from memory what you understood of its steps and where it might go wrong, then compare with the lesson. The effort of recalling is the real exercise, and an AI assistant can cut it short if you ask it too early. On how fast skills expire, and why the skill of learning has itself become capital, see the article "Learn How to Learn: What Does AI Really Add?".

Start with something that works: a path in five steps

The free fast.ai course teaches "top-down": you begin with a model that works on a real problem, then open the box to understand how it was built, rather than beginning with algebraic symbols. It asks for about a year of coding experience, preferably in Python, and secondary-school mathematics. This is one teaching philosophy among several, and the source does not show it to be superior to others, but it is an example for anyone who wants to learn by doing.

1. Choose a small problem that interests you, such as classifying texts or predicting a simple value.
2. Build a modest solution that works, even if it is incomplete.
3. Understand what you built: why did it err? What data did it learn from?
4. Test yourself without help after each session, then review a few days later.
5. Use the language model as a tutor: ask it for hints and questions, not the full solution.

In Saudi Arabia

The Ministry of Education announced, in an item published in 2025, that the Samai initiative, begun by SDAIA with the assistance of the ministries of Education and of Human Resources and Social Development, had exceeded its target of empowering one million citizens in AI, with about 70% of participants employed and 30% students. This is a welcome widening of awareness. But a certificate is a beginning; the measure is what you can build and explain.

A conclusion you can test

Learning is a blessing that needs patience and regularity. Knowledge is sought through effort, and success in it is hoped for those who take the means and ask God for ease. Set aside a regular hour, and for one week try testing yourself before you ask the tool. See the difference.