How to learn in the age of AI / SI

Sources checked 19 October 2026. Noura is an imagined graduate, and her assignments, notebook and the exercises are illustrations. The recommendation is an editorial position, not a universal finding from the studies. SI refers to hypothetical superintelligence. Practice schedules are suggestions.

Imagine Noura, a new graduate, preparing her first conversion report. Last month, 4% of visitors bought something. This month, 5% did. An AI assistant gives her the calculation and a sentence for the report: conversion rose by one percentage point, a relative increase of 25%.

She checks the numbers against the explanation. They make sense. She sends the report.

A week later, another rate moves from 6% to 9%. Her colleague asks how large the improvement is. “Three per cent,” she says, then hesitates. He writes a question beside her calculation: “Three per cent of what?”

The first report had no arithmetic error. Noura had followed every line on the screen. She had also managed to finish it without choosing a denominator herself. The second assignment exposed something the first one never required her to do.

If you are a junior or a fresh graduate, my advice is to keep generative AI out of your learning during the foundation stage. That includes explanations, hints, study modes, code completion and debugging. Use courses, books, documentation and people who can teach you. Give yourself the chance to encounter the question in Noura's margin before a colleague has to ask it.

What the studies found

Bastani and colleagues' 2025 trial at a Turkish high school found that students using a general GPT-4 assistant performed 48% better on assisted maths practice than the control group. On a subsequent exam without help, they performed 17% worse. A tutor with learning safeguards largely avoided that exam penalty, without an established exam benefit. [1]

The strongest objection to a blanket rule comes from successful AI tutoring. Kestin and colleagues studied 194 Harvard physics students across two lessons. Students using a deliberately designed AI tutor achieved higher immediate test scores than those in an active-learning class. The result does not establish long-term retention, but it rules out saying that AI assistance always prevents learning. [2]

I still favour a clear rule for beginners. Advice such as “use AI only when it helps you understand” asks the learner to judge whether an explanation has produced understanding. Noura thought she understood the first calculation. Keeping the chatbot closed removes that decision while she is learning how to judge her own work. This is my reason for recommending it, rather than a conclusion established by the experiments.

The rule needs a teaching arrangement behind it. Someone must choose material at the right level, provide examples and review attempts. A beginner left alone with an impossible assignment will not acquire the missing prerequisites by being stubborn. If a line makes no sense, get instruction. If the task is too large, make it smaller.

Books and human explanations can be consumed passively too. Closing the chatbot will achieve little if you replace it with two hours of watching solutions. The course needs places where you stop, attempt something and find out what you missed.

Maths: account for the step you took

Noura's colleague asks her to write both answers to the new problem. The increase from 6% to 9% is three percentage points. Relative to the old rate, it is 3 ÷ 6 × 100 = 50%. She writes “old rate” beneath the 6.

Then he reverses the numbers. A fall from 9% to 6% is also three percentage points, but the relative decrease is 3 ÷ 9 × 100 ≈ 33.3%. The denominator has changed. She needs to explain why before doing another example.

These are deliberately small calculations. They leave little room to hide a mistaken choice behind a long solution. Try the discount or unit-cost version below if rates already feel familiar. The question about the base will follow you into both.

For algebra, use the same habit of accounting for each operation. In 3x + 6 = 18, subtract 6 from both sides to get 3x = 12, then divide both sides by 3 to get x = 4. Subtracting the same amount from equal quantities preserves their equality. “Move the six across and change its sign” is a convenient shorthand once you understand that reason.

Write the reason beside the line. Check by substitution: 3 × 4 + 6 = 18. Then change the problem. In 5x − 10 = 20, you must add 10 to remove the negative constant. A person copying the previous operation will subtract again.

If you are new to this, you can start with some steps already worked. Renkl, Atkinson and Große studied gradually removing solution steps as learners moved from examples to problem solving. [6] Our exercise applies that idea: study the full solution, finish a partially worked one, then try a different equation. The activity itself has not been tested as a teaching programme.

Once those steps feel easy, inspect this shorter solution: x² = x; divide both sides by x; therefore x = 1.

Substitute 1 and it works. The answer still misses something.

Zero also satisfies the original equation: 0² = 0. Dividing by x required x ≠ 0, so that route excluded zero before it had been considered. Factoring instead gives x(x − 1) = 0, which retains both possibilities: x = 0 or x = 1.

This is why checking the answer you found and checking whether you found every answer are different jobs. As the maths gets harder, record the conditions beside the operations. In a proof, name the definition or theorem that permits a step. In a probability calculation, write down what is being counted and under which assumptions.

Keep the wrong line. A teacher can do more with “I divided here and lost zero” than with “I don't understand algebra.” If even the worked solution is unreadable, ask which earlier topic you need. A problem with dividing both sides may begin with fractions, several pages back.

Code: investigate the order worth exactly 20

Noura's next assignment is a small program that adds orders worth at least 20. This is a useful first project because the requirement fits in a sentence and the result can be calculated by hand. She can keep the interpreter, debugger, test runner and documentation open. AI generation and AI debugging stay off.

Here is one implementation:

def eligible_total(amounts, minimum):
    total = 0
    for amount in amounts:
        if amount >= minimum:
            total += amount
    return total

Before running eligible_total([12, 20, 35], 20), write the result you expect and track total after each item. If you have never written a loop, start with an introductory lesson before this exercise. Python's own tutorial explains for and if, but explicitly assumes some general programming knowledge. It is useful as a reference, not a complete first course. [8]

Now replace >= with >. With [12, 35], the changed program returns the same answer as before. It appears to work. Noura needs an order worth exactly 20 to discover what she excluded.

Write a one-line test case on paper: input [20], minimum 20, expected result 20. The altered program returns 0. The requirement says “at least,” so the mistake is in the comparison. A larger sample would take longer to inspect without making the mistake any clearer.

There is another experiment worth doing. Move total = 0 inside the loop and predict the result before running it. Each iteration now discards what was accumulated before it; with an empty list, the variable never receives a value. Then restore the code and write a function that counts qualifying orders instead of adding their amounts. You will need a different update to the accumulator.

The value of this work is visible in the questions you can answer afterwards. Which input reveals the bug? Why does the initial value sit outside the loop? What must change when the requirement changes? Keep a record of the answer and the example that convinced you.

A 2026 preprint by Shen and Tamkin gives one reason to take this practice seriously. They studied 52 developers learning the unfamiliar Python library Trio. Rounded immediate quiz averages were 50% with AI assistance and 67% without it; debugging showed the largest gap. The difference in completion time was not statistically significant. This was a small, short-term study, not evidence of permanent skill loss. [3]

Programming teachers have a sequence for work like the exercise above: predict, run, investigate, modify and make, or PRIMM. Sentance, Waite and Kallia reported better post-test results for students taught this way in a classroom study without random assignment. [7] It supports considering that teaching sequence, with the limits of a school study kept in view.

Build outward gradually. Five expenses in a file are enough for a next project: read them, validate them and report the total. Try a missing value, a malformed number and an empty file. Write expected results from the requirement before running the code. If the test repeats your mistaken assumption, both the test and the program can agree and still be wrong.

How to organise the learning

Noura now has something worth keeping from her assignments: a small record of decisions she got wrong. It might look like this.

A useful study session should leave enough working to make an entry like one of these. Here is how to arrange it.

Choose a course with a next page

Pick one sequence and follow it long enough to encounter its exercises. OpenStax's free Elementary Algebra 2e has worked examples and practice problems; Harvard's free CS50P course is designed for people with or without programming experience. [10][11] Use the lessons and assignments, leaving AI assistants off. The Python documentation can answer a specific language question once you have a starting vocabulary.

You do not need to collect five courses before beginning one. Choose a chapter, an exercise and a place to get feedback. If you cannot access a teacher, use checked solutions and the course's discussion community. Show a narrow question with your attempt; a stranger can help more readily with one failing comparison than with an entire unfinished project.

Make a prediction you can check

Write the output, the next algebraic step or the denominator before looking at the answer. If a blank page stops you, study one example and complete the last step of another. You can increase the amount you do yourself as the material becomes familiar.

A prediction gives the review something to examine. “I expected 20, but got 0” is more useful than “the program is confusing.” You can now locate the first point where what happened differed from what you expected.

Ask for the missing explanation

Use the reference, then bring the unresolved step to a teacher, colleague or study partner. If the notation is new, ask for an explanation before continuing.

I expected [result], but got [actual result]. Here is the smallest example that still fails. I checked [reference] and tried [attempt]. Can we go through [specific step], then let me finish it?

Afterwards, make the correction yourself and attempt a different example. A helpful explanation should be followed by something you do with it.

Leave a task for another day

Before ending the session, choose a new problem for your return. Attempt it before reopening the worked solution. Roediger and Karpicke found benefits from recall on delayed tests of studied passages; Dunlosky and colleagues' review rated practice testing and distributed practice highly. [4][5] Neither study establishes the exact minutes or dates in our planner.

Use the templates as a starting arrangement. Change the time to fit the task, and leave room for a review if you get stuck.

At work, this requires an agreement with your manager. Ask for a task you can complete and review, plus a time to discuss the part you cannot explain. Managers who prohibit AI during learning also need to provide instruction and time. Rewarding only the speed of the finished output makes the learning arrangement difficult to sustain.

You can assess progress without declaring yourself an expert. Solve a changed problem, explain a choice and locate an error across several sessions. A fresh graduate might manage those steps for a loop while still needing a full lesson on recursion. Plan from the skill in front of you; the job title does not describe every gap.

If a future machine can do the whole assignment

SI, or superintelligence, refers here to a hypothetical system with intellectual abilities far beyond human experts. [9] The learning studies above did not test such a system. They cannot settle what jobs would remain or when that future might arrive.

Consider a narrower thought experiment. Suppose tomorrow's assistant calculates every rate correctly and produces faultless code. Noura can still have questions about the assignment. What counts as a visitor? Should cancelled orders be included? Did conversion rise because the product improved, or because different people arrived? Correct arithmetic alone does not answer those questions. A better system might help investigate them, but handing her its answer would still tell us little about what she can investigate herself.

Learning maths gives her a way to examine quantities, assumptions and evidence. Learning programming lets her turn a rule into behaviour she can inspect. These are reasons to study the subjects while she is responsible for work that uses them. They are not a promise that any particular skill will guarantee a job.

Return to the notebook, with the chatbot closed. Change Noura's requirement to “orders worth at least 50.” Write down what should happen for [49, 50, 51] and for an empty list. Then change the code, run it and explain any disagreement before you move on.

Sources

• [1] Bastani et al., Generative AI without guardrails can harm learning (PNAS, 2025). Relative effects, each against its own control outcome. Author manuscript.

• [2] Kestin et al., AI tutoring outperforms in-class active learning (Scientific Reports, 2025). Designed tutor; two lessons; immediate assessment.

• [3] Shen & Tamkin, How AI Impacts Skill Formation (2026 preprint). Authors’ report. The rounded means differ by 17 percentage points. Prompting-pattern observations were exploratory, not causal.

• [4] Roediger & Karpicke, Test-Enhanced Learning (2006). Retrieval and delayed retention of prose.

• [5] Dunlosky et al., Improving Students’ Learning With Effective Learning Techniques (2013). Review of ten learning techniques.

• [6] Renkl, Atkinson & Große, How Fading Worked Solution Steps Works (2004). Experiments on gradually removing worked steps; not an AI trial.

• [7] Sentance, Waite & Kallia, Teaching computer programming with PRIMM (2019). Nonrandom classroom comparison; predates AI tutoring. Author copy.

• [8] Python documentation, More Control Flow Tools. Tutorial audience: programmers new to Python, not complete programming beginners. Language reference, not learning-outcome evidence.

• [9] Alfonseca et al., Superintelligence cannot be contained: Lessons from Computability Theory. Cited for its definition of hypothetical superintelligence, not as evidence about learning.

• [10] OpenStax, Elementary Algebra 2e: solving equations. Free course text with worked examples and exercises.

• [11] Harvard, CS50’s Introduction to Programming with Python. Free OpenCourseWare; designed for students with or without prior programming experience. A resource suggestion, not a claim of superior learning outcomes.