How to Use AI Without Letting It Think for You

SIGNAL OVER NOISE

Artificial intelligence can expand human intelligence—or quietly disable it.

A study involving nearly 1,000 high-school students in Turkey demonstrates the difference. Students practicing mathematics with a ChatGPT-style assistant performed dramatically better during practice: their scores increased by 48 percent. But when the AI was removed for the exam, they scored 17 percent lower than students who had practiced without it. A more conservative, preregistered analysis found a somewhat smaller decline, but the central result remained: students appeared to perform better with AI while learning less.

Many had simply asked the system for answers and copied them.

This exposes one of the most important distinctions of the AI era: successful completion is not the same as learning.

The students completed more problems correctly. They probably felt more capable. The visible evidence suggested progress. But the intelligence responsible for that performance was partly outside them. Once the external intelligence disappeared, so did the apparent improvement.

AI had functioned as an answer machine—not a learning machine.

The Disappearing Struggle

Learning requires cognitive effort. We have to retrieve information, tolerate confusion, make predictions, test possible solutions, recognize mistakes, and try again. Those moments of difficulty are not unfortunate obstacles standing between us and learning. They are often the mechanism through which learning occurs.

Answer-giving AI can remove that mechanism.

When a system instantly summarizes the reading, writes the code, solves the equation, or produces the essay, it creates an extraordinarily convincing simulation of competence. The assignment is finished. The answer may be correct. The user experiences relief and forward motion.

But whose intelligence produced the result?

That is the question we should ask every time we use AI: Am I using this system to amplify my thinking, or to avoid thinking?

I call AI an intelligence amplifier. Its highest purpose is not to replace human intelligence but to extend its reach—to help us notice patterns, examine assumptions, explore alternatives, make connections, and understand ideas that might otherwise remain inaccessible.

An amplifier, however, needs a signal.

If I bring curiosity, judgment, experience, and an emerging idea into the exchange, AI can strengthen all of them. If I bring nothing and ask it to produce the finished answer, it may generate an impressive output without strengthening me at all.

A Different Kind of AI Tutor

The Turkish study also tested another version of the assistant. Instead of simply supplying answers, it provided hints and guided students toward the solution. With that version, students’ practice performance improved substantially, but the later exam penalty disappeared.

That difference matters.

The problem is not AI itself. The problem is how the interaction distributes the thinking.

An AI tutor should not immediately eliminate confusion. It should make confusion productive. It can ask:

• What have you tried?

• What do you think the answer might be?

• Which part do you understand?

• Where exactly did you get stuck?

• What principle might apply here?

Instead of producing the entire solution, it can offer one clue. If that is not enough, it can offer another. After explaining the answer, it can ask the learner to apply the same idea to a new problem without assistance.

Used this way, AI provides something few students have ever had: a patient, responsive tutor available at any hour, willing to explain the same concept in five different ways without embarrassment, impatience, or judgment.

That is revolutionary. But only if the learner remains intellectually present.

Keep the Intelligence on Your Side

The most useful rule may be the simplest: before asking AI, attempt an answer.

The attempt does not have to be good. It can be confused, incomplete, or completely wrong. Its purpose is to activate your own reasoning before encountering the system’s reasoning.

Then use AI to interrogate the gap:

• What did I understand correctly?

• Where did my reasoning break down?

• Why does the correct approach work?

• Can you explain it with a different example?

• What assumption am I missing?

• Give me a similar problem, but do not solve it for me.

Afterward, close the chat and reconstruct the idea independently. Explain it in your own words. Write the code yourself. Solve a comparable problem. Return to it several days later without notes and see what remains.

That final step is essential because AI-assisted fluency can be deceptive. A conversation may feel clear while it is happening because the model is carrying the structure. The real test is whether you can reproduce the reasoning when the scaffolding is gone.

Answer or Ability

There is nothing inherently wrong with asking AI for an answer. Sometimes the answer is all we need. If I need to translate a sentence, format a document, locate an error, or complete a task I will never need to perform independently, efficiency may be the appropriate goal.

But when the objective is education, the question changes.

Do I need the answer—or do I need the ability?

If I need the ability, AI should not always make the work easier. It should make the work more fruitful. It should help me operate slightly beyond my current capacity while ensuring that I remain the person observing, deciding, attempting, and understanding.

This distinction will become increasingly important as AI enters classrooms, workplaces, and professional training. We cannot evaluate these systems only by asking whether people complete tasks faster or produce better immediate results. We must also ask what happens when the system is removed.

• Can the student still solve the problem?

• Can the employee still exercise judgment?

• Can the programmer understand the code?

• Can the writer defend the argument?

• Can the professional recognize when the AI is wrong?

If not, the system has not amplified intelligence. It has created dependence while disguising that dependence as competence.

The great promise of artificial intelligence is not that human beings will no longer have to think. It is that we may be able to think further, faster, and more expansively than before.

But that promise is not automatic. We have to protect the human role in the exchange.

AI can supply information, explanation, challenge, feedback, and perspective. It can illuminate the road and help us navigate difficult terrain. But if we want knowledge that remains after the machine is gone, we still have to do the walking.