A question worth thinking about
Does AI make us
better thinkers?
A better answer on the screen. But what changes in the person?
The short answer
It can. Better output and better thinking are different achievements.
AI can help people perform some tasks better, and carefully designed tutoring can improve learning. It can also help someone finish an exercise while leaving them less able to solve the next one alone. [1] [2] [3]
Our interpretation: use AI to challenge and develop your reasoning, and check what you can still explain when the assistance is gone.
What do we mean by “better”?
Writing a stronger report, learning a concept and judging whether a claim is true are different outcomes. A tool can improve one without improving the others. Before asking whether AI helps, decide which outcome you care about.
Performance with help
Can you produce a better result while using the tool?
Learning that stays with you
Can you explain, apply and question the idea independently?
Follow the evidence
Three studies. Three different tests.
These are selected studies, not a systematic review. Their results belong to particular people, tasks and AI systems; they are not a verdict on every use of AI.
01 / Task performance · Randomised experiment
AI helped on some consulting tasks and hurt on another.
In an experiment with 758 consultants, GPT-4 users completed suitable tasks 25.1% more quickly on average. On a separate task designed to exceed the model’s capabilities, AI users were about 19 percentage points less likely to reach the correct solution.
The limit: this measured work produced with assistance, not lasting improvements in the consultants’ thinking. The negative result involved one deliberately selected task.
Dell’Acqua et al. · Organization Science, 2026 Read the original consulting study02 / Independent learning · Randomised experiment
Better practice scores concealed weaker independent performance.
In a Turkish high-school maths study involving nearly 1,000 students, those given a general GPT-4 interface scored 48% higher during assisted practice but 17% lower on subsequent unaided exams than the control group. A tutor with teacher-designed safeguards largely mitigated that harm.
The limit: this was one school and short-term maths performance. The guarded tutor did not show a statistically significant unaided exam benefit; that does not prove equivalence or improvement.
Bastani et al. · PNAS, 2025 Read the original mathematics study03 / Structured tutoring · Randomised crossover study
A purpose-built tutor improved learning on specific physics lessons.
A Harvard undergraduate physics study involving 194 students compared a structured AI tutor with active classroom learning across two lessons. Students scored higher on the tests after the AI-tutored lessons.
The limit: a carefully designed tutor is different from an unrestricted chatbot. Immediate tests on two topics do not establish long-term retention, wider critical-thinking gains or a replacement for teaching.
Kestin et al. · Scientific Reports, 2025 Read the original physics tutoring studyWhat these studies leave open
Whether everyday AI use strengthens judgement over months or years; whether learning transfers to unfamiliar problems; and which designs work for which learners. Those are further questions, not conclusions established here.
Choose a direction
Where does your curiosity take you?
Make the strongest case for AI
An argument to examine
Thinking need not happen entirely inside one person’s head. Writing and calculation tools already help us tackle problems we could not manage unaided. If AI helps you generate alternatives, find objections and test assumptions, the person working with the tool may reason better.
The challenge is to show that the combined judgement is reliable. A persuasive explanation alone is insufficient. The consulting study illustrates why the task matters.
Examine the strongest objection
An objection to examine
What if the judgement needed to check an AI answer is the very skill you stop practising? Delegation could make today’s task easier while weakening your ability to spot tomorrow’s error. The maths experiment gives a concrete reason to take this possibility seriously.
This objection is strongest when developing independent skill is the goal. It does not establish that all delegation is harmful, or that every AI design has the same effect.
Apply this to the way you use AI
A practical inference to try
- Attempt: write your own answer and where you are uncertain.
- Challenge: ask AI for a hint, an objection or a counterexample.
- Check: inspect the evidence behind an important claim.
- Transfer: close the tool and tackle a new example yourself.
Treat this as a routine to test, not a proven recipe. The useful signal is what you can subsequently explain and do, rather than how fluent the exchange felt.
Pause for 30 seconds
What would convince you that learning happened?
A colleague writes a much stronger report using AI. They conclude: “This tool is making me a better thinker.”
What extra evidence would you want before accepting that conclusion?
Compare your reasoning
First clarify the claim. If they mean better work with AI, independently assessing the report’s accuracy and reasoning would help. If they mean improved personal skill, ask them to explain its reasoning, spot a planted error, or solve a fresh problem without the tool.
A later test and a comparison with their earlier ability would strengthen the evidence. Attributing the change specifically to AI would also require a fair comparison with learning without it. One impressive report cannot settle all of those questions.