By Costin Liculescu, creator of InsightArea
A software developer encounters a difficult bug. Instead of investigating it, she pastes the error into an AI assistant. The assistant proposes a change, writes the code and generates tests. Everything passes. Ten minutes later, a colleague asks why the fix works and whether it could break another part of the system. The developer cannot explain either.
The immediate task is complete. But something important may not have happened: the developer did not necessarily learn how to diagnose that class of problem.
This distinction matters far beyond software engineering. Students can generate essays without developing an argument. Managers can request recommendations without identifying their assumptions. Anyone can receive a polished explanation and mistake the ease of reading it for genuine understanding.
The problem is not that AI makes work easier. It is that producing a good answer and developing the ability to produce or evaluate an answer are different goals. We should use AI differently depending on which goal matters.
Does AI actually weaken critical thinking? What the research says
In a September 2026 Nature feature on critical thinking in the age of AI, journalist Helen Pearson examined concerns that generative AI may allow learners to bypass the mental work involved in understanding a problem. The article describes research in which access to an AI tool improved students’ performance on mathematics problems while they had the tool, but their subsequent independent performance was worse than that of students who had never used it.
That finding is a warning, not a universal verdict. As the Nature feature makes clear, the evidence about generative AI’s long-term effects on learning and critical thinking is still limited and mixed. Different tools, tasks and ways of using them can produce different outcomes. AI can also support learning when it provides feedback, asks useful questions or encourages active problem-solving rather than replacing it.
Meanwhile, research on learning offers a more established direction. Retrieving information from memory, checking our understanding, considering alternatives and monitoring our own reasoning are valuable learning behaviors. The Education Endowment Foundation’s evidence review rates metacognition and self-regulation as high-impact approaches in school settings, although implementing them well matters and school-based estimates should not be treated as promises for every adult using AI.
What follows is a practical framework inspired by that evidence and the questions raised in Nature. The individual habits below have not all been tested as one unified AI-use program. They are a way to put well-established principles of active learning and careful reasoning into daily practice.
Eight practical habits for staying mentally sharp while using AI
1. Think first, even when your first idea is incomplete
Before opening an AI chat, write down what you think the problem is, what you already know and what you would try first. You do not need a correct answer. You need a starting position that you can examine and revise.
Example: Before asking AI to debug an API timeout, identify whether you suspect network latency, database contention or a dependency. Write one observation that would support each possibility. Then ask AI to challenge your reasoning.
Try this: “Here is my initial explanation and the evidence I have. Identify gaps in my reasoning, but do not give me a complete solution yet.”
Thinking first is especially important when your goal is to become better at the work rather than merely finish it. If the task is routine, your expertise is already established and the result can be reliably checked, full delegation may be sensible.
2. Make AI a Socratic tutor rather than an answer machine
AI assistants are often designed to be helpful by answering immediately. But an immediate answer can short-circuit the very steps a learner needs to practice. You can change the interaction by asking for questions, feedback and progressively stronger hints.
Try this prompt:
“Help me solve this problem myself. Ask one question at a time. If I am stuck, give me a small hint. Challenge unsupported assumptions. Reveal the full solution only after I have attempted it or explicitly ask for it.”
This approach works for understanding algorithms, evaluating an argument, learning mathematics or interpreting a scientific result. It is less appropriate when time-critical professional work requires a verified answer now; choose the interaction based on the stakes and your actual objective.
3. Explain the answer without looking at it
A clear AI response feels easy to understand while it is on the screen. That feeling is not a reliable test of learning. Close the chat, put the source aside and explain the central idea from memory.
Example: You ask AI how a database index speeds up a query. After reading the explanation, answer three questions without help: What does an index change? Why is it useful? When might it make performance worse?
If you cannot explain the mechanism, you have found a specific knowledge gap. Reopen the material, resolve that gap and repeat. This is not punishment; it is a more useful measure of understanding than the feeling that an explanation sounded convincing.
4. Deliberately search for alternative explanations
Critical thinking frequently means resisting the first plausible story. Both humans and AI systems can settle too quickly on an explanation that fits the available facts.
Example: Website conversions drop after a redesign. Was the design responsible? Perhaps traffic quality changed, a tracking script failed or a seasonal effect altered demand. A convincing narrative is not yet a causal explanation.
Try this: “Give me three competing explanations. For each, identify the strongest supporting evidence, what would count against it and the cheapest useful test to distinguish it from the others.”
This habit is valuable in science, engineering, business and everyday judgment. The key is not to invent endless possibilities, but to identify plausible alternatives that could genuinely change your decision.
5. Verify important claims outside the original source
An unfamiliar website can look authoritative. So can an AI answer that confidently cites a study. Formatting, fluency and apparent specificity do not establish reliability.
One useful technique is lateral reading: leave the page and investigate the source from independent places. Who published the claim? What expertise or incentives do they have? Can you locate the original evidence? Do other reliable sources interpret it the same way?
Example: An article announces that a supplement “improves memory by 40%.” Find the actual study, ask what was measured, how many participants were involved, whether the result was replicated and whether the headline reflects the finding. Use AI to help locate and analyze sources, but inspect crucial evidence yourself.
The more consequential the decision, the stronger the verification should be. Not every minor fact requires a literature review; health, financial and professional decisions often require more.
6. Practice retrieval and revisit what matters
Reading a helpful explanation three times is not equivalent to retrieving the idea independently. When you want knowledge you can use later, test yourself without assistance and return to the material after some time has passed.
Example: After studying a new programming concept, solve one small problem from memory. Try a different example the next day and another later in the week. The purpose is not to memorize every detail, but to make central concepts accessible without constantly reopening the AI conversation.
AI can generate practice questions, vary problem types and provide feedback. It should not answer the practice questions for you before you have tried.
7. Develop critical thinking inside specific fields
There is no single abstract “critical thinking” switch that automatically transfers from philosophy to medicine, engineering and investing. Effective judgment depends heavily on domain knowledge.
A good software engineer needs to understand debugging, testing, dependencies and architectural trade-offs. A careful reader of medical research needs some understanding of study design, statistical uncertainty and causal inference. Even a sophisticated general thinker can struggle in an unfamiliar field.
Try this: Identify the recurring judgments your field requires. Practice those judgments on real examples. Ask AI to construct contrasting cases, explain failure modes and critique your answers using the standards of the discipline.
For ambiguous work, this may mean learning to distinguish a missing requirement from a technical obstacle, identifying which decision is reversible, and asking the right person the right question. For a related method, see our guide to starting difficult tasks when the next step is unclear.
8. Periodically test what you can do without AI
AI-assisted performance and independent competence are different measures. If you rely on AI for an activity you genuinely want to master, occasionally remove the assistance and see what remains.
Example: Once a week, debug a small problem or evaluate a short argument without AI. Compare your reasoning, not just whether the final answer is correct. Which assumptions did you test? Did you notice counterexamples? Where did you get stuck?
Then use AI as a reviewer: “Evaluate my solution. What did I miss? Which step was weak? Suggest a different problem that tests the same skill.”
Do not turn this into a rule against productive automation. If your objective is to deliver a routine artifact rather than build a skill, independent rehearsal may add little value. The relevant question is what capability you want to retain.
A simple decision rule: automate the task or practice the skill?
Before using AI, decide which of these situations describes you:
- I need a dependable result. Delegate what the system can do well, and verify the output proportionately to the risk.
- I need to learn the underlying concept. Ask for explanations, hints and practice problems; test your understanding without assistance.
- I need to make an uncertain judgment. Develop your initial view, use AI to explore alternatives, investigate key evidence and retain responsibility for the decision.
- I need to develop professional expertise. Alternate assisted work with independent attempts, feedback and reflection.
This distinction avoids two unhelpful extremes: treating all AI assistance as cognitive damage, and assuming that anything completed efficiently has also been learned.
A 15-minute daily exercise to keep your reasoning active
Choose one genuine question from your work, studies or reading. It should be challenging enough to require thought, but small enough to make progress in one sitting.
- Minutes 1–5: Think independently. Define the question, write your initial explanation and name one uncertainty or alternative.
- Minutes 6–10: Use AI for critique. Ask the assistant to challenge your argument, provide a counterexample or suggest a useful test.
- Minutes 11–15: Reconstruct and revise. Close the assistant and explain the best version of your answer from memory. Record what changed and what you still need to verify.
These timings are a convenient exercise design, not a scientifically established optimum. A complex problem may need much longer; a short daily practice is simply easier to repeat.
If you struggle to start, avoid waiting for a feeling of readiness. A small scheduled attempt is often more useful than promising yourself a perfect study session later. Our earlier analysis of movement, creativity and cognition explores another way to approach demanding thinking tasks.
Frequently asked questions
Does using ChatGPT make you less intelligent?
There is not enough rigorous long-term evidence to make that general claim. Some studies raise concerns about overreliance and reduced independent performance in particular settings; others suggest benefits when AI is used to support active learning. The task, tool and way it is used all matter.
Should students avoid AI when studying?
Not necessarily. AI can provide examples, feedback and explanations. But students should still practice the reasoning and retrieval required to demonstrate independent understanding, and follow their institution’s rules about permitted assistance.
What is the most useful prompt for developing critical thinking?
Start with: “Here is my reasoning. Do not agree automatically. Identify my weakest assumption, ask a question that could change my mind, and help me check the evidence.” The effectiveness depends on your willingness to do the intellectual work, not on the wording alone.
Can AI improve critical thinking instead of weakening it?
Potentially, yes. It can expose alternatives, generate counterexamples and offer targeted feedback. But a productive interaction must still involve the user’s judgment and active processing. The existence of those capabilities does not guarantee better learning outcomes in every setting.
The InsightArea takeaway
The ability to get a polished answer is becoming cheaper. The ability to understand the answer, detect its limitations and recognize when the question itself is wrong remains valuable.
The sensible response is neither to reject AI nor to surrender every difficult cognitive task to it. Use AI to reduce unproductive effort while preserving the forms of effort that build the capabilities you care about. Delegate execution when your goal is productivity; retain meaningful reasoning when your goal is learning or judgment.
That distinction may be more important than any individual prompting technique.
Sources and further reading
- Helen Pearson, “How to stay smart in the age of AI: the science of critical thinking,” Nature, 23 September 2026. Main reporting source and context for the uncertainties discussed.
- Education Endowment Foundation, “Metacognition and self-regulation,” Teaching and Learning Toolkit. Evidence about explicit strategies for planning, monitoring and evaluating learning, principally in school settings.
- Education Endowment Foundation, “Metacognition and Self-Regulated Learning,” guidance report, second edition, November 2025.
Costin Liculescu writes at InsightArea about artificial intelligence, psychology, scientific reasoning and the practical implications of new technologies.
Be First to Comment