Is ChatGPT Making Us Mentally Lazy? What the Research Actually Says

Research raises useful questions about effort and learning. It does not justify a blanket claim that ChatGPT is damaging everyone’s ability to think.

Start by defining the worry

Someone asks a chatbot to draft a routine email. Someone else copies a solution they were meant to learn. Both avoided work. Only one necessarily avoided the practice that was the point of the task. Calling both mentally lazy makes the conversation louder and the diagnosis less precise.

There are at least three possible concerns: using less effort right now, developing a habit of accepting answers, and losing a skill over time. These require different evidence. A report about effort cannot establish long-term skill loss by itself. A disappointing exam result cannot explain every kind of professional AI use.

What the much-discussed worker survey found

A 2025 CHI study of 319 knowledge workers collected examples of workplace AI use and self-reports about critical-thinking effort. Greater confidence in AI was associated with less reported critical thinking; greater task-specific self-confidence was associated with more. Participants also described verification and integration as important parts of their work.

This is evidence about reported experiences and associations. It did not randomly assign years of AI use, repeatedly test general reasoning ability, or show that a person’s brain deteriorated. Self-reports can illuminate workflows while leaving causality and measurement unresolved. Those limits belong beside the finding, not in tiny print beneath an alarming headline.

A learning experiment gives a different kind of evidence

The Bastani and colleagues classroom study tested GPT-4-based math support with different designs. Its standard interface produced poorer unassisted outcomes than the control group, while safeguards largely mitigated that harm. Crucially, the safeguarded condition did not demonstrate a positive unassisted-learning effect. The published correction concerned affiliation details.

Random assignment makes this stronger evidence about that intervention than a simple survey. It still cannot establish what happens to every adult who uses a chatbot. Subject, age, assessment, prior instruction and tutor design matter. The sensible response is to examine how assistance changes practice, rather than announce a universal personality flaw.

Less effort can be a successful design outcome

A tool that makes a difficult document readable may free a reader to examine the argument. Dictation may let someone spend attention on ideas rather than typing. A summary may help a busy worker decide which full report requires reading. Reduced effort on one operation can leave more room for another.

The useful question is which operation disappeared. If software handles transcription while you compare competing explanations, your reasoning may remain active. If it selects the explanation, supplies the evidence and writes the conclusion before you engage, little independent judgment may be left. A single time-saved measure cannot distinguish those workflows.

Watch for an acceptance habit

Try observing a behavior rather than labeling yourself. Do you submit an answer you cannot explain? Do you treat a citation as real without opening it? Do you ask for a judgment before identifying your priorities? These are specific habits you can change even while the long-term research remains unsettled.

For a week, keep a brief tool-use log: task, purpose, what the tool supplied, what you checked. Include successful uses. Otherwise the exercise becomes a search for guilt. The goal is to discover where assistance expands your work and where you are letting its appearance of completeness end an inquiry prematurely.

Preserve the operation you want to learn

If you want to learn an argument, reconstruct it before requesting a polished version. If you want to learn a method, attempt a small problem before asking for a hint. If your job is to evaluate evidence, open the sources and compare the claims yourself. Assistance can follow these actions without eliminating them.

A prompt asking for hints is a useful request, not a guarantee that the system will behave like a carefully engineered educational tutor. Read what it actually returns. If it reveals too much, set it aside and choose a new example. Your workflow needs a recovery step as well as a good intention.

A small experiment you can run on your own workflow

Choose two similar, low-stakes learning tasks. On one, request an immediate explanation. On the other, attempt an answer before consulting help. Later, try a related question with the explanations closed. Record both the time spent and what you can explain. This is a personal observation, not a controlled scientific study.

The comparison can still expose a useful practical issue. Perhaps the first approach helps when you lack basic vocabulary, while the second helps when you already have enough knowledge to begin. Perhaps fatigue or task difficulty explains the difference. Avoid treating one result as proof. Use it to formulate a better question about when assistance is serving your purpose.

A better question than “Am I becoming lazy?”

Ask: what did I stop practicing, and do I still need that skill? You may deliberately stop drafting repetitive status updates from scratch. You may deliberately keep checking arithmetic, interpreting evidence or outlining original arguments. The distinction is practical and personal; it does not require pretending every mental operation deserves equal time.

Our guide to cognitive offloading explains the wider pattern. For a small test today, attempt a Thinking Gym challenge independently, then use assistance only to examine an alternative approach. Notice the specific contribution each stage makes.

Audit one shortcut

Choose a recent AI-assisted task and identify its actual goal.

List the operations you delegated: recall, drafting, calculation, evidence evaluation or decision.

Circle one operation you still need to be able to perform independently.

Design a two-minute independent check for that operation on your next task.

Sources and context

The impact of generative AI on critical thinking (Lee et al., CHI 2025) — Survey of 319 knowledge workers; self-reported effort and associations, not a causal test of declining skill.

Generative AI without guardrails can harm learning (Bastani et al., PNAS 2025) — Randomized classroom study in one Turkish high school; tool design and unassisted assessment matter. The published correction concerns an author affiliation.

Published by CreativePulse. Research, writing and production assisted by AI. The practical exercises are editorial suggestions; research limits are stated in the article. Editorial standards.

Continue reading

Are We Getting Smarter With AI, or Better at Looking Smart?

What Is Cognitive Offloading, and When Does It Become a Problem?

Try a short Thinking Gym exercise. Your response stays in your browser.