Key takeaway: No, not in the cartoon "your brain melts" sense. What changes is how often you practice hard cognitive work. Tools that draft, summarize and decide for you shrink the daily workout that used to build skill, so people feel lazier when the real shift is less deliberate practice.
Does using AI make you lazier? No. It makes you practice less, which feels the same from the inside and is a different problem.
I keep hearing the worry phrased as a confession. I think ChatGPT is making me dumb. I can't write an email without it anymore. I used to be able to do this.
The confession assumes something broke. Most of the evidence points somewhere less dramatic and more uncomfortable: nothing broke, you just stopped going to the gym, and the gym stopped sending reminders.
The workout you stopped doing
Hard thinking used to come bundled with the job. You wanted the report, so you wrote the report. You wanted to understand the spreadsheet, so you stared at it until it gave up. The effort was the toll, not a virtue.
Evan Risko and Sam Gilbert defined cognitive offloading in 2016 as using physical action to change what a task demands of your head, so it costs you less.¹ A shopping list counts. So does a phone reminder. So, now, does asking a model to summarize the forty page deck you were supposed to read before the meeting.
Cognitive offloading is not laziness. It is the brain doing what it was built to do: spend effort only when it has to.
In six experiments published in 2010, Wouter Kool, Matthew Botvinick and colleagues let people choose freely between task paths that differed in how much mental control they demanded. People leaned toward the easier path, and the bias held even when they had not noticed which path was harder.² Call it the law of least mental effort. Daniel Kahneman's slow, deliberate System 2 has always been the expensive mode, the one we switch on when fast, automatic System 1 cannot cope.³
So here is the actual shift. For most of history, the task forced System 2 on you. Now a tool can stand in for it on demand, and the task no longer insists.
The Microsoft Research and Carnegie Mellon survey that started much of this conversation caught the shift in people's own words. Hao-Ping Lee, Advait Sarkar and colleagues asked 319 knowledge workers about 936 real uses of generative AI at work. Workers said the tools cut the effort of critical thinking, especially on routine, low stakes tasks under time pressure. Higher confidence in the AI went with less critical thinking. Higher confidence in their own ability went with more.⁴
The data describes reps disappearing, not brains degrading.
What the studies actually measured
The best evidence here does not come from brain scans. It comes from a simple, cruel design: give people AI, then take it away and see what is left.
Hamsa Bastani and colleagues ran that experiment with nearly a thousand high school math students, published in PNAS in June 2025. Students with a standard ChatGPT style tutor did 48% better on practice problems. Then the AI went away for the exam, and those students scored 17% lower than classmates who never had it. A second tutor, built to give hints instead of answers, raised practice scores by 127% and largely avoided the exam penalty.⁵
Read those numbers twice. The tool did not make the students stupid. It did their practice for them, and practice was the whole point.
Anthropic found the same shape in adults this January. Judy Hanwen Shen and Alex Tamkin had 52 mostly junior developers learn an unfamiliar Python library, with or without an AI assistant. The AI group was not reliably faster. On a quiz minutes later it averaged 50%, against 67% for people who coded by hand, and the widest gap was on debugging.⁶ The people who kept high scores used the assistant to ask conceptual questions, not to write the code for them.
A randomized study of 117 university students in the British Journal of Educational Technology supplies the best phrase in this whole literature. Yizhou Fan and colleagues found ChatGPT helped students raise essay scores but produced no extra knowledge gain or transfer. They warned of "metacognitive laziness": letting the tool do the planning and checking you would otherwise do yourself.⁷
Then there is the study everyone shares. MIT Media Lab's "Your Brain on ChatGPT" put EEG caps on 54 students writing essays with an LLM, a search engine, or nothing. Connectivity was weakest in the LLM group, and those writers struggled to quote essays they had finished minutes earlier.⁸ It is vivid. It is also a small preprint with a methods critique already in circulation, so I treat it as a picture of the problem, not the proof.
Across the stronger studies the pattern holds: AI lifts performance while you use it and lowers what you learn when it does the work you were supposed to practice.
The colonoscopy problem
If this were only about students, you could file it under homework.
In August 2025, Krzysztof Budzyń and colleagues reported in The Lancet Gastroenterology and Hepatology on endoscopists at four Polish centres that had started using AI to flag polyps. When those doctors worked without the AI, their adenoma detection rate fell from 28.4% before the tool arrived to 22.4% after.⁹ These were experienced specialists, not teenagers. The study is observational, so confounding is possible. The direction is still hard to ignore.
That is what effort atrophy looks like in a profession. Not a collapse. A slow handover of vigilance to a box that beeps, and a slightly worse doctor on the days the box is off.
The consulting version is subtler. In the Harvard and BCG experiment with 758 consultants, GPT-4 users finished 12.2% more tasks and worked 25.1% faster on work inside the model's abilities. On a task deliberately placed outside them, AI users were 19 percentage points less likely to reach the right answer, and their wrong answers read as more persuasive.¹⁰ Dell'Acqua's earlier name for this is falling asleep at the wheel.
I think of it as polish without traction. The output looks like effort happened. Sometimes it did not.
The case for the lazy brain
I have to argue with my own headline here, because the headline is too comfortable.
Offloading is how civilization works. Plato has Socrates complaining in the Phaedrus that writing would breed forgetfulness. Calculators were going to end arithmetic. Most of us cannot find our way across a city without a phone, and we have also mostly stopped getting lost. Some skills deserve to fade. Nobody mourns computing square roots by hand.
The evidence base is also thinner than the headlines. Michael Gerlich's widely shared 2025 study of 666 people in the UK found heavier AI use went with lower critical thinking scores, with offloading as the link.¹¹ It is correlational and leans on self report. People who already avoid hard thinking may simply reach for AI more. The Microsoft survey is self report too, and it found workers spending more effort on high stakes tasks, because now they had to verify, integrate and steward what the machine produced.⁴
So for some people, AI moves effort from producing to checking. That is real cognitive work, and arguably the work that matters most once output is cheap.
And here is the complication I cannot dodge. For some people the plain word is the right one. The Bastani students were not victims of a subtle design flaw. Handed a box that solves the problem, many used it to solve the problem. If that is not laziness, it is laziness's well dressed cousin. The tool did not create the impulse. It removed the friction that used to stop it.
Calling it "less practice" instead of "laziness" is not an excuse. It is a diagnosis, and it points at a fix.
Effort as a setting
If effort used to be a requirement and is now a setting, the question is who controls the setting.
Right now, mostly nobody. The default interface hands you the answer, because the answer is what you asked for. The students who kept their skills used a tutor designed to withhold it. The developers who kept their scores asked "why" instead of "write this." The people who hold on treat the tool as a sparring partner, not a ghostwriter.
That is a design choice and a workplace choice. We made a related argument about career entry at https://thinknarratives.com/articles/ai-will-not-take-your-job-it-is-taking-the-ladder: the first draft and the first ugly spreadsheet used to be where juniors learned. If the tool does all of those, the ladder thins, and so does the skill on the rungs that remain.
It shows up in places that look unrelated, too. When we looked at people asking chatbots to read their sleep data, at https://thinknarratives.com/articles/stop-asking-a-chatbot-how-you-slept, the confident summary replaced the slower habit of noticing your own patterns. Offloading the noticing is still offloading.
The gym analogy writes itself, which is fitting, because soon almost nothing else will have to.
What optional effort will look like by 2028
By 2028, effort will be a feature you switch on, not a cost you cannot avoid. Learning modes, hint only tutors and "explain before you answer" settings will be standard in the big assistants and in school deployments, because the evidence that unguarded tools hurt learning is already strong enough for institutions to act on. Professions that depend on vigilance, starting with medicine, will add AI off practice to training, the way pilots still practice flying by hand.
Most people will not pick the harder setting on their own. The ones who do will be noticeably better at the work the machine still gets wrong, and that gap will count for more in hiring than raw output.
AI will not make you lazy. It will make effort optional, and optional things are the first to go.
FAQ
Does using AI make you lazier? Not in the sense of damaging your brain. It lets you skip the effortful practice that builds and maintains skill. Studies of students and junior developers find better performance with the tool and weaker learning once it is removed.
Is ChatGPT making me lazy? It can make you practice less if you use it for answers rather than for checking your thinking. In one PNAS field study, students with an unrestricted ChatGPT style tutor scored 17% lower on an unassisted exam. A hint only tutor largely avoided that drop.
Does AI make you stupid? There is no good evidence that AI lowers intelligence. The stronger findings are narrower: less learning when AI does the task, and in one colonoscopy study, worse detection by doctors once the AI was off. Popular EEG results are early and small.
Does using AI reduce critical thinking? It can reduce how often you use it. A Microsoft Research survey of 319 knowledge workers found that higher trust in AI went with less critical thinking and higher self confidence with more. Workers also shifted effort toward verifying AI output.
Can AI make you less productive? Yes, on tasks outside what the model handles well. In a Harvard and BCG experiment, consultants using GPT-4 were faster and better inside its abilities, but 19 percentage points less likely to be correct on a task outside them.
How do I use AI without losing skills? Attempt the work first, then use the tool to critique it. Ask for hints, explanations and reasons rather than finished answers. In Anthropic's coding study, developers who asked conceptual questions kept far more of what they learned.
Footnotes
¹ Evan F. Risko and Sam J. Gilbert, "Cognitive Offloading," Trends in Cognitive Sciences 20(9), 2016, 676 to 688. Definition of cognitive offloading as physical action that reduces a task's cognitive demand.
² Wouter Kool, Joseph T. McGuire, Zev B. Rosen and Matthew M. Botvinick, "Decision Making and the Avoidance of Cognitive Demand," Journal of Experimental Psychology: General 139(4), 2010, 665 to 682. Six experiments; a bias toward the less demanding option that did not depend on awareness of the demand difference.
³ Daniel Kahneman, Thinking, Fast and Slow, Farrar, Straus and Giroux, 2011. Dual process model: fast System 1, effortful System 2.
⁴ Hao-Ping (Hank) Lee, Advait Sarkar and colleagues (Microsoft Research and Carnegie Mellon), "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers," CHI 2025, April 2025. 319 workers, 936 examples; confidence in AI linked to less critical thinking, self confidence to more; less effort on routine tasks, more on high stakes ones; shift toward verification and stewardship.
⁵ Hamsa Bastani, Osbert Bastani, Alp Sungu and colleagues, "Generative AI without guardrails can harm learning: Evidence from high school mathematics," PNAS 122(26), 25 June 2025. Nearly 1,000 students; practice grades up 48% (GPT Base) and 127% (GPT Tutor); GPT Base students 17% lower on the unassisted exam; safeguards largely mitigated the harm.
⁶ Judy Hanwen Shen and Alex Tamkin, "How AI Impacts Skill Formation," arXiv 2601.20245, and Anthropic, "How AI assistance impacts the formation of coding skills," 29 January 2026. Randomized trial, 52 developers; quiz 50% (AI) vs 67% (hand coding), p=0.01; speed gain not statistically reliable; largest gap on debugging.
⁷ Yizhou Fan and colleagues, "Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance," British Journal of Educational Technology, 2025. 117 students; ChatGPT improved essay scores but not knowledge gain or transfer.
⁸ Nataliya Kosmyna and colleagues, "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task," MIT Media Lab, arXiv 2506.08872, June 2025. 54 participants; weakest EEG connectivity and poorest recall in the LLM group. Preprint; statistics and sample size have been criticized.
⁹ Krzysztof Budzyń and colleagues, "Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study," The Lancet Gastroenterology and Hepatology, August 2025. Non AI adenoma detection fell from 28.4% (226 of 795) to 22.4% (145 of 648).
¹⁰ Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, Karim R. Lakhani and colleagues, "Navigating the Jagged Technological Frontier," Organization Science (working paper released 18 September 2023). 758 BCG consultants; 12.2% more tasks, 25.1% faster inside the frontier; 19 percentage points less likely correct outside it, with more persuasive write ups.
¹¹ Michael Gerlich, "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking," Societies 15(1), 6, January 2025. 666 UK participants; frequent AI use negatively correlated with critical thinking, mediated by offloading; self reported measures.
Sources
- Risko and Gilbert, Cognitive Offloading (2016): https://doi.org/10.1016/j.tics.2016.07.002
- Kool et al., Avoidance of Cognitive Demand (2010): https://pmc.ncbi.nlm.nih.gov/articles/PMC2970648/
- Kahneman, Thinking, Fast and Slow (2011): https://us.macmillan.com/books/9780374533557/thinkingfastandslow/
- Lee et al., CHI 2025: https://dl.acm.org/doi/10.1145/3706598.3713778
- Bastani et al., PNAS 2025: https://www.pnas.org/doi/10.1073/pnas.2422633122
- Anthropic, AI assistance and coding skills (2026): https://www.anthropic.com/research/AI-assistance-coding-skills
- Fan et al., BJET 2025: https://doi.org/10.1111/bjet.13544
- MIT Media Lab, Your Brain on ChatGPT (2025): https://www.media.mit.edu/publications/your-brain-on-chatgpt/
- Budzyń et al., Lancet Gastroenterology and Hepatology 2025: https://www.thelancet.com/journals/langas/article/PIIS2468-1253(25)00133-5/fulltext
- Dell'Acqua et al., Organization Science: https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- Gerlich, Societies 2025: https://www.mdpi.com/2075-4698/15/1/6
- Thinknarratives, AI Will Not Take Your Job: https://thinknarratives.com/articles/ai-will-not-take-your-job-it-is-taking-the-ladder
- Thinknarratives, Stop Asking a Chatbot How You Slept: https://thinknarratives.com/articles/stop-asking-a-chatbot-how-you-slept