Research so far suggests that AI assistants can reduce effort and skill practice when people lean on them without checking or thinking first, but the evidence does not show that AI makes people lazy by nature. Studies from MIT, Microsoft Research, and the University of Pennsylvania link heavy, unguarded use to weaker recall, less critical thinking, and lower scores once the tool is removed. The same studies show that how the tool is designed and used changes the result.
Key takeaways
- An MIT-led preprint of 54 participants found weaker brain connectivity and poorer recall of their own essays among people who wrote with a chatbot.
- A CHI 2025 survey of 319 knowledge workers found that higher trust in AI went with less critical thinking.
- In a field experiment with nearly 1,000 high school students, unrestricted GPT-4 access improved practice results but left students worse off on a later exam.
- A tutor version built to give hints, not answers, largely avoided that drop.
- Automation bias is an older, well-documented risk: people accept machine output too readily when it is usually right.
What "lazy" and "vulnerable" mean here
Laziness is a judgment about motive, and studies cannot measure motive well. What they can measure is effort, memory, and performance without the tool. Vulnerability is easier to define: a person is vulnerable when they cannot tell that an AI answer is wrong, or cannot do the task when the tool is missing. The question is therefore narrower than the headline version. It asks whether using AI changes how much thinking people do, and whether that matters when the tool fails.
Three studies, side by side
| Study | Who took part | Main finding | Main caution |
|---|---|---|---|
| Kosmyna et al., MIT Media Lab and collaborators, preprint submitted June 10, 2025 | 54 people writing essays, split into chatbot, search engine, and no-tool groups; 18 returned for a fourth session | Chatbot users showed the weakest connectivity in EEG recordings, the lowest sense of ownership of their essays, and difficulty quoting their own work | Small sample, one task type, and a preprint on arXiv |
| Lee et al., Microsoft Research and collaborators, CHI 2025 | 319 knowledge workers sharing 936 examples | More trust in generative AI went with less critical thinking; more self-confidence went with more | Survey-based, so it shows reported experience, not measured behavior |
| Bastani et al., University of Pennsylvania and collaborators, PNAS, July 2025 | Nearly 1,000 high school math students in Turkey | Unrestricted GPT-4 raised practice performance but left students worse off on an exam without it | One subject, one school system, a limited period |
The MIT essay study in detail
In the MIT paper, participants wrote essays over three sessions with a chatbot, a search engine, or no tool, while EEG recorded brain activity. The no-tool group showed the strongest and most widely spread connectivity, the search group fell in the middle, and the chatbot group showed the weakest. In a fourth session, 18 participants switched conditions. Those who moved from the chatbot to no tool showed reduced alpha and beta connectivity, which the authors read as under-engagement. Those who moved the other way, from no tool to the chatbot, showed higher recall and activation similar to search users. The pattern suggests that order matters: using AI after doing the thinking looked different from using it instead of thinking. The authors themselves call for deeper inquiry, and 54 people cannot settle a question about a whole population.
Trust and verification at work
The Microsoft Research survey asked workers about real tasks. Rather than a simple decline, the authors describe a shift in the character of critical thinking toward information verification, response integration, and task stewardship. A worker who checks a draft instead of composing it is still thinking, just about different things. The risk appears when trust is high and checking stops.
Automation bias is not new
Psychologists studied this long before chatbots. In a 2010 review in the journal Human Factors, Raja Parasuraman and Dietrich Manzey describe complacency and automation bias, in which people under-monitor automated systems or follow imperfect advice, producing omission and commission errors. They report that the effect appears in both novices and experts and was not removed by training or instruction. Generative AI brings the same pattern to everyday writing and study.
Habits that keep skills in use
- Attempt the task first, then use AI to critique or extend the attempt.
- Ask for hints or questions rather than finished answers when learning a skill.
- Verify names, numbers, and quotations against a primary source before reuse.
- Schedule some work with no assistant, the way pilots keep up manual flying.
- In teaching, test students without the tool as well as with it.
FAQ
Does using ChatGPT damage the brain?
No study shows lasting damage. The MIT preprint reports differences in brain activity and recall during essay writing in a small group, and its authors call for further work.
Are students who use AI learning less?
It depends on the use. In the Penn-led experiment, students with unrestricted access did worse on a later exam, while those with a safeguarded tutor largely avoided that harm.
What is automation bias?
It is the tendency to accept the output of an automated aid without enough checking, which leads to errors when the aid is wrong or silent.
Can AI improve work as well?
Practice performance in the high school study rose with AI access, so results with the tool can improve. The concern is what remains when the tool is removed.
Related reading
- A beginner path for learning AI
- AI and human intelligence compared
- Why human judgment stays in the loop
- Humans and AI on creative tasks
More in AI Basics.
Last verified: October 2026
Sources
- Kosmyna N. et al., Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, arXiv:2506.08872, 2025
- Lee H.-P. et al., The Impact of Generative AI on Critical Thinking, CHI 2025, Microsoft Research, 2025
- Bastani H. et al., Generative AI without guardrails can harm learning: Evidence from high school mathematics, Proceedings of the National Academy of Sciences, vol. 122(26), 2025
- Without Guardrails, Generative AI Can Harm Education, Knowledge at Wharton, August 27, 2024
- Parasuraman R., Manzey D., Complacency and Bias in Human Use of Automation: An Attentional Integration, Human Factors, 2010