Drafting emails, summarizing documents, solving equations, looking up facts, generating options, checking arguments: tasks that once required sustained mental effort now resolve in seconds with a single prompt. According to new research from Carnegie Mellon, MIT, Oxford, and UCLA, the tools may be reshaping cognitive habits humans spent decades building.
Scientists have started asking what the brain does with itself when it no longer has to work quite so hard. The answer, emerging from multiple independent research programs in 2025 and 2026, is more concerning than most public conversation acknowledges.
The Landmark Four-University Study
A preprint posted in late April 2026, covered extensively by outlets including Fast Company, pooled three randomized controlled trials conducted by researchers at Carnegie Mellon University, the University of Oxford, MIT, and UCLA. The paper carries a title that leaves little room for ambiguity: AI Assistance Reduces Persistence and Hurts Independent Performance.
The study enrolled 1,222 adults in tasks built around two established cognitive science instruments: a graded set of fraction-based math problems and a reading-comprehension battery. Half worked without assistance, while the other half had access to an AI assistant powered by OpenAI’s GPT-5.
The design included a deliberate shock. After AI-assisted participants worked through the majority of the problems, the AI helper was removed without warning for the test’s final three problems.
What Happened When the AI Disappeared
While the AI was available, the assisted group outperformed the controls. The moment the assistant disappeared, the assisted group’s solve rate dropped to 71 percent, against 77 percent in the control group. A group that was performing better suddenly performed worse, on problems of equivalent difficulty, with no change in the available time.
According to co-author Rachit Dubey, “Once the AI is taken away from people, it’s not that people are just giving wrong answers. They’re also not willing to try without AI.” Researchers found that just 10 to 15 minutes of exposure to AI tools reduced cognitive function and problem-solving abilities.
Participants who asked the chatbot for direct answers suffered the steepest drop in solve rates and skipped nearly twice as many questions. Those who asked for hints or clarification stayed close to the no-AI control group. Delegating the answer is categorically different from delegating part of the process.
The Cumulative Erosion
The researchers argue that how you prompt the model matters more than whether you use it at all. Each individual interaction feels efficient and reasonable. The cumulative erosion of independent problem-solving capacity, across hundreds of such interactions, does not announce itself until the AI is gone and the user finds themselves staring at a problem they cannot begin.
Current AI systems are built to deliver immediate, complete responses. The goal is to satisfy the user’s request as effectively as possible in the moment — not to build the user’s independent capability over time. Those are two very different design objectives.
The MIT “Cognitive Debt” Research
Researchers led by MIT Media Lab’s Nataliya Kosmyna conducted the first comprehensive study examining what happens to human brains when people use AI writing assistants. The study used advanced brain monitoring technology to track 54 participants as they wrote essays using either AI tools, search engines, or no external help. The findings revealed concerning changes in brain connectivity patterns and cognitive performance that persisted even after AI use stopped.
The work, titled “Your Brain on ChatGPT,” tracked writers across repeated essay sessions using EEG headsets. Researchers observed reduced neural connectivity, poorer memory of written text, and more homogenized prose among heavy AI users. Lead author Kosmyna labeled the phenomenon “cognitive debt.”
The EEG data produced a clear gradient. Brain-only participants showed the strongest, most distributed neural networks. Search-engine users showed moderate engagement. LLM users showed the weakest connectivity. Self-reported essay ownership was lowest in the LLM group and highest in the brain-only group. LLM users also struggled to accurately quote their own work.
Participants who used AI to write essays could not, immediately afterward, accurately recall what was in those essays. Authorship and memory are normally tightly coupled. When the AI does the generating, that coupling appears to break down.
Limitations Worth Noting
The MIT study’s sample size of 54 participants is relatively modest for the strength of the claims it generated in popular coverage. Reviewers noted that some results could be interpreted more conservatively, with primary concerns focusing on study design considerations including the limited sample size, the reproducibility of the analyses, methodological issues related to the EEG analysis, and inconsistencies in the reporting of results. The preprint had not completed formal peer review as of mid-2026. Replication in larger samples is needed before the specific effect sizes are treated as settled numbers.
The four-university behavioral study, by contrast, enrolled 1,222 participants across three separate randomized controlled trials, a sample size that is considerably more robust for the type of behavioral outcomes it measured.
Cognitive Offloading: The Underlying Process
The concept binding these studies together is cognitive offloading, the process by which humans externalize mental tasks to tools, devices, or environments to reduce the effort required in the moment. Cognitive offloading is not new. Writing itself is a form of cognitive offloading. So are calculators, navigation apps, and calendars. The relevant question is not whether offloading occurs, but what repeated offloading does to the underlying capacity.
When a person solves a math problem manually, they are not just getting the answer. They are reinforcing the neural pathways associated with mathematical reasoning, building tolerance for the discomfort of not immediately knowing, and developing persistence. When a model provides the answer instantly, none of that reinforcement occurs. The answer is obtained, the session ends, and the capacity to generate that answer independently remains exactly where it was before, or, as the four-university study suggests, slightly weaker.
By automating tasks and providing information fast, AI tools leave the user with less cognitive load to devote to more complex parts of the task. However, such a decrease can also result in cognitive offloading, when individuals have become overly dependent on AI and lost the ability to think critically and engage in depth. The initial reduction in cognitive load is experienced as a benefit. The downstream erosion of capacity is experienced much later, if at all.
Age, Education, and Differential Vulnerability
A 2025 study published in Societies, drawing on surveys and in-depth interviews with 666 participants across diverse age groups, found a significant negative correlation between frequent AI tool usage and critical thinking abilities, mediated by increased cognitive offloading. The same study found that higher educational attainment was associated with better critical thinking skills regardless of AI usage, and younger participants exhibited higher dependence on AI tools alongside lower critical thinking scores compared to older participants.
Younger users, such as those aged 17 to 25, appear more vulnerable to overreliance, perhaps because they have less developed reasoning habits or are less aware of their cognitive strategy. The age group most likely to be using AI tools extensively for academic work is also the group whose independent reasoning habits are still forming. The tools are arriving precisely when the brain is most in the business of consolidating how it thinks.
If you’re already thinking about how technology shapes developing minds, the question of screen time and children’s thinking is worth considering alongside this research.
The Decisive Variable: How, Not Whether
The most practically important finding across all of this research is also the one that received the least attention in popular coverage: the effect is not uniform across all AI users. It is concentrated in a specific type of user.
Researchers found that participants “who asked the AI for direct solutions saw the largest decline in solve rate and the largest increase in skip rate.” Those who used the AI for partial guidance, hints, or explanations rather than complete answers did not show the same deterioration. In the reading-comprehension and math trials, the gap between answer-seeking users and hint-seeking users in post-AI performance was substantial enough to represent a genuinely different behavioral profile.
The distinction maps onto a broader principle in cognitive science: the generation effect. When a person generates an answer, even a partially correct one, through their own effort, the memory trace for that information is stronger and the associated reasoning process is reinforced. When a person reads or receives a correct answer without attempting to generate it themselves, the retention and skill-building are significantly weaker. AI that provides complete answers bypasses the generation process entirely. AI that provides partial guidance forces the user to complete the generation step, preserving much of the cognitive work that makes the interaction educationally valuable.
The question is not whether to use AI, but whether the way it is being used is preserving or replacing the cognitive effort that builds durable capability. The phrase AI brain independence is becoming a useful shorthand for exactly this distinction.
Read More: Steve Jobs Banned iPads for His Kids – The Surprising Reason Might Change Your Mind
What the Research Actually Tells Us
For users who delegate the full cognitive task to AI, the evidence of measurable decline in independent performance is now multi-study and consistent. For users who engage AI as a scaffold, using it to check direction rather than supply answers, the same decline does not appear to materialize.
The four-university study, the MIT EEG research, and the critical thinking data collectively demonstrate that the effect is real, it is rapid, and it is not inevitable. It follows from a specific pattern of use. The researchers are not arguing that AI is dangerous. They are arguing that AI as currently designed, optimized for immediate and complete answers rather than for the long-term development of user capability, creates a structural incentive toward the type of use that produces cognitive cost. The cost accumulates across interactions that each feel, individually, entirely reasonable.
The researchers who ran the four-university study included a line that did not make it into most coverage: “each incremental act feels costless, until the cumulative effect becomes overwhelming to address.” Ten minutes is already enough to produce a measurable effect, and daily use across months compounds that effect in ways no single experimental session can capture.
For parents thinking about how their children engage with these tools, for employers watching their teams use AI across entire working days, and for individuals trying to protect the reasoning capacity they’ve spent years building, the question isn’t whether to engage with AI at all. It is whether the engagement is one that leaves the human brain more capable, or one that quietly borrows against it.
Disclaimer: This information is not intended to be a substitute for professional medical advice, diagnosis, or treatment and is for information only. Always seek the advice of your physician or another qualified health provider with any questions about your medical condition and/or current medication. Do not disregard professional medical advice or delay seeking advice or treatment because of something you have read here.
AI Disclaimer: This article was created with the assistance of AI tools and reviewed by a human editor.