Is AI Making Us Dumber? What the Research Actually Shows

What the research actually says about AI and cognition

The core finding, repeated across multiple studies, is that when AI handles cognitive work, the human brain does less of it. That reduced effort costs you something later. The mechanism is cognitive offloading, and it is not new. What is new is the category of tasks being handed over.

The MIT Media Lab study (2025) divided 54 participants into groups and had them complete essay tasks with brain activity monitored via EEG across 32 regions. The ChatGPT group showed the lowest neural engagement of all groups. More specifically, their brain activity declined over repeated sessions. They were not just producing less engaged work once. They were getting progressively less engaged the more they used the tool, reported Psychology Today in July 2025.

The FGV/UFRJ randomised controlled trial (2025, Social Sciences & Humanities Open) is methodologically stronger because it is preregistered. One hundred and twenty participants completed a learning task. Forty-five days later, without warning, they were tested on retention. ChatGPT group: 58%. Traditional study group: 69%. The researchers attributed the gap to weaker initial encoding. When AI does the synthesis, the brain processes information less deeply. Prior AI experience did not close the gap.

"Excessive reliance on AI-driven solutions contributes to cognitive atrophy and a measurable shrinking of critical thinking abilities." (MIT Media Lab study, Harvard Gazette, November 2025)

Seventy percent of 319 workers in a separate NYU professor study reported using less cognitive effort for reading comprehension when ChatGPT was available (TEDxSherbrooke, 2025). That figure is not a measure of laziness. It is rational adaptation: if a tool handles the task, most people let it. The cognitive cost arrives later.

A University of Technology Sydney (UTS) report from March 2026 reviewed the literature and named the pattern a "performance paradox." Unstructured AI use improves immediate task outputs. Students and workers produce better-looking work. Long-term learning suffers because the struggle through a task is precisely what builds durable knowledge. Remove the struggle and you weaken the encoding.

The Harvard Gazette's November 2025 coverage of the MIT findings drew a direct comparison to GPS navigation, where regular use measurably weakened spatial memory in ways users did not notice until they tried to navigate without it. Researchers working on AI existential risk, including those discussed in the will AI destroy humanity article, note that cognitive dependency is one of the underexamined near-term risks sitting alongside the longer-term concerns.

Is AI making people dumber?

For most people using AI without structure, yes, in specific measurable ways. Not general intelligence. Specific skills: memory retention, the ability to critically evaluate information, and independent reasoning without AI assistance.

The mechanism is cognitive offloading. Humans have always offloaded cognitive tasks to external tools. Notes, calendars, calculators. The brain does not maintain memory or perform calculations when external systems handle them reliably. This is efficient. What changes with AI is the category of tasks being offloaded. Calculators took arithmetic. AI takes reasoning, synthesis, analysis, and argument construction. These are exactly the tasks that build higher-order thinking when humans perform them.

A student who has AI write their essay skips the cognitive work that would have built their writing ability. The essay looks better. The writing skill does not improve. Over a degree program, that gap compounds.

The UTS March 2026 report distinguished two types of cognitive offloading with meaningfully different outcomes:

The most specific finding from the UTS research was something called "illusion of competence." Participants in unstructured AI use groups rated their own understanding as high because the AI output was fluent. They believed they had grasped the material. They could not reproduce the reasoning without AI assistance.

That gap between perceived and actual competence is the specific risk the research keeps returning to. It is harder to detect than a wrong answer, and harder to correct. According to the PMC cognitive offloading study (2025), AI reduces what cognitive load researchers call germane load, the mental effort of working through a problem deeply, and that reduction is the mechanism connecting AI use to long-term skill decline.

Is technology making us dumber? What history actually shows

The calculator debate is older than ChatGPT by 50 years. When electronic calculators entered classrooms in the 1970s, educators warned that students would lose arithmetic skills. They were right about arithmetic and largely wrong about broader reasoning. Mental arithmetic declined. Mathematical thinking did not. The tools shifted which skills mattered.

GPS navigation is the most direct analogy for the AI debate. Regular GPS users develop measurably weaker spatial memory than those who navigate without assistance. A 2020 study by Javadi et al. found that hippocampal activation during navigation was lower in GPS-dependent drivers than in active navigators. The Harvard Gazette made the comparison explicit in November 2025: heavy AI use risks "dulling minds the way GPS dulled our sense of direction."

Two things from the technology history are worth holding onto. First, cognitive offloading has always shifted which skills atrophied and which grew. Writing weakened oral memory but created entirely new modes of extended reasoning. Second, the transition costs were real even when long-term outcomes were neutral. People who grew up with calculators but never learned to estimate are worse at spotting computational errors than those who learned arithmetic first.

The AI debate follows the same pattern, with one difference: scope. Calculators took one task. GPS took one task. AI can take reasoning, judgment, synthesis, communication, planning, and creative production at the same time. The range of skills potentially affected is wider by a meaningful degree.

"AI risks turning knowledge work into shallow prompting, bypassing the deep work that actually builds expertise." (Cal Newport, The New Yorker, 2024)

The number most guides don't show

The FGV/UFRJ RCT showed an 11-point retention gap after one learning session, tested 45 days later. A student using AI for the majority of their coursework across a four-year degree accumulates retention deficits across dozens of subject areas. The researchers found the gap did not narrow with prior AI experience. The brain does not adapt and compensate over time. Familiarity with the tool does not solve the problem.

At 70% of 319 workers reporting reduced cognitive effort for reading comprehension when AI is available (NYU professor study), and the US knowledge workforce at approximately 65 million people, roughly 45 million workers may be operating in a reduced-engagement cognitive mode for reading tasks on any given workday. That number follows directly from the research figures. It is not in the literature itself.

Is AI itself getting dumber? The model quality question

Some searches for "is AI getting dumber" are not about human cognition at all. They ask whether AI models like ChatGPT are becoming less accurate or capable over time. That is a different question with a different answer.

There is evidence of capability drift in specific model versions, but no general decline across the field. Researchers at Stanford and UC Berkeley published a 2023 study tracking GPT-4 and GPT-3.5-turbo performance on specific tasks over several months. GPT-4's accuracy on identifying prime numbers dropped from 84% in March 2023 to 51% in June 2023. OpenAI attributed the changes to ongoing safety fine-tuning that altered model behaviour on certain task types.

The Epoch AI benchmark tracking project (2023-2025) has recorded 2-5% drops on MMLU scores for specific model versions following update cycles. These are not model collapses. They are fine-tuning side effects on particular task categories.

The broader picture is different. GPT-5.2, Claude Sonnet 4.6, and Gemini 2.5 Pro all outperform their predecessors on most standard benchmarks. The overall capability frontier is rising. Individual model versions can drift after specific updates. The field as a whole is not declining.

For users who suspect a specific AI tool has gotten worse, the most reliable test is to benchmark it on a task they completed themselves six months ago and compare outputs directly. Subjective impressions of AI degradation are common and are often correct about a specific model version at a specific point in time. For a full breakdown of where current frontier models sit relative to the AGI threshold, see what is AGI explained.

How to stay cognitively sharp while using AI

AI use type matters more than AI use frequency. The UTS March 2026 report identified five habits that preserve cognitive function in regular AI users, and the FGV/UFRJ research team added a sixth based on their retention findings.

Draft first, then use AI. Producing your own attempt before consulting AI forces the initial encoding that builds durable knowledge. Even a rough draft engages the reasoning circuits that passive AI consumption bypasses. The FGV/UFRJ researchers found that participants who formed an initial memory trace before using AI retained more at 45 days.

Use AI for extraneous load, not germane load. Extraneous cognitive load is the mental overhead of formatting, grammar, and surface presentation. Germane load is the effort of working through the actual argument or problem. Offload the first. Keep the second.

Verify outputs verbally. After using AI, close the window and try to explain the output in your own words without looking. If you cannot, you have not processed the information deeply enough to retain it. This test reveals the illusion of competence before it creates a problem.

Apply retrieval practice. Review AI-generated material with time gaps. Test yourself on the content 24 hours later without the AI. This works for the same reason flashcards work: spaced retrieval strengthens memory traces that passive reading does not.

Calibrate to task type. Use AI where the output is the point (a boilerplate email, a format conversion, a grammar check). Avoid AI where the process is the point (understanding a concept, developing an argument, making a judgment call). The distinction is easy to lose in a workflow built for speed.

For readers thinking about which jobs are most exposed to AI task displacement versus which skills remain durable, the jobs AI cannot replace article covers automation risk scores for 12 career categories. The roles with the lowest automation risk share a property relevant here: they require the exact kind of embodied, relational judgment that suffers most from cognitive offloading.

Frequently Asked Questions

Related Articles