The fear that technology is making humans less capable is older than the internet. Writing was once accused of weakening memory. Calculators were blamed for undermining arithmetic. Search engines were said to hollow out knowledge. GPS raised concerns about spatial navigation. Each time, the panic was partly exaggerated and partly correct.
The same pattern is now unfolding around generative AI. The claim spreading through schools, workplaces, and social media is blunt: AI is making us dumber. But that framing is too crude. Intelligence is not a single muscle that either grows or collapses every time a new tool appears. It is a set of cognitive routines: remembering, searching, comparing, planning, writing, revising, resisting easy answers, and persisting through ambiguity.
The more accurate concern is not that AI is destroying intelligence. It is that AI is changing which cognitive skills people continue to exercise — and which ones they quietly stop practicing. That distinction matters. It shifts the debate away from moral panic and toward a more useful question: what happens when a tool no longer just helps retrieve information, but begins to generate the argument, the structure, the language, and sometimes the conclusion itself?
The old internet already changed memory
The scientific foundation for this debate predates ChatGPT, with a study published in Science titled “Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips”. The researchers found that when people expected information to remain accessible later, they were less likely to remember the information itself and more likely to remember where it could be found.
This is often summarized as the “Google effect,” but the mechanism is more interesting than the slogan. The study did not prove that search engines make people stupid. It suggested that people adapt their memory strategies to the environment. When information is easy to retrieve externally, the brain reallocates effort. It remembers pathways, locations, and access points instead of storing every fact internally.
That is not necessarily cognitive decline. It is cognitive offloading. Humans have always offloaded cognition. Calendars offload memory. Maps offload spatial planning. Notebooks offload working memory. Search engines offload factual recall. In many cases, this is beneficial. External tools extend cognition by freeing mental resources for higher-order tasks. The problem appears when offloading becomes substitution: when the tool does not merely support a skill, but replaces the need to practice it.
Generative AI intensifies this issue because it does not only store or retrieve. It composes. It summarizes. It argues. It chooses tone. It generates code. It proposes plans. It simulates expertise. In other words, it can occupy the parts of cognition that previous tools mostly left untouched.
AI does not just retrieve answers. It performs the task
A calculator does not explain what a student should think about a novel. A GPS does not write the final paragraph of a legal memo. A search engine does not automatically transform a vague idea into a polished strategy document. Generative AI can do all of those things. That difference explains why the cognitive stakes feel higher.
With earlier tools, the user still had to assemble the output. With generative AI, the output often arrives already assembled. The temptation is not just to save time, but to skip the difficult middle stage where learning usually happens. That middle stage is cognitively expensive. It includes confusion, retrieval, false starts, revision, and the slow conversion of vague intuitions into explicit reasoning. These are not inefficiencies. They are part of how skills are built.
This is where recent research becomes relevant. The evidence does not justify the claim that AI is broadly reducing intelligence. But it does suggest that certain forms of AI use can reduce cognitive engagement, weaken independent performance, and change the user’s relationship to their own work.
The writing problem: less effort, less ownership
One of the most discussed studies in this area is “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task”, a 2025 preprint by Nataliya Kosmyna and colleagues. The researchers divided participants into groups that wrote essays using an LLM, a search engine, or no external tool. They also used EEG to measure brain activity during the task.
The results were striking but should be interpreted carefully. Participants who wrote with no tool showed the strongest and most distributed brain connectivity. Search-engine users showed moderate engagement. LLM users showed the weakest connectivity. The LLM group also reported the lowest sense of ownership over their essays and had more difficulty accurately quoting from their own work later.
This does not mean ChatGPT “damages the brain.” The study was relatively small, published as a preprint, and focused on a specific task: essay writing. A later comment on the paper raised methodological concerns, including sample size, reproducibility, EEG analysis, and interpretation of the results.
Still, the core signal is hard to ignore: when a model does much of the writing, the user may become less cognitively entangled with the final product. That matters because writing is not only a communication act. It is a thinking process. Drafting forces prioritization. Revision exposes weak logic. Sentence construction clarifies relationships between ideas. When AI performs those steps too early or too completely, the user may receive a finished text without having gone through the cognitive friction that normally produces understanding. The risk, then, is not that AI writes badly. The risk is that AI can write well enough to let people bypass the work that would have made them better thinkers.
The learning problem: short-term performance can hide long-term weakness
The same pattern appears in education. A 2025 study in PNAS, “Generative AI Without Guardrails Can Harm Learning”, examined the use of GPT-4-based tutors in math education. The researchers conducted a field experiment with nearly a thousand high school students and compared different forms of AI assistance.
The key result was not that AI is bad for learning. It was that design matters. Students using AI support could improve performance during practice. But when the AI was removed, those who had relied on a more direct, ChatGPT-like interface underperformed compared with students who did not receive the same kind of help. By contrast, a tutor designed with learning safeguards — one that pushed students toward reasoning instead of simply supplying answers — reduced that harm.
This distinction is essential. AI can be a tutor or a shortcut. The interface determines which role it plays and a good tutor does not merely answer. It withholds, prompts, redirects, tests, and scaffolds. It creates productive struggle. A bad tutor optimizes for immediate completion. It removes friction so efficiently that the student may finish the task without developing the underlying skill. This is the core paradox of generative AI in education and professional work: the tool can improve output while weakening the process that would have improved the user.
The persistence problem: easy answers change behavior
Another important piece of evidence comes from the 2026 preprint “AI Assistance Reduces Persistence and Hurts Independent Performance” by Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel Bakker, and Rachit Dubey. Across randomized controlled experiments involving 1,222 participants, the authors found that AI assistance improved short-term performance but reduced later performance when participants had to work without AI. It also made participants more likely to give up.
The most important part is the timescale. These effects appeared after brief interactions with AI, around ten minutes in some conditions. That does not prove permanent cognitive change, but it does suggest that AI can rapidly alter task behavior. If a tool trains users to expect immediate answers, then difficult problems may start to feel less like normal cognitive work and more like interface failure.
Persistence is not a decorative virtue. It is part of skill acquisition. Many complex abilities depend on staying with a problem past the first moment of confusion. Coding, technical writing, design, and strategic planning all require tolerance for incomplete understanding. If AI becomes the default response to friction, users may not lose intelligence. They may lose stamina. That is a subtler but more practical concern. A person can remain intelligent while becoming less willing to engage in the slow, uncomfortable behaviors through which intelligence becomes competence.
GPS shows the pattern: unused skills decay
The concern is not unique to AI. Navigation research shows a similar mechanism. In “Habitual Use of GPS Negatively Impacts Spatial Memory During Self-Guided Navigation”, Louisa Dahmani and Véronique Bohbot found that greater lifetime GPS use was associated with worse spatial memory during self-guided navigation. A follow-up over three years suggested that heavier GPS use predicted steeper decline in hippocampus-dependent spatial memory.
A later systematic review on GPS use and navigation ability found negative effects on environmental knowledge and sense of direction, while also noting methodological weaknesses in parts of the literature.
The lesson is not that GPS should be abandoned. GPS saves time, reduces stress, and improves safety in many contexts. But it also changes what users practice. A driver who never builds a mental map may become less capable of navigating without assistance. The skill does not vanish because the person becomes less intelligent. It decays because it is no longer exercised.
AI may do something similar to writing, reasoning, memory, and problem-solving. Not automatically. Not universally. But plausibly, especially when the tool is used as a substitute rather than a scaffold.
The wrong question is whether AI makes us stupid
The public debate tends to collapse everything into a binary: AI is either an amplifier of intelligence or a machine for cognitive decline. The evidence supports neither extreme.
Generative AI can help experts move faster, explore alternatives, detect blind spots, translate between domains, and compress tedious work. For people with strong prior knowledge, AI can function as leverage. It can make expertise more productive.
But for learners, novices, and distracted professionals, the same tool can become a bypass. It can produce the appearance of competence before competence exists. It can generate fluent output without requiring the user to build the internal model behind that output. This is why AI’s cognitive impact depends heavily on how it is used.
Using AI to challenge a draft is different from using it to write the draft. Asking AI to generate counterarguments is different from asking it to decide what one believes. The same model can either deepen cognition or replace it. The difference is not the tool alone. It is the workflow.
The future skill is not avoiding AI. It is knowing when not to delegate
The practical conclusion is not to reject AI. That would be both unrealistic and intellectually lazy. Generative AI is becoming infrastructure: embedded in search, productivity software, code editors, classrooms, customer support, operating systems, and research workflows. The real question is governance at the level of cognition.
Some tasks should be delegated because they are low-value, repetitive, or mechanical. Others should not be delegated too early because they are the very tasks through which skill is formed. Summarizing a document after reading it is different from summarizing it instead of reading it. Getting feedback on a paragraph is different from outsourcing the paragraph. Asking for hints is different from asking for answers.
The most important educational and professional habit may be deliberate friction: preserving moments where the human still has to retrieve, reason, draft, compare, and struggle before the machine intervenes.
This distinction is reinforced by the PNAS study on AI tutoring, which found that systems designed with learning safeguards led to better outcomes than those that simply provided direct answers. Students who relied on more answer-oriented AI performed worse once the tool was removed, while those using structured guidance retained more of what they learned.
AI is a mirror for cognitive discipline
AI is not making us dumber in any simple, universal, scientifically settled way. The evidence is still early, uneven, and task-specific. Some of the most provocative findings come from preprints that require replication and stronger longitudinal data. There is not yet enough evidence to claim that generative AI causes broad cognitive decline.
But the early science does point to a real risk: cognitive outsourcing can produce short-term fluency while weakening the habits that create durable skill.
The danger is not that AI will think for everyone. The danger is that it will become convenient enough that many people stop noticing which parts of thinking they have stopped doing.
A more precise way to see it is this: AI doesn’t have to erode intelligence to reshape how we think. It simply has to make some kinds of effort feel optional. When that happens, the skills tied to that effort begin to fade quietly, often disguised as efficiency — until the moment the tool is gone, the context shifts, or the person has to rely on their own reasoning again.
Discover more from TekCrispy
Subscribe to get the latest posts sent to your email.



TekCrispy