“AI Dependence Is Shutting Down Human Thought”: The ‘AI Paradox’ Grips Campuses and Workplaces

Repeated acceptance of chatbot answers, wholesale outsourcing of thought
Faster assignment completion at the expense of memory and verification practice
Cognitive debt spills into the labor market, widening hiring and wage disparities

A report has found that university students are becoming excessively reliant on artificial intelligence (AI) and experiencing what researchers call “cognitive surrender,” abandoning the process of thinking for themselves. As students delegate more assignments to AI, the cognitive activity devoted to memory, reasoning, and verification appears to weaken. If such learning habits become entrenched, their ability to define problems independently and identify errors may also deteriorate. Concerns are mounting that disparities in reasoning skills formed at university could translate into differences in workplace competence after graduation, ultimately determining hiring prospects and wages.

AI Dependence Fuels an “Illusion of Learning”

According to the New York Times (NYT) on September 15, local time, the Massachusetts Institute of Technology’s (MIT) research committee on AI use warned in a report last month that students may fall into an “illusion of learning” by immediately turning to AI for answers whenever they encounter even minor difficulty. The report also noted a growing tendency for students to isolate themselves by consulting chatbots instead of attending professors’ office hours or studying with their peers. Researchers also identified cognitive surrender, in which students transfer to AI the mental burden required for the reasoning process. In an experiment involving 54 participants conducted by the MIT Media Lab, the ChatGPT group exhibited the weakest neural connectivity among those who used ChatGPT, those who used search engines, and those who used no tools. ChatGPT users also demonstrated weaker recall of the sentences they had written and a diminished sense of ownership over their work.

Research findings also suggest that AI may impair actual learning capacity. According to a recent study of 26,000 middle and high school students in China, homework scores increased by 18% after students began using AI, but their scores on tests conducted in controlled environments where AI was unavailable declined by 20%. The results indicate that although AI helped students complete their assignments, it may have weakened the deep learning required to solve problems independently. Natalya Kosmyna, an MIT researcher who participated in the study, explained that the students had “effectively not used the memory networks that help them remember at all.”

Passive Acceptance of Information Proliferates

Academic achievement among U.S. students also appears to be weakening substantially. In an editorial published on September 14, the Wall Street Journal (WSJ) stated, “Last year’s Programme for International Student Assessment (PISA) found that American students lagged far behind their Chinese counterparts,” adding that “the PISA results show that declining educational competitiveness represents a more fundamental problem for the United States in its innovation rivalry with China.” Researchers at the U.S. think tank Center for Economic and Policy Research (CEPR) identified “effort displacement” as the principal cause of learning loss. Their analysis found that approximately 80% of the learning loss stemmed from students completing assignments at abnormally high speed.

It has long been established that people who are reluctant to expend cognitive effort are more inclined to accept information exactly as it is presented. Generative AI has merely raised the threshold at which users feel verification can be omitted. Because answers are presented in fluent prose and well-ordered logic, the cues needed to detect errors and uncertainty have become less visible, prompting users to delegate even judgments in unfamiliar fields to chatbots. Conditions have thus emerged for the passive acceptance of information, once observed primarily among specific groups, to spread into an everyday mode of thought.

Table 1. Impact of AI Dependence on Learning Ability

Eight in 10 Accept AI Answers Even When They Are Wrong

Stephen Shaw and Gideon Nave of the University of Pennsylvania’s Wharton School characterized this phenomenon as “System 3,” in which an external AI intervenes in human intuition and deliberation. In a preprint, the researchers administered 9,593 Cognitive Reflection Tests to 1,372 participants and found that 79.8% accepted the answers provided by AI even when they were incorrect. The correct-response rate, which stood at 45.8% when participants solved problems without AI, increased to 71% when AI supplied the correct answer but plunged to 31.5% when it provided an incorrect one. This marked the emergence of the “Scissors Effect,” in which human reasoning performance becomes contingent on the model’s accuracy.

Participants remained confident even after accepting incorrect answers. In an experiment imposing a 30-second time limit, the accuracy rate among AI users on questions for which AI supplied an incorrect answer fell to 12.1%. Even when participants received financial rewards for accurate judgments and immediate feedback, the rate of cognitive surrender reached 57.9%. Participants with greater trust in AI were more likely to accept incorrect answers, while a strong need for cognition and high fluid intelligence acted as protective factors. Even after mechanisms designed to encourage verification were introduced, the first answer supplied by AI remained the anchor for subsequent judgment.

“Cognitive Debt” Obscured by the Speed of Generation

As such dependence accumulates, cognitive disparities in the AI era are expected to become increasingly polarized. Researchers at Microsoft Research and Carnegie Mellon University analyzed 936 cases of AI use involving 319 knowledge workers and identified an association between greater trust in AI and less frequent engagement in critical thinking. Those who were confident in their own professional competence, by contrast, engaged more actively in verifying information, integrating results, and managing tasks. Even when using the same tool, one group directs AI’s judgment while the other becomes confined to approving its generated answers.

Moreover, the repercussions of AI dependence are accumulating as “cognitive debt.” Margaret-Anne Storey, a professor at the University of Victoria in Canada, argued in a recent paper that as AI rapidly produces large volumes of code, development teams may lose their understanding of that code, while design objectives and technical constraints may also disappear from the documentation. Outputs continue to accumulate even as those responsible become unable to explain how they work or how they should be modified. As these gaps widen, error response and maintenance are again delegated to AI, inevitably eroding the organization’s decision-making capacity.

Workers Without Reasoning Skills Risk Being Squeezed Out of the Labor Market

These cognitive gaps are translating directly into displacement pressure in the labor market. Liu Shengyu, a machine-learning systems engineer at Chinese AI company DeepSeek, recently went so far as to raise the possibility of changing careers. A graduate of Peking University’s Turing Class, Liu ranked among the top performers in a global university supercomputing competition and participated in developing the core computational technology behind DeepSeek V4.1, placing him among the field’s most elite engineers. He predicted that AI could perform the complex systems-optimization work currently under his responsibility within the next six months to one year. This suggests that even engineers working at the frontier of AI development are beginning to perceive the risk of displacement.

Changes across industry are already visible in hiring and wage indicators. Google Chief Executive Sundar Pichai said in April that 75% of the company’s new code was being generated through AI. The proportion has tripled from 25% in October 2024 in less than two years. The market value of workers equipped with AI skills has also surged. According to the “2026 Global AI Jobs Barometer” published by global accounting and consulting group PricewaterhouseCoopers (PwC), which analyzed more than 1 billion job postings across 27 countries, jobs requiring AI skills carried an average wage premium of 62%. This represented a five-percentage-point increase from 57% last year. Companies have begun to recognize AI proficiency as a distinct qualification and a basis for higher compensation.

The threshold for entry-level recruitment has also risen sharply. PwC’s analysis of 2.4 million entry-level jobs in the United States found that positions with high exposure to AI were seven times as likely as conventional entry-level roles to require judgment, leadership, creativity, and face-to-face communication skills. These so-called “high-skill entry-level jobs” have increased by 35% since 2019, while conventional entry-level positions have declined by 10%. Job seekers possessing only degrees and certifications are being screened out at the application stage, while hiring opportunities are increasingly concentrated among candidates who can demonstrate both subject-matter expertise and the ability to use AI.