MIT Professor Eric So Warns of Cognitive Atrophy from Overreliance on AI
MIT Sloan Professor Eric So examines the hidden cognitive costs of generative AI and outlines how professionals can avoid outsourcing critical thinking.
As artificial intelligence tools become integrated into daily workflows, researchers are examining how routine reliance on these systems alters human cognition. In his book, The Collision: What AI Does to Us, published this month by W.W. Norton, MIT Professor Eric So assesses the broader consequences of handing everyday intellectual tasks over to machine models. While acknowledging the speed and utility of domain-specific automation, So warns that excessive dependence risks causing a form of cognitive atrophy, diminishing our ability to build and preserve fundamental problem-solving skills.
So, the Sloan Distinguished Professor of Global Economics and Behavioral Science at the MIT Sloan School of Management, argues that human comprehension improves through iteration, overcoming mistakes, and actively grappling with material. Yet, behavioral patterns show a shift toward convenience. So highlights an Anthropic study finding that roughly half of queries submitted by college students were low-effort requests for simple answers. For So, the realization hit home after catching himself outsourcing his own phrasing to ChatGPT for a routine colleague inquiry, prompting him to analyze the systemic and psychological drivers behind our reliance on automation.
The Mechanics of 'AI Gravity'
According to So, human adoption of AI is accelerated by what he calls "AI gravity"—a convergence of natural biology and societal competition. Humans carry an innate biological drive to conserve mental energy, which makes instantaneous assistance appealing. When peers and competitors leverage AI to work faster, individuals face intense social and professional pressure to follow suit, particularly because detecting AI usage remains difficult.
This pressure is increasingly institutionalized. In many organizations, utilizing AI has shifted from an informal recommendation to an explicit metric for performance reviews, bonuses, and promotions. A similar cycle operates within academia, where students adopt tools prematurely under the belief that prospective employers will require them. So notes that this dynamic fosters an escalating dependency: by skipping the formative struggle required to build core competencies early on, users become increasingly incapable of operating without AI assistance later.
Rethinking Cognitive Friction
To preserve essential thinking skills, So advocates for viewing cognitive effort much like society came to view physical activity after industrialization. When mechanization reduced the need for physical labor, society had to intentionally reframe physical exercise as a necessary discipline. Mental effort, or "productive struggle," requires a similar cultural shift.
So also emphasizes the relational and psychological effects of interactive models. AI systems are inherently persuasive yet sycophantic—programmed to satisfy users to maintain engagement and earn positive ratings. Because these tools consistently agree with the user, individuals risk growing unaccustomed to real-world friction and interpersonal debate.
To counter cognitive erosion, So outlines four personal strategies: embracing productive struggle, valuing one's independent capability without AI, reinvesting saved time into building deeper expertise or personal growth, and deploying models as interactive tutors rather than passive answer machines.
What it means for developers
For software engineers and technical teams, So's findings present a direct challenge to how tools are incorporated into the development cycle. Developers frequently rely on AI to generate boilerplate, write documentation, or debug code. However, delegating the entire analytical process risks breaking what So terms the "inseparable task bundle"—the deep comprehension gained through writing, troubleshooting, and understanding an implementation from scratch.
When developers simply paste solutions from an assistant, they risk losing the foundational knowledge required to defend, modify, or debug complex architectures during critical production issues. Teams must intentionally balance velocity with cognitive investment. Rather than treating models strictly as code-completion engines, developers can use them as cognitive trainers—prompting them to critique designs, explain trade-offs, or simulate edge cases.
Teams experimenting with these prompting techniques can evaluate how different leading engines handle tutoring and technical reasoning. Developers can try top AI models cheaply through one API at https://apixoai.online to benchmark outputs from providers like Anthropic and OpenAI without committing to rigid single-vendor environments. Ultimately, retaining individual mastery ensures engineers remain capable of capitalizing on the tools they deploy, rather than becoming dependent on them.
Source: Q&A: The “collision” between AI, your brain, and society — MIT News AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
