GPS killed our ability to navigate. This is not a figure of speech. A longitudinal study published in Scientific Reports found that habitual GPS users showed measurable decline in hippocampal-dependent spatial memory over three years, and critically, the decline was caused by GPS use rather than the reverse (Dahmani & Bohbot, Scientific Reports, 2020). In 2011, psychologist Betsy Sparrow demonstrated the “Google Effect”: when people expect future access to information, they have lower recall of the information itself and enhanced recall for where to find it. The internet had become transactive memory, an external system we know how to query rather than knowledge we actually hold (Sparrow, Liu & Wegner, Science, 2011). GPS eroded spatial reasoning. Search engines eroded factual retention. AI extends this pattern to the cognitive functions that matter most for self-governance: analysis, evaluation, and the capacity to detect when you are being manipulated.
The early evidence is concerning. A 2025 study of 666 participants found a significant negative correlation (r = -0.68) between frequent AI tool usage and critical thinking abilities, mediated by cognitive offloading (Gerlich, Societies, 2025). A laboratory experiment published in the British Journal of Education Technology assigned 117 students to write essays with or without ChatGPT access. The AI group offloaded the thinking itself, even when the tool was prompted to assist rather than replace. The researchers called this “metacognitive laziness” (Hechinger Report/BJET, 2024). Each individual delegation is rational. The danger is aggregate and generational: a population that habitually delegates judgment may progressively lose the capacity to exercise it.
The consequences are professional, democratic, and self-reinforcing. In the ACCEPT trial, endoscopists who used AI-assisted polyp detection for six months saw their adenoma detection rate drop from 28% to 22% when AI was removed: six months of assistance produced a 20% decline in unaided diagnostic accuracy (Fortune/Lancet GI, 2025). This gap between measurement and reality is structural. The most rigorous labor market studies, including Anthropic’s March 2026 Economic Index, deliberately downweight augmentative AI use (half weight) relative to full automation (full weight), because augmentation preserves the job. But cognitive atrophy operates through precisely the augmentation channel: the worker keeps the position while progressively losing the capacity to perform it unaided. Displacement metrics are designed to measure job loss, not skill erosion, which means the atrophy risk is systematically invisible to the methodology best positioned to detect it. Aviation provides the longer track record. Air France Flight 447 killed 228 people in 2009 because pilots who rarely flew manually could not recover from a stall when the autopilot disengaged. FAA data showed over 60% of aviation accidents involved challenges with manual control tied to automation management (RAeS/Airline Ratings, 2021). Medicine and aviation at least have licensing regimes that mandate independent competence. Most professions do not.
The democratic consequences connect directly to Shade #5. Cognitive atrophy is what makes information collapse lethal. A population with strong analytical skills can survive a polluted information environment; a population without them cannot. The combination of scalable fabrication and degraded analytical capacity is more dangerous than either alone.
The self-reinforcing dimension may be the most dangerous, because the damage is invisible while the tool is available. A 2026 Aalto University study found that AI flattens the Dunning-Kruger curve: participants consistently overestimated their performance when using AI regardless of accuracy, and the most AI-literate users showed the greatest overconfidence (Computers in Human Behavior/Live Science, 2026).
A March 2026 study by Boston Consulting Group and the University of California, Riverside, published in Harvard Business Review, identified a distinct short-term mechanism that compounds the long-term atrophy. Surveying 1,488 full-time U.S. employees at large companies, the researchers found that 14 percent of AI-using workers reported what they termed “AI brain fry”: mental fog, difficulty focusing, slower decision-making, and headaches after intensive AI oversight. The core finding is a discordance between generation speed and verification speed. AI produces dense, substantive output in seconds. Validating that output properly requires the same deep cognitive engagement the human would have used to produce it, often more, because verification demands both comprehension of the content and comparison against contextual knowledge the AI may lack. The human is not doing less thinking. They are doing harder thinking, at a pace set by the machine rather than by their own cognitive rhythm. Workers with high AI oversight demands expended 14 percent more mental effort, reported 12 percent more mental fatigue, and experienced 19 percent greater information overload (Fortune, March 2026). Productivity peaked at two to three simultaneous AI tools and dropped beyond four, hitting the same cognitive ceiling as conventional multitasking. Those experiencing brain fry reported 33 percent more decision fatigue and 39 percent more major errors (The Decoder, March 2026). The study drew a critical distinction: burnout is emotional exhaustion from sustained overwork, and AI can reduce it by automating routine tasks. Brain fry is cognitive exhaustion from oversight overload, and AI causes it by producing output faster than humans can meaningfully verify. The same tool alleviates one form of strain and creates another. Organizations that respond to AI productivity gains by expecting more output per worker (as Block, Shopify, and Klarna have done) are accelerating the cycle: more AI output, more oversight demand, more cognitive depletion, more errors, less capacity to catch the errors. The generation-verification gap will widen as models become more capable, because more capable output requires more sophisticated verification while arriving at the same or greater speed. The biological processing capacity of the human overseer does not scale with the computational capacity of the tool.
The decay compounds across generations. When educators rely on AI-generated materials without critical adaptation, they lose the analytical skills they are supposed to transmit (IJRSI, 2025). A teacher who has lost the capacity for independent evaluation cannot teach it. The cognitive infrastructure of self-governance, once lost across a generation, requires skills to rebuild that the generation no longer possesses.
This concern is ancient, and the honest response to it deserves space. Socrates warned that writing would destroy memory. Critics said the same about calculators and the internet. In every case, humanity adapted: we lost some cognitive capacities and gained others. The Google Effect itself proved difficult to replicate in a 2018 large-scale study, suggesting the original findings may have been overstated (Nature, 2018 replication). The question is whether the current transition follows the same pattern or represents something qualitatively different. Previous tools automated discrete tasks: arithmetic, recall, navigation. AI automates the integrative functions, synthesis, evaluation, reasoning, that sit at the top of cognitive hierarchies. A student who uses a calculator still has to understand what to calculate. A student who asks ChatGPT to evaluate an argument has outsourced the evaluation itself.
The governance response is education reform. Finland’s national media literacy curriculum, integrating critical thinking about information sources from primary school onward, has been cited as a model (Frontiers in Communication, 2025). The deeper challenge is institutional: schools and professional training must distinguish between AI-assisted thinking, where the human retains evaluative control, and AI-replaced thinking, where the human accepts outputs without engaging the cognitive processes that produced them. The governed outcome is modest (+1) because every employer, platform, and productivity incentive pushes toward more delegation, and the cognitive costs are invisible until they accumulate.
Key tension: Every employer, platform, and productivity incentive pushes toward more cognitive delegation to AI. The costs are invisible until they accumulate, and by then the capacity to recognize the loss may itself have atrophied.
September 2026
The MIT Ad Hoc Committee on AI Use released its final report on August 13, 2026 and published the public summary on August 25. The committee’s finding is that generative AI “can produce credible solutions and provide reasonable responses to almost any written assignment in our undergraduate curriculum, including essays, math and science problems, proofs, and coding assignments.” The report documents “diminishing critical thinking, weakening memory, eroding confidence, and undermining mastery,” names the phenomenon “cognitive surrender,” and finds AI “accelerating isolation” through reduced office hours and fewer study groups. The committee adopted as an institutional principle that “productive struggle is essential to learning” and that the goal of AI in education is “augmentation, not automation, of human thinking.” The report cites a 2026 systematic review of 67 studies that concluded structured AI use in inquiry supports critical and creative thinking, while unstructured use produces cognitive offloading and weaker thinking, giving the shade’s developmental-versus-substitutive distinction a peer-reviewed evidence base (MIT Ad Hoc Committee on AI Use, Final Report, August 13, 2026; Forbes, August 25, 2026).
The atrophy has reached the builders. Anthropic’s Jack Clark returned from paternity leave in February 2026 to find that colleagues “hardly wrote code anymore” but instead managed five or six Claude instances, “which sometimes managed several more Claudes.” OpenAI researcher roon (Tarun Gogineni) wrote in September that “corrigibility has become a matter of faith” because Astra’s code uses “crazy meta-programming and abstruse primitives” that its own creators cannot readily follow: “we have really no choice but to ask another astra to read/use the outputs of astra 1.” Aryaman Arora, another OpenAI researcher, described Astra’s code as “uniquely unreadable.” The pattern this shade described for students and junior workers has now reached the people who build the systems. This is a recursive form of the metacognition problem: the population capable of correcting the substitutive pattern is itself losing the capacity to read what the systems produce (TIME, “Inside the Race to Make AI Build Itself,” August 7, 2026; Zvi AI #186, September 17).
The shade’s governed score moves from plus one to plus two on the basis that MIT’s response is an operating institutional example of the developmental-scaffolding approach the shade’s governance-side prose calls for.