Is AI Making Us Dumber? From Socrates to GPS: What Science Actually Knows About Cognitive Atrophy

THE THESIS. No, AI is not making us dumb. But it is doing something real and measurable: it atrophies the specific skills we delegate and stop practicing. That distinction matters. All the serious evidence points the same way: the loss is specific, not general. The driver on GPS loses spatial memory, not intelligence. What's different about generative AI is scale: it is the first tool that can take over not one isolated function (remembering, calculating, navigating), but the thinking process itself. So the right question isn't 'AI, yes or no?' It's 'which skills do we keep in manual mode?' And that is a matter of design and habit — not destiny.

SOCRATES, THE CALCULATOR AND THE CABBIE. This panic is 2,400 years old. In Plato's 'Phaedrus', Socrates warns that writing will 'produce forgetfulness' because people will stop exercising memory. He was right in the small: we memorize less than a Greek oral poet did. He was wrong in the large: outsourcing memory made science, law and accumulated knowledge possible. The calculator replayed the script. The reference meta-analysis (Hembree and Dessart, 1986, covering 79 studies) found calculator use did not harm math skills — it slightly improved them — with a single exception in fourth grade. GPS is the more uncomfortable precedent. In 2000, Eleanor Maguire showed that London taxi drivers, forced to memorize the entire city, had an enlarged posterior hippocampus — the brain region central to memory and navigation. And a 2020 longitudinal study in Scientific Reports found the reverse: heavier habitual GPS use predicted steeper decline in spatial memory over time. The historical pattern is clear: the delegated skill atrophies; general intelligence does not. 'Use it or lose it' is literal.

THE MIT EXPERIMENT, MINUS THE HEADLINES. The study that set the debate on fire is 'Your Brain on ChatGPT' (MIT Media Lab, June 2025). Fifty-four university students wrote 20-minute essays while monitored with EEG — electroencephalography, a technique that measures the brain's electrical activity. Three groups: brain-only, search engine, and ChatGPT. The result: the ChatGPT group showed the weakest neural connectivity during the task, their essays were strikingly homogeneous and, per the study itself, 83% could not quote a sentence from their own text minutes later, versus about 11% in the other groups. The authors coined the term 'cognitive debt': like technical debt in software, it accumulates when you outsource too much thinking.

Now, what the study does NOT say. Lower brain activation during a task does not mean 'getting dumber': if you delegate work, you work less in that moment — that's nearly a truism, measured with electrodes. The sample is small (54 people, and only 18 reached the decisive session), the task is artificial, and a critical commentary by University of Vienna researchers challenges the EEG analysis, the reproducibility, and conclusions that outrun the data. The authors themselves publicly asked the press not to run 'brain damage' headlines. And the most valuable finding got buried: in the fourth session, participants who first wrote unaided and then used ChatGPT showed high activation and good recall. Order matters: think first and delegate later, and there's no debt; delegate from minute one, and there is.

WHEN ATROPHY TRULY MATTERS: DOCTORS AND PILOTS. The most serious signal doesn't come from student essays. In August 2025, The Lancet Gastroenterology & Hepatology published a study of 19 veteran endoscopists (over 2,000 colonoscopies each) at four Polish centers. After months of using AI-assisted detection, their adenoma detection rate — adenomas are precancerous polyps — in colonoscopies WITHOUT the machine fell from 28.4% to 22.4%: six absolute points, a 20% relative drop. That's not a minor stat: this rate is directly tied to colon cancer prevention. Mandatory caution: it's observational, it cannot prove causation, and its authors say so. But it is the first evidence that AI exposure may degrade a patient-relevant clinical outcome.

Aviation has been living this movie for 30 years. In 1997, Captain Warren VanderBurgh coined 'children of the magenta' for pilots who manage the autopilot but have lost their manual flying skills — named after the magenta line the flight computer draws. In 2009, Air France Flight 447 fell into the Atlantic with 228 people aboard: the autopilot disengaged over an iced-up probe, and the crew could not handle a basic aerodynamic stall. The FAA responded with official alerts urging more hand-flying. The lesson is central to this debate: aviation did not remove the autopilot — which has made flying safer than ever — it mandated manual practice. That is the model.

THREE SCENARIOS, NOT ONE. Every use of AI falls into one of three patterns, and it pays to separate them because headlines blur them together. One: substitution. You delegate the whole task and stop practicing it; the skill atrophies (the endoscopist, the GPS). Two: scaffolding. AI supports you while you learn, exposing you to more cases and better feedback; the skill grows (the MIT group that thought first and delegated later; the doctor who argues with the machine's diagnosis instead of accepting it). Three: redistribution. You drop the mechanical work and reinvest the time in strategy: the programmer who stops writing boilerplate and moves to designing the architecture. Workplace data suggests that today, with passive use, the first pattern dominates. A Microsoft and Carnegie Mellon study (CHI 2025, 319 knowledge workers, 936 real-world cases) found that more confidence in AI meant less critical thinking applied; more confidence in oneself meant more. A 2026 randomized controlled trial with 1,222 participants found AI help boosts immediate results but erodes persistence: after just ten minutes of assistance, people gave up faster on hard problems once left alone. And Gerlich's famous correlation (666 participants, 2025) between heavy AI use and weaker critical thinking is real, but it is just that — a correlation: perhaps those least inclined to think are those who delegate the most.

OUR READ. AI-driven cognitive atrophy is real, specific and — this is the key point — avoidable. We are not facing a collective lobotomy, but an accelerated version of the GPS dilemma with one difference of scale: this time the shortcut covers almost every mental task. In the short term the risk is serious and asymmetric: it hits hardest those still learning — the junior who never develops judgment — and the professions where errors cost lives. Companies celebrating AI productivity today may discover in three years that they have stopped producing seniors. In the long term, we remain optimistic — with conditions. Writing didn't make us dumb: it made us scientists. AI can do the same for thinking — free us from the mechanical so we can work on what we love — if we copy aviation: deliberate practice in manual mode for critical skills, tools designed to make you think before they answer, and one personal rule the MIT evidence supports: think first, delegate second. Socrates' question is still alive 2,400 years later. So is the answer: every memory technology took away a muscle and handed us a lever. This time the lever is enormous. All the more reason to choose carefully which muscles we refuse to let go.