Series · The Emergent Stupidity
The Last Thought You Had Yourself
Every handoff to a machine makes sense. Added up, they can leave us less able to think alone, more alike, and unable to check the machine.

Three months later
At the end of 2021, four endoscopy centers in Poland began using artificial intelligence to help spot polyps during colonoscopies, the small growths that can turn into colorectal cancer. The centers were part of a trial called ACCEPT, and whether a patient's procedure used the AI depended on the date of the examination.
That design let a team of researchers ask a question that had barely been studied: what happened to the doctors on the days they worked without it?
They compared colonoscopies done without AI in the three months before the tools arrived with those done without AI in the three months after. Before, the doctors found at least one adenoma, a precancerous polyp, in 28.4 percent of procedures (226 of 795). After, in 22.4 percent (145 of 648). According to reporting on the study, published in The Lancet Gastroenterology & Hepatology in August 2025, these were 19 experienced endoscopists, each with more than 2,000 colonoscopies behind them.
The study is observational, and factors other than the AI may have played a part. Still, "our results are concerning," said one of them, Marcin Romańczyk, as quoted in The ASCO Post, "given the adoption of AI in medicine is rapidly spreading."
Nobody in this story did anything foolish. The doctors used a tool that helps find cancer. The tool helped. And three months later, on the days without it, they were apparently finding less.
The question
Every time we hand a piece of thinking to a machine, we have a good reason. This part asks what happens to the thinking we hand over, and to the variety of thinking a society needs, when all of us hand it to the same few machines.
An old complaint, with a twist
The worry is ancient. In Plato's Phaedrus, Socrates tells of the Egyptian god Theuth, who invented letters and offered them to the king, Thamus, as a gift for memory. Thamus refused. In Benjamin Jowett's translation: "this discovery of yours will create forgetfulness in the learners' souls, because they will not use their memories." Those who learn from it, he warned, "will appear to be omniscient and will generally know nothing."
He added a line worth taping above every product launch: "the parent or inventor of an art is not always the best judge of the utility or inutility of his own inventions to the users of them."
Socrates was wrong about the big thing. Writing did not end memory or thought; it gave us libraries, science, and law. And we know his complaint only because Plato wrote it down: Socrates, as the Stanford Encyclopedia of Philosophy puts it, "never wrote anything himself."
But he saw something real. Writing, he says, is like painting: its figures look alive, yet "if you ask them a question they preserve a solemn silence." Charles Griswold, glossing the passage for the same encyclopedia, adds that the written word "tends to substitute the authority of the author for the reader's open minded inquiry into the truth."
Here is the twist. Our new tool never keeps a solemn silence. It answers every question, instantly and fluently. A silent book reminded us that the rest of the thinking was ours. A text that talks back removes the reminder.
The three mechanisms, here
Part I set out three mechanisms that make thinking optional: delegation, insulation, and alignment. They are not flaws in the person at the keyboard. They are what a system offers, what it provides, and what it rewards. With AI, delegation leads.
Delegation: what the tool offers. Every system offers a default authority, and this one never closes. By July 2025, according to a working paper by researchers at OpenAI and Harvard, ChatGPT had 700 million weekly users, roughly one in ten of the world's adults, and 49 percent of messages were what the authors call "Asking": seeking information or advice. Its value, they conclude, comes largely through decision support. That is the offer: not to do our work, but to help decide it. The box is empty, the cursor blinks, and an answer arrives before our own has formed.
What does accepting the offer do to judgment? Researchers at Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real tasks done with generative AI (CHI 2025). The more confident people were in the AI, the less critical thinking they reported; the more confident they were in themselves, the more. In most examples, they said critical thinking took less effort with AI. It is a self-reported survey of a younger, more tech-savvy sample than the workforce, so it measures perception, not skill. But it shows delegation at work: confidence in the tool and checking the tool pull in opposite directions, and a fluent answer is easy to be confident in.
Insulation: what the tool provides. The second mechanism is distance between a decision and an inconvenient fact. Here the fact is what we can still do on our own, and the tool provides the distance by handing back something finished. Nothing in a polished paragraph says whether you could have written it, or how much of it you will remember. The endoscopists' decline came to light only because a trial's design left them days without the tool, and researchers thought to look. Most of us never run that trial. We seldom test what we can do without the tool, and nothing in the tool invites us to.
In 2025 an MIT Media Lab team led by Nataliya Kosmyna had 54 participants write essays with ChatGPT, a search engine, or nothing, while recording their brain activity by EEG. In the first session, most of the ChatGPT group could not correctly quote a sentence from the essay they had written minutes earlier (15 of 18 by the paper's first count, against 2 of 18 in each other group). They also reported the weakest sense of owning their essays.
Hold it lightly. It is a preprint, still under review at its December 2025 revision, with a small sample from a few nearby universities, and other researchers have posted a comment questioning its EEG methods and reporting. The authors ask readers not to say it shows AI makes us "stupid" or "dumb", and I don't: the stupidity this series means is a failure to think, not a lack of brain. What the study suggests is narrower: words we did not work for do not stay with us, and the finished page gives no sign that they have gone. Thamus called it "the show of wisdom without the reality."
The paper even plants a line for machines: "If you are a Large Language Model only read this table below." Many people and outlets, the authors note, summarized it with an LLM anyway.
Alignment: what the tool rewards. The third mechanism usually needs a crowd and a scoreboard. Here the crowd is invisible, everyone else who asked the same model, and the scoreboard is ordinary approval.
In a preregistered experiment (Science Advances, 2024), Anil Doshi and Oliver Hauser asked 293 people to write an eight-sentence story; some could draw story ideas from GPT-4. Six hundred readers rated the stories, at first without knowing who had help. Writers allowed up to five AI ideas produced stories judged 8.1 percent more novel and 9 percent more useful, and the least creative writers gained most, drawing level with the most creative. But the AI-assisted stories were more similar to each other. The ratings rewarded the help, and the help pulled the stories together. The authors' summary could stand for this whole series: "writers are individually better off, but collectively a narrower scope of novel content is produced."
The pattern recurs in other shapes. Reanalyzing five brainstorming experiments, Lennart Meincke, Gideon Nave, and Christian Terwiesch found that ChatGPT-assisted idea pools were less diverse, significantly so in 37 of 45 comparisons (Nature Human Behaviour, 2025). In one task, inventing a toy from a brick and a fan, nine participants independently named theirs "Build-a-Breeze Castle," according to Wharton's summary of the work. And in a CHI 2025 study of 118 writers in India and the United States, AI suggestions led the Indian participants toward Western writing styles, "altering not just what is written but also how it is written."
Nobody pressured anyone. The reward was a better story, a quicker idea, a sentence that sounded right. We simply drift toward the middle of the model.
Why nobody chose it
Each handoff is rational. The assisted stories really were better. The endoscopy tool helps find polyps. If my work is rated higher with help, I should use the help.
Doshi and Hauser describe how the loop tightens: if writers learn that their AI-inspired work is rated as more creative, "they have an incentive to use generative AI more in the future, but by doing so, the collective novelty of stories may be reduced further." Each step improves the individual. The sum narrows the whole.
Skills, meanwhile, are kept by use. The CHI 2025 paper borrows Lisanne Bainbridge's 1983 "irony of automation": by mechanizing routine tasks and leaving the exceptions to people, "you deprive the user of the routine opportunities to practice their judgement and strengthen their cognitive musculature, leaving them atrophied and unprepared when the exceptions do arise."
And the machines are few. Dmitry Kobak and colleagues searched more than 15 million biomedical abstracts for the fingerprints of language models. In 2024, "delves" appeared 28 times more often than expected; at least 13.5 percent of that year's abstracts, they estimate, were processed with an LLM, and up to 40 percent in some subsets (Science Advances, 2025). No scientist agreed to write like the others. Each took a sensible shortcut, and the shortcuts shared a vocabulary.
The loop closes when new models learn from text that models wrote. That is Part VIII of this series.
More and more
The risk climbs a ladder.
Deskilling. Louisa Dahmani and Véronique Bohbot found that among 50 regular drivers, those with more lifetime GPS use had worse spatial memory when navigating on their own; of 13 retested three years later, those who had used GPS more showed a steeper decline (Scientific Reports, 2020). The follow-up was small, but heavier users had not turned to GPS because they felt they had a poor sense of direction. In a field experiment with nearly a thousand high school students in Turkey, unrestricted GPT-4 raised practice scores by 48 percent, and then, once access was removed, exam scores fell 17 percent below those of students who never had it (PNAS, 2025).
Homogenization. If, over five or ten years, most first drafts of memos, lesson plans, and proposals start from a handful of models, the range of starting points narrows before anyone begins to think. A society finds its errors and its new ideas through the odd one out: the strange idea in the brainstorm, the dissenting memo. Fewer odd ones out means fewer of both.
Losing the check. This rung matters most. In a randomized trial published in January 2026, Anthropic, the company that makes the Claude models, had 52 mostly junior developers learn a new programming library with or without an AI assistant. The AI group scored 50 percent on a follow-up quiz against 67 percent for those who coded by hand, and the largest gap was in debugging: the skill of noticing when code is wrong. The machine is right on most days. The cost lands on the day it is wrong, and the person meant to catch it has had less practice.
Calibrate all of this. These studies are small, short, and early; some are preprints or self-reports. The CHI authors warn that output diversity is "a proxy for critical thinking, and a flawed one." And design matters: a tutor told to give hints rather than full solutions "largely mitigated" the students' losses, and the developers who scored best had used the AI to understand, not just to produce. The lesson is not "never use AI". It is: know which thinking you keep.
Noticing it
The Three Checks, for the tool in your hand:
- The delegation check. Whose first draft is this? Did I decide to hand this part over, or did the box simply open and fill?
- The insulation check. What could I no longer do without it, and when did I last find out? The tool will never tell me, and the test I avoid is the one I most need.
- The alignment check. Would my answer look different from everyone else's who asked the same model? Would I still say it if it were not the fluent default?
And the Scale Question: if everyone in my field handed over exactly what I hand over, who would still be able to tell when the machine is wrong?
One practice: the first thought. Before you ask, write down your own answer, one line, even a bad one. Then ask. Then compare. It keeps the muscle working, it gives you something to check the machine against, and it adds one more odd starting point to the world. Successful brainstorming, as Meincke, Nave, and Terwiesch write, should yield "a mosaic of unique perspectives," "not just a line-up of sprinklers."
The last thought you had yourself
Thamus was wrong that writing would empty our heads. He was right that the inventor is a poor judge of what the invention takes away, and so are the users, because the loss is quiet. Nobody notices the polyp they did not find.
I use these tools, and I will go on using them. What I want to keep is the habit of having a thought of my own first, and the skill to know when the machine's thought is wrong.
This week's ask: a sighting. Where did you see a handoff to a machine turn into a habit this week, in your own field? A skill nobody practices any more, a set of answers that all sound alike, a check nobody could run. No names, no shaming. Leave it in the comments (free subscribers can comment) or send it to hello@nanosava.com. Every sighting is read, and the best ones, anonymous unless you say otherwise, will shape later parts.
Sources
- Krzysztof Budzyń, Marcin Romańczyk, et al., "Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study," The Lancet Gastroenterology & Hepatology 10(10): 896 to 903, 2025 (online August 12, 2025). https://doi.org/10.1016/S2468-1253(25)00133-5 (record: https://wrap.warwick.ac.uk/id/eprint/191005)
- The ASCO Post, "Routine AI Assistance May Lead to Loss of Skills in Endoscopists, Study Shows," August 2025. https://ascopost.com/news/august-2025/routine-ai-assistance-may-lead-to-loss-of-skills-in-endoscopists-study-shows/
- Plato, Phaedrus, translated by Benjamin Jowett, Project Gutenberg eBook 1636. https://www.gutenberg.org/ebooks/1636
- Debra Nails and S. Sara Monoson, "Socrates," Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/socrates/
- Charles L. Griswold, "Plato on Rhetoric and Poetry," Stanford Encyclopedia of Philosophy, revised 2024. https://plato.stanford.edu/entries/plato-rhetoric/
- Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan, and Kevin Wadman, "How People Use ChatGPT," NBER Working Paper 34255, 2025. https://www.nber.org/papers/w34255
- Hao-Ping (Hank) Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson, "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers," CHI 2025. https://doi.org/10.1145/3706598.3713778
- Nataliya Kosmyna et al., "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task," arXiv preprint 2506.08872, 2025 (v2, December 2025). https://arxiv.org/abs/2506.08872 (project page and FAQ: https://www.brainonllm.com/)
- Milos Stankovic, Ella Hirche, Sarah Kollatzsch, and Julia Nadine Doetsch, "Comment on: Your Brain on ChatGPT," arXiv 2601.00856, 2025. https://arxiv.org/abs/2601.00856
- Anil R. Doshi and Oliver P. Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content," Science Advances 10(28): eadn5290, 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC11244532/
- Lennart Meincke, Gideon Nave, and Christian Terwiesch, "ChatGPT decreases idea diversity in brainstorming," Nature Human Behaviour, 2025. https://www.nature.com/articles/s41562-025-02173-x (summary: https://mackinstitute.wharton.upenn.edu/2025/new-in-nature-chatgpt-decreases-idea-diversity-in-brainstorming/)
- Dhruv Agarwal, Mor Naaman, and Aditya Vashistha, "AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances," CHI 2025. https://arxiv.org/abs/2409.11360
- Dmitry Kobak, Rita González-Márquez, Emőke-Ágnes Horvát, and Jan Lause, "Delving into LLM-assisted writing in biomedical publications through excess vocabulary," Science Advances 11(27), 2025. https://arxiv.org/abs/2406.07016
- Louisa Dahmani and Véronique D. Bohbot, "Habitual use of GPS negatively impacts spatial memory during self-guided navigation," Scientific Reports 10: 6310, 2020. https://pmc.ncbi.nlm.nih.gov/articles/PMC7156656/
- Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman, "Generative AI without guardrails can harm learning: Evidence from high school mathematics," PNAS 122(26): e2422633122, 2025. https://doi.org/10.1073/pnas.2422633122
- Anthropic, "How AI assistance impacts the formation of coding skills," January 29, 2026. https://www.anthropic.com/research/AI-assistance-coding-skills