COBURNICUSMost of us can’t name five phone numbers from memory anymore. I can’t. I handed them to my phone years ago, and honestly, it was a great trade. I don’t miss them.
But we kept trading. We gave up directions, then a good chunk of our attention. Now we’re handing over the part where we read the thing, think about the thing, and decide what we believe about it. Each trade made sense on the day we made it. I made this film to follow the latest one all the way down, because at the bottom of it is a question I can’t shake: who checks the machine when nobody knows anything?
The phone took storage
The story about phones wrecking our memory is a little too neat. “Digital amnesia” was coined in 2015 by Kaspersky, a cybersecurity company, off surveys it commissioned. A famous 2011 Science paper on looking things up had its first experiment fail to replicate, twice (Camerer et al., 2018; Hesselmann, 2020). The phone changed what our brains bothered to keep, and we still did our own thinking.
The feed competes for attention
The biggest review so far, in Psychological Bulletin in 2025, covered 71 studies and nearly 100,000 people, though only 14 looked at thinking skills. Heavier short-video use was linked to poorer attention and weaker impulse control, and feeling you couldn’t stop was a much stronger signal than hours watched.
The chatbot goes after the thinking
The essay you didn’t write. Researchers at the MIT Media Lab (Kosmyna et al., a preprint) had 54 people write essays with ChatGPT, a search engine, or nothing. Right after writing, 83 percent of the ChatGPT group couldn’t accurately quote a single sentence from their own essay, against about 11 percent in the other groups. The researchers asked journalists not to call this “brain rot,” and lead author Nataliya Kosmyna has said they didn’t find any. They found people who weren’t really there when the essay was written.
The surprise test. In a randomized trial in Brazil (Barcaui, Social Sciences & Humanities Open, 2025), 120 undergraduates studied unfamiliar material, half with ChatGPT. Forty-five days later, the 85 who were still around got a surprise test. The traditional group scored 68.5 percent and the ChatGPT group 57.5 percent, eleven points lower. The chatbot group finished faster and kept less.
The practice that didn’t transfer. At a high school in Turkey (Bastani et al., PNAS, 2025), nearly a thousand students practiced math with plain ChatGPT, with a tutor version that gave hints instead of answers, or with nothing. On practice problems the ChatGPT group did 48 percent better. On the real exam, with no AI in the room, they did 17 percent worse than students who never had it. The tutor version, the one that made them do the work, didn’t cause that drop.
The experts. In Poland, 19 experienced doctors, each with more than 2,000 colonoscopies behind them, found precancerous growths in about 28 percent of the procedures they did on their own before AI polyp detection arrived. In the three months after it arrived, their rate without the AI averaged about 22 percent (Budzyń et al., Lancet Gastroenterology & Hepatology, 2025).
Surrender. Steven Shaw and Gideon Nave at Wharton call it cognitive surrender: adopting the machine’s answer as your own with barely any scrutiny. In their experiments, people working logic problems with a sometimes-wrong chatbot still went along with it about four times in five when it was wrong. On their own they got about 46 percent right. A wrong chatbot dragged that to 31 percent.
None of this shows a chatbot shrinking anybody’s brain. No study has found that. What older studies show is that adult brains change with practice. London cab trainees who passed the Knowledge grew gray matter in the back of the hippocampus (Woollett and Maguire, 2011), and people who learned to juggle grew gray matter that partly faded when they stopped (Draganski et al., 2004).
Where do the experts come from?
I’ve been writing software for a long time, and I learned most of it by being the most junior person in the room, getting the boring work, and having somebody senior look at my code and sigh. The boring work was the curriculum. It’s also what AI is good at.
Stanford’s Digital Economy Lab, using payroll records, found in its August 2026 update that employment for 22- to 25-year-olds in the most AI-exposed jobs was about 19 percent below where it would be if it had kept pace with peers in less-exposed jobs. A year earlier the gap was 15 percent by the same measure. It shows up mostly as fewer young people being hired, and IBM said it planned to triple US entry-level hiring this year.
On a spreadsheet, skipping beginners looks fantastic for about ten years. Then the seniors retire.
Two things that didn’t make the film
The research turned up a stronger version of the tutor result. In a Harvard physics course (Kestin, Miller et al., Scientific Reports, 2025), 194 students in a crossover trial learned more than twice as much from an AI tutor as from an active-learning class, in less time. Read next to the Turkey study, it raises a question about how much depends on whether the tool makes you do the work.
And a correction to a popular headline: the McGill GPS study (Dahmani and Bohbot, Scientific Reports, 2020), which found heavy GPS users had worse spatial memory three years later, never scanned anyone’s brain. “GPS shrinks your hippocampus” is not something that paper showed.
The case against
People have made this argument for about 2,400 years. In Plato’s Phaedrus, Socrates tells the story of an Egyptian king who warns that writing will weaken memory and give people only the appearance of wisdom. Writing gave us libraries and science in exchange. Many of these studies are small or early, and in a small follow-up session of the MIT study, people who wrote on their own first and then switched to ChatGPT stayed engaged and used it more thoughtfully.
I use this stuff every day. What still worries me is that a book says the same thing to everybody and you can see who wrote it. A chatbot is private, it changes without notice, and a small number of companies control it. In Nature Medicine, Alber et al. showed that swapping a thousandth of one percent of a training set for medical misinformation produced models that spread more harmful medical errors while scoring just as well on standard benchmarks.
So here’s the question I’m still sitting with. Pick something you asked a machine this week. Could you tell if the answer was wrong? And if you couldn’t, who could?
Sources
- Sparrow replication, Camerer et al. (2018): https://www.nature.com/articles/s41562-018-0399-z
- Hesselmann replication, PeerJ (2020): https://pmc.ncbi.nlm.nih.gov/articles/PMC7651475/
- Short-video review, Psychological Bulletin (2025): https://psycnet.apa.org/fulltext/2026-89350-001.pdf
- Kosmyna et al., Your Brain on ChatGPT (MIT, preprint): https://arxiv.org/abs/2506.08872
- Barcaui, ChatGPT as a cognitive crutch (2025): https://doi.org/10.1016/j.ssaho.2025.102287
- Bastani et al., PNAS (2025): https://www.pnas.org/doi/10.1073/pnas.2422633122
- Budzyń et al., colonoscopy deskilling (2025): https://wrap.warwick.ac.uk/id/eprint/191005/
- Shaw & Nave, cognitive surrender (2026): https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646
- Woollett & Maguire, London cab trainees (2011): https://www.cell.com/current-biology/fulltext/S0960-9822(11)01267-X
- Draganski et al., juggling (2004): https://pubmed.ncbi.nlm.nih.gov/14737157/
- Dahmani & Bohbot, GPS and spatial memory (2020): https://www.nature.com/articles/s41598-020-62877-0
- Stanford Digital Economy Lab, August 2026 update: https://digitaleconomy.stanford.edu/news/canariesaug26/
- Kestin, Miller et al., AI tutor in physics (2025): https://www.nature.com/articles/s41598-025-97652-6
- Alber et al., Nature Medicine (2025): https://www.nature.com/articles/s41591-024-03445-1
Subscribe to get the next film and the research behind it. The full documentary and the rest of the films are on YouTube: https://www.youtube.com/@coburnicusmind?sub_confirmation=1