Cognitive debt: the question the MIT study never asked

Aug 24

83% of them couldn't quote a single sentence they had written 60 seconds earlier. Nobody checked whether the writing was any good.

That stat comes from a preprint out of MIT, shared widely by Daniel Pink. Researchers put 54 people in EEG scanners over four months. One group wrote essays with ChatGPT, one used a search engine, one worked unaided. The unaided writers showed the strongest brain connectivity. The ChatGPT group showed the weakest neural engagement, and most of them could not recall their own sentences a minute after hitting submit. Researchers named the effect cognitive debt.

Jenni and I spent a good chunk of Episode 71 on it, and I was the skeptic in the room.

Start with the 83%. Of course they could not recite a sentence. Hand me something I read 60 seconds ago and ask me to quote a line from it and I probably could not either. Your brain handles information differently when it is composing than when it is reading or regurgitating. A gap between those two states is not a discovery.

My real question is the one the study did not answer. The researchers compared brain activity across three groups. They did not compare the essays. The people whose brains worked hardest may have produced something worth reading, or they may have produced garbage, because all they had to draw on was whatever was already in their heads. Working hard and working well are different outcomes, and this study measured one of them.

Jenni pushed back, and her argument is the stronger half of this. She is not disputing any individual essay. Her point is that the mechanism is the finding: hand over the thinking and the capacity to think degrades, and you will not notice the day it starts. She sat through a keynote this year from someone who lives inside these tools, and the talk had no thread holding it together, just soundbite after soundbite. She wondered whether that was the cost showing up on stage.

My counter is that she has no baseline. I have sat through scattered presentations since long before anyone had ChatGPT.

Where we land together is on how, not whether. Using these tools to pressure-test your thinking is a different act from using them to skip the thinking, and only one of those runs up a debt. So the question worth sitting with is what you do with the hour the tool just gave back to you. Put it into thinking harder about the message and you are fine. Put it into producing more of everything and that is where the debt accrues.

Also in this episode:

Jenni opened with new research from Michael Gill at Oxford's Saïd Business School, drawn from more than 250 studies across four decades, arguing that treating every broken rule as a punishable offense misses what is going on. He sorts workplace rule breaking into four types, and only one is really about self-interest. The others cover people bending a policy that blocks the outcome it was written to protect, people pressured into harmful shortcuts, and people doing the right thing for a customer. I told the story of a CEO who emailed me about how disappointed he was before he had asked anyone a single question. HR came around afterward wanting facts. The verdict was already out by then.
Gallup's latest on AI and culture reports that 24% of employees think it improved and 25% think it got worse, which is close to reporting that nothing happened. The manager data is where the story sits: employees whose manager champions AI are far more likely to say it transformed their work, 33% versus 4%. Jenni asked the question worth asking. Gallup sells manager training, and every culture problem their research finds seems to resolve into invest in managers. Is that what the data shows, or is it just always true of every workplace shift anyone has measured?
Tanya Pikula makes the case that writing was never really the job, and that the work has always been judgment. I agree, and I would frame it as bottlenecks. AI broke the writing bottleneck the way it broke the coding one, and judgment is the next bottleneck in line. Jenni took it somewhere less comfortable. Internal communications has spent years telling each other the job is advising rather than producing content, and we had that conversation almost entirely among ourselves. Nobody carried it to the HR conferences or the business conferences. So stakeholders still see the output, and the output is exactly the part AI can now do. Jenni writes and speaks on this if you want more of it. We also got into whether grammar is elitist, which we did not resolve, and probably will not.

Watch or listen to the full episode on Frequency, with timestamps, links to every story and the full transcript.