Last week my kid asked ChatGPT to explain a homework question, then asked me if the answer was right. I did not have a clean reply. I use these tools all day at work. I also want him to still be able to sit with a hard sentence when the screen is closed.
I just wrote Still Mortals about the other side of this: OpenAI's chief scientist calling the models an alien mind, and a few labs growing something that might one day be conscious. This note is about us. A paper posted on 3 September 2026 by Ricard Solé, Michael Levin, David Krakauer, and colleagues asks what happens to a population of people as ChatGPT-like tools spread.
Their title is "Large-Language Models as a Cognitive Virus." That sounds like an infection. They say they do not mean a parasite. They mean a habit that jumps from person to person, the way a catchy phrase does, or a workplace rule. Once your team expects every email to be drafted by a model, you look slow if you write the first line yourself.
They sort people into three piles, which maps onto what I see in IT.
- No model, or almost none. You still draft, search, and decide the old way.
- You use it and you still think. You ask for a draft, then you check, cut, and own the result. If the tool is down, you can still close the ticket.
- You cannot skip it. The first move is always the model. Checking feels optional. A day without it is a lost day.
People move between those piles. A colleague shows you a trick, a vendor includes the copilot in the license, a manager asks why the write-up took two hours. That pushes you toward the model. Practice, pride, or a bad hallucination can pull you back. The paper adds one more idea I recognize from offices: independent work is easier to keep when the people around you still do it. If everyone else pastes, writing from scratch starts to look like a hobby.
That last idea is what makes their math able to snap. In their equations, once enough people live in the third pile, going back can take more pressure than it took to get there. Think of an inbox that now assumes every reply is model-assisted. Turning the tool off is not a private preference. It is a fight with the workflow.
They distinguish using the model as a tutor, where you rebuild the argument in your own words, from letting it write, judge, and remember for you. I called the first of those partnership in the last note. The second is how a team quietly forgets how to do the job.
What this paper does not measure
They did not survey a country. They did not time how long it takes a helpdesk to write a ticket with and without a model. The numbers in the figures are chosen to show the shape of the equations. Change those numbers and the cliff becomes a gentle slope. So we do not know if a real snap exists. We do not know, if it does, whether we are near it.
That is the current state. The paper says lock-in is possible. It does not say it has happened. Reading their curves as proof would be like reading "alien mind" as a soul.
What I can do on Monday is smaller, and I want it anyway. Can this person still close the ticket if the model is down. Is checking the output part of the job. Did we accept a draft, or did we finish the work. Those questions do not settle the alien-mind argument in Still Mortals. They only tell us whether we can still think when the model is not in the room.