Outsourced thinking: is AI making us dumb?

Joe Client was right, and I knew it before he finished the sentence.

It was early in my consulting career. I had walked into his office proud of a pitch deck I had spent days building. The charts were clean. The market was sized. The competitors were mapped. Nearly all of it had come out of Capital IQ, Factiva and PitchBook, and I had laid it out in the order that made it look like a point of view.

I am calling him Joe Client in this piece, and the name fits in a way I did not expect back then. All these years later, he is a client again. We have serviced his business recently.

Joe Client flipped through it, stopped early and said the thing I have never forgotten.

"This is basically regurgitated data."

He did not say it to be cruel. He said it because he had read the same databases himself, or paid people who had. What he wanted from me was the part that was not in any of them. What would you do? Where is this wrong? What happens on Monday morning if we follow this? Couldn't I have run these exact same searches (today's version of prompts) myself?

I did not have good answers that day. I had pages. He was right. No critical thinking, just regurgitated data.

I tell that story now because I hear Joe Client's voice every time someone hands me a beautiful AI draft at work.

The draft reads well. The structure is right. The Claude boxes have that familiar look. The sentences sound like a person who did the reading. Sometimes a person did. Sometimes the person asked a model, skimmed the answer, fixed two lines and sent it on, the same way I once pulled three databases into a deck and called the assembly thinking.

So here is the question I keep arguing with myself about. How is AI much different?

In one way, it is not different at all. We have been passing off pulled information as our own work for a long time. Ask my phone.

I do not remember routes anymore. I open Google Maps for places I have been a dozen times, follow the blue line, make the turns and get there. Ask me an hour later to give someone those directions and I would struggle. The trip happened. The route never stuck.

Phone numbers went the same way. I have three memorized: my wife's, my mom's and my dad's. Everyone else lives in my contacts. If my phone died, I could not call most of the people I talk to every week.

Maps did not make me dumb. Contacts did not either. They took small jobs off my brain, and I let them, and most days I would make the same trade again.

Joe Client's deck was a bigger job, and that is the difference I keep coming back to.

Capital IQ still made me do the handling. I had to pull the numbers, decide what went in, leave the rest out and stand in a room while a client asked me where a figure came from. I handled the material badly that day, but I handled it. Joe Client could catch me because he could ask a second question, and then a third, and watch the deck fall apart in my hands.

AI removes the handling.

I can ask for the answer, the argument and the counterargument before my coffee gets cold. I can present the result without ever having held the pieces. The second question is still a risk. The third one is worse. And the voice in the draft sounds so sure of itself that it is easy to borrow the confidence along with the words.

That borrowing is what the researchers are starting to measure.

A study out of ZHAW in January 2025, by Michael Gerlich, ran 666 people in the UK through a critical thinking assessment. People who used AI often scored lower. The link ran through offloading, handing the mental work over, and it hit the youngest group hardest, ages 17 to 25.

At MIT Media Lab, Nataliya Kosmyna's team watched it happen in real time. They put EEG caps on 54 people writing essays. One group used ChatGPT, one used a search engine, one used nothing. The ChatGPT group's brain connectivity was the weakest of the three, up to 55 percent lower than the no-tools group on some measures. After they finished, 83 percent of that group could not quote a single sentence from their own essay.

It was a small study, 18 people a group, published as a preprint in June 2025. It did not test intelligence. I am not going to pretend it settles anything. But I recognize the feeling it describes. I have turned in work I could not quote back either. Mine went to Joe Client.

The workplace version is quieter. In 2025, Microsoft Research and Carnegie Mellon asked 319 knowledge workers about 936 times they had used AI at work. The more someone trusted the tool, the less critical thinking they said they did. Their job had changed shape around them, from solving the problem to checking the answer and stitching pieces together. Supervising sounds like thinking. Some days it is. Some days it is just approving.

Which brings me back to my house.

My 9-year-old and my 8-year-old are growing up with a tool that can do the word problem, write the paragraph and explain the answer in plain language, all before they have had to sit with any of it. When I was stuck on homework, the stuck part lasted. I carried a wrong answer around until a teacher, a parent or a book corrected it. The correction landed because the struggle came first.

They may never need the struggle. That is the sales pitch, and some of it is true. Why should a kid waste an evening the way I wasted evenings?

Then I read the study from China.

Researchers followed 26,811 students in grades 7 to 12 for 30 months as generative AI worked its way into homework. Homework scores rose 18 percent. Homework time fell 30 percent, from about 64 minutes to about 45. Then the closed-book exams came, the ones where the tool stays outside the room. Scores there fell 20 percent within six months.

Better homework. Less learning. Nobody in that story had a Joe Client in the room asking the second question, until the exam did it for them.

I do not think the answer is to ban any of this. I use AI at work every day and candidly manage a team of humans and another team of agents. My team uses it. My boys use it. My boys ask me what Muse is when they see a billboard promoting Meta's latest tool on the way to school. I would rather they learn it with me than sneak it without me.

I think the answer starts with admitting which jobs we are handing over.

A route is a job I can live without. A phone number, same. The reasoning behind a recommendation is a different job. That reasoning is the thing Joe Client was buying, the thing an exam is testing, and the thing I need when a client, a partner or one of my boys asks me the second question.

Socrates worried about this trade around 370 BCE, when the new tool was writing. He said writing would "create forgetfulness in the learners' souls because they will not use their memories." He was right about the memory. He missed what writing would build. Every tool since has had the same split in it, from stones and abacuses to slide rules on the Apollo missions to the calculator schools tried to ban in the 1970s.

AI has the split too. The open part is which side grows in a kid who gets the answer first and the struggle second, or never.

I owe Joe Client for my side of it. He embarrassed me early enough that the lesson stuck. I learned to ask of my own deck what he asked of it. Where did this come from? What do I actually believe? What would I do on Monday?

That he is back on our client list now, all these years later, is the part of the story I like most. The bad meeting was not the end of the relationship. It was the start of me learning how to do the work he was asking for.

Our kids may need someone to play Joe Client for them on purpose, while they are still learning. So may the young analysts starting work now, whose first decks will be better than mine ever were, and whose Joe Client moment may never come if the deck is good enough to get through the meeting.

That is what I want to work out next. What do we make people do the slow way, on purpose, in a world that can do it for them in ten seconds?

Until then, a simpler question, the one I started with at my own kitchen table. What have you stopped remembering since you started handing the work over?

My original LLMs - large, loud models, ages 9 and 8 - in the back seat, mid-jam to "Du bist gut genug." They have had the back seat for years. The AI kind gets the same seat now, and the same voting rights. None.

A note on how this was written.

I wrote this post first. The story is mine, the opinions are mine, and the first draft was mine, word for word.

Then I used AI to make it punchy. It stayed long anyway, which tells you who was in charge. We went through multiple rounds. I cut lines, put lines back, argued with suggestions and kept the parts I could defend out loud.

That is how we use AI internally. Recommendations do not start with a prompt. The human sits in the front seat and drives. AI sits in the back seat with ideas, edits and the occasional bad suggestion you have to ignore. The critical thinking stays in the front seat, on every round, or the whole arrangement falls apart.

Joe Client would have spotted the other version in about four pages. Today's Joe Client is just as relevant in today's world.