I use artificial intelligence every day. It writes my code, drafts my emails, plans my weeks, and tags my photos. I'm not here to argue against any of that. The tools are good and they're getting better. The point of this post is narrower. There's a small set of activities where letting the model do the work is quietly making us worse at the job, and the clearest example is the one most office workers spend half their week doing. Meetings.
The pattern is familiar by now. The pre-read goes out the night before. The model summarizes it for you. The meeting starts. The model takes notes. Halfway through, somebody pastes the transcript into another tool and asks for the key points. After it's over, the model writes the follow-up email. You sat in the room. You said a few things. You're not sure you could explain the decision in your own words an hour later, and if someone asked you what changed your mind, you'd have to scroll a transcript to find out.
That's the part that should worry us.
What the Research Actually Found
A team at the Massachusetts Institute of Technology Media Lab ran a study in 2025 that put electroencephalogram caps on people while they wrote essays. Some used a search engine. Some used artificial intelligence. Some used nothing. The group using artificial intelligence had the lowest brain engagement of the three, with a roughly 55% reduction in measured neural connectivity compared with the unaided group. After they finished, 83% of the artificial intelligence users couldn't recall a single sentence they had just submitted. The students who wrote unaided could quote themselves back almost verbatim. Time covered the headline numbers, and USA Today summarized the methodology.
A separate study of 666 adults published in MDPI Societies found a strong negative correlation between artificial intelligence usage and critical thinking scores. The mechanism was cognitive offloading. The more people used the tools to do the thinking, the less practice they got at thinking, and the less able they were to do it when the tools weren't available. Polytechnique Insights calls this cognitive atrophy. Stanford SCALE calls it lazy thinking. Different names for the same thing.
None of this means the tools are bad. It means there's a category of cognitive work where outsourcing it costs you something the receipt doesn't show. Reading is in that category. Thinking is in that category. Most of what happens in a good meeting is in that category.
The Pre-Read Is the Meeting
If a meeting has a fifteen page deck attached, the meeting is the deck. The thirty minutes on the calendar are for the parts that can't be written down. Disagreement. Trade-offs. Body language. The five seconds of silence after someone says the thing nobody wanted to say. The deck is the input. The conversation is the output.
When I let the model summarize the deck for me, what I get back is a competent abstract. What I don't get is the slightly off chart on page eleven. The footnote that doesn't quite match the headline. The assumption tucked inside an aside on page seven that the entire recommendation depends on. Models are bad at noticing what's missing. They're bad at noticing the thing that doesn't fit. They're bad at the small wait-a-minute reaction that you only get from sitting with a document long enough to feel it.
If I walk into the meeting having read the summary, I'm a passenger. I can nod along. I can ask one of the three questions the summary primed me to ask. I can't push back on the part the summary glossed over because I don't know it's there. The other people in the room who actually read the document have a different meeting than I do.
Reading the pre-read is the cheap part of the job. Skipping it raises the cost of every minute that follows.
Take Your Own Notes
I'm a big fan of meeting transcription. The audio is searchable. The action items are captured. People who couldn't make it can catch up. The model is genuinely better than I am at logging who said what.
That's not what notes are for.
Notes are how you think while someone else is talking. The act of choosing what to write down is the act of deciding what matters. If I write down the number, it's because I think the number is the point. If I write down the question instead of the answer, it's because the question is the part I want to come back to. By the time I've decided what to put on the page, I've already done a small piece of synthesis. I'm not just storing information. I'm processing it.
When the transcript writes itself, that step doesn't happen. I show up to the meeting with no filter, sit through it with no filter, and walk out with no filter. The transcript has every word. My head has a feeling. The two don't talk to each other, and a week later when the topic comes up again, I have to go read the transcript like a stranger.
People who take their own notes remember the meeting. People who let the model take notes remember they were in a meeting.
Don't Let the Summary Decide
The most common failure I see is post-meeting. Somebody pastes the transcript into a tool and asks it to extract the decisions and the action items. The output is sensible. It also flattens the meeting. It can't tell that the senior person on the call was clearly uncomfortable with the proposal but let it pass because there wasn't time. It can't tell that two of the action items are real and the third was a polite way to end the call. It can't tell that the decision was tentative, and the moment somebody quotes the summary in a follow-up email, the tentative decision becomes the official decision.
I've watched this play out enough to know it's not theoretical. The summary becomes the meeting. The nuance that was in the room dies somewhere between the transcript and the recap. By the time anyone notices, the project is moving on the basis of something nobody actually agreed to.
If a meeting matters, somebody who was in the room has to write the recap. Use the transcript as a check. Use the model to clean up the prose. But the editorial choices about what was decided, what's still open, and what people actually meant should come from a human who was there. Otherwise you're not running the meeting. You're rubber-stamping a software interpretation of it.
What I Actually Do
Three rules, none of them clever.
One. I read the pre-read. The whole thing. I don't speed-read it and I don't summarize it. If a deck is fifteen pages, I block fifteen minutes. If it's a memo, I read it twice. The first read tells me what it says. The second tells me what it's missing. I don't show up to a meeting I haven't done the homework for, and I push back when other people do.
Two. I take my own notes. Pen on paper most of the time, a clean text file when I have to type. Short. Bullets. Questions in the margin. I'm not trying to capture what was said. I'm trying to capture what I thought about what was said. The transcript can have everything else.
Three. If I have to send a recap, I write it before I look at the transcript. Then I check the transcript to make sure I didn't miss anything material. The recap reflects what I think happened. The transcript catches anything I missed. The order matters. If you start with the transcript, the recap is just a polished version of the model's interpretation, and we're back where we started.
The Part That's Worth Saying Out Loud
The reason any of this matters isn't the meetings themselves. Meetings are a stand-in. Reading carefully, taking notes, writing your own recap. These are the same skills that let you understand a hard document, hold a complicated argument in your head, and form an opinion you can defend. If you stop practicing them at work, you stop being good at them everywhere else too. The book gets harder. The conversation gets shallower. The judgment call gets fuzzier.
The Massachusetts Institute of Technology study found that the brain activity people lost while using artificial intelligence didn't fully come back when the tools were taken away. Natalie Wexler put it well. The skill you don't use is the skill you don't have. If your brain doesn't get the reps, eventually it can't do the lift.
I'll keep using the tools. I use them constantly. But the parts of the job where the work is the thinking, I'm going to keep doing myself. The meeting is one of those parts. So is the pre-read, the notes, and the recap. Letting the model do them isn't a productivity hack. It's a slow trade where you give up something you can't easily get back, and the only thing you save is fifteen minutes you were going to spend reading the deck anyway.