An MIT study on cognitive debt changed these five things about how we work.
The short answer: AI works best when you protect the order of the work. People who reason through a problem first and then bring AI in to sharpen it get better output while still building capability. People who prompt first get a finished artifact and no room to build the muscle. As a leader, your job is to apply the first pattern, because the second can be easy to miss until it fails in front of a client.
A post came across my feed claiming that four months of ChatGPT use dropped users' neural connectivity by 47 percent. The number turned out to be wrong and the graphic had nothing to do with the research, but the study underneath it is real, and one of its findings has changed how I think about the way my team works.
What is cognitive debt?
Cognitive debt is the term Nataliya Kosmyna and her colleagues at the MIT Media Lab use for what accumulates when you hand the thinking part of a task to a tool. Fifty-four people wrote essays using ChatGPT, Google search, or no tools, with EEG caps recording their brain activity across three sessions.
The result matters: 83 percent of the ChatGPT group couldn't quote from the essay they had finished minutes earlier, and that group reported the lowest sense of ownership over their own work. The people who used no tools reported the highest.
This is worth knowing before you share the finding: the paper is still a preprint, the sample is small, with only eighteen people completing the final session, and researchers in Vienna and Dresden published a formal commentary raising questions about the method and the reporting. The authors have asked people to stop attaching words like brain damage to their work, and to treat it as one early signal in a much older line of research on offloading, not as settled science.
The signal is still worth acting on, because it puts a name to something I recognize from our work. Sometimes the document exists but the thinking behind it hasn't happened.
Why is this a leadership problem and not a character problem?
Because you can't see it in the final work.
A finished document looks exactly the same whether or not there was reasoning behind it. Same structure, same grammar, often better polish. You find out which one you have when a client pushes on it and the person who wrote it has no understanding underneath the words.
That moment arrives late and it arrives in public. When the operations leader in the third chair wants to know what breaks on their end once this is live, and who absorbs it when it does. When procurement is already three steps ahead on year-three pricing. And the most technical person in the room catches the one claim that isn't quite right and stops believing in every statement after it. None of that is a writing problem, and it can’t be fixed in the moment.
There's a second cost that shows up further out. The experiences that build judgment are the ones AI is best at removing. Tasks like working a messy client problem before anyone knows the answer, getting the argument wrong, then rewriting it from the middle. A person three years into their career who skips that a hundred times will produce fine work, but will never develop the instinct to know when the work is wrong. That's the version of this I'm most concerned about, and I say that as someone actively handing more work to people earlier in their careers than I used to.
Does the order of work matter when using AI?
Yes, and it's the most useful thing in the study.
In the fourth session, participants who had written unaided first and then used AI showed a network-wide spike in connectivity, higher than the AI-first group across every frequency band measured. Same tool, opposite outcome, because the order changed.
It’s a small sample, so hold it loosely. I still use it, because it matches what I watch happen in our own work. Sequence is the whole variable, and it's the one a leader can set.
Five things we do about it at illi
We build B2B messaging for industrial, manufacturing, construction, and B2B technology companies, which means our work has to hold up in front of engineers who know their own operation better than we ever will. AI has made us faster. These are the five practices that keep it from making us hollow.
1. Draft before you prompt. Get the argument down first, on paper, badly. Then bring the tool in to pressure test it, find the hole, and argue the other side harder than the client will. The rough first pass is not wasted time. It's the thing that makes the second pass yours.
2. Get what your team knows out of their heads and into a document. We've built our buyer research, voice rules, and archetypes into instruction files the tools read before they write anything. AI starts from what we know about technical buyers instead of the internet's average opinion about manufacturing marketing.
That wasn't based on an afternoon of prompt writing. It's years of buyer interviews and decades of judgment that used to live only in a few senior people’s minds, written down for the first time. The tool was never the hard part. The hard part was pulling what my team knows into a form somebody else could use, including a machine.
If you do one thing on this list, do this one. Every team has expertise that lives only in the instincts of three or four people, and right now that expertise is the entire difference between your AI output and your competitor's. Undocumented, it does nothing for you. Written down, it compounds, and it keeps compounding after those three people are in a meeting or on vacation or gone.
3. Defend it out loud before it ships. Whoever owns the work explains it to the team without reading from it, then describes what they considered and rejected to get there. If what they rejected isn’t clear yet, we don't have a message yet, we have a draft.
4. Protect the reps on purpose. For people early in their careers, some passes are assigned unaided by AI, and we say why. Not as a rule about tool use, as a rule about which muscles we're building this quarter. The struggle is the training, and if I remove all of it I'm optimizing this month's output against next year's capability.
5. Say where you used it. Out loud, without apology. When you hide it, you lose the ability to see how work gets made, and quality review becomes guesswork. Naming it makes the conversation about craft instead of compliance.
What I'm still working out
I'm in the middle of moving from doing most of the work to leading the people who do it, and it's the hardest transition I've made. AI complicates it in a specific way: it makes doing it myself faster than teaching someone else, right at the moment I'm supposed to be teaching. The pull toward just handling it is stronger than it was two years ago, and giving in to it would cost me the thing I'm trying to build.
I don't have that solved. What I have is a bias toward setting the order of the work, protecting the reps for the people who need them, and refusing to accept a document nobody can defend, no matter how clean it reads.
For the record, I wrote this with AI in the loop. I read the paper first, then the commentary, then built the argument, then asked for help tightening it. That order is the point.
If you're leading a team through the same thing, I'd like to hear what you run in practice, not what your policy document says. Especially the practice that holds up on a Thursday when the deadline just moved.
Sources
- Kosmyna et al., "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task" https://www.media.mit.edu/projects/your-brain-on-chatgpt/publications/
- Project FAQ and author guidance: https://www.media.mit.edu/people/nkosmyna/overview/
