Researchers at the MIT Media Lab had people write essays three ways: with an AI assistant, with a search engine, and with nothing but their own head. In the last session, they switched the groups around. The people who had done the thinking first and then used AI carried their work with them. The people who leaned on AI from the start and then had to work on their own came up short.
Same tool. Opposite outcomes. The variable was never the tool. It was whether the thinking had already happened.
That study is small, and it has not been through peer review yet, so hold it loosely. But it is not the only one pointing this way. A 2025 study out of Switzerland asked 666 people how much they used AI and how much they trusted it, then gave them a critical thinking test. The heavier users scored lower. The researchers think the reason is offloading: the more we hand over the thinking, the less we practice it.
What I take from that is not a warning about the use of AI tools. Instead, it is a clear signal about what the tools reward.
AI is a multiplier, and a multiplier does not care what it is multiplying. Give it clear thinking and it gives you leverage. Give it a vague ask and it gives you a polished version of the vagueness, on time, in the right format, convincing at a glance, and wrong in ways that take an audit to surface.
So the skill worth building is not prompting. It sits upstream of the prompt: framing the problem so it is worth solving, naming the constraints that make an answer right for us and not just right in general, and having a point of view on what good looks like when the work comes back. Anyone can generate ten options now. Knowing which one is right, and being able to say why, is the job.
That changes the shape of the work more than the amount of it. Engineers hit this first because their output was the easiest to generate, and researchers who studied them found the same trade: less code written, more code checked. Supervisory work is the term that has stuck. The rest of us are close behind.
There is one more move here, and I think it is the one most of us are missing.
We talk constantly about the prompt. We almost never talk about the rejection. When someone with real expertise looks at what came back, spots what is wrong, and can say why, something new gets created in that moment. Not the rejection itself, but the reasoning behind it. That reasoning is a constraint that did not exist five minutes earlier, and it is worth more than the output that triggered it. The output gets thrown away. The constraint stays, and you can hand it to the tool up front the next time, so the same correction does not get made twice. That is what compounds.
Right now, most of that value disappears. The correction happens in a chat window, the window closes, and the next person starts from nothing. We are making expert calls faster than we ever have and writing almost none of them down.
The fix is not glamorous. Write down why: not just what was decided, but what made it the right call and what would have changed it. I still catch myself skipping this part, usually on the days I am moving fastest, which are exactly the days it matters most.
We have been working this way on Ask Hubert, an internal assistant my team is building. The documentation lives in the repo with the product: markdown files holding the strategy, the backlog, the reasoning behind decisions we have already argued through, and the standards the work has to meet. When someone new comes to the work, or an agent is pointed at it, the reasons are already sitting there.
The design system I lead runs on the same approach. The documentation sits in the repo where agents can read it and on a documentation site where people can. We built it to teach rather than to hand out parts, with the decisions behind it out in the open, and it improves every time someone uses it and finds where it falls short.
AI will happily make us faster. Whether it also makes us sharper depends on the part it cannot do for us: deciding what is worth building, judging what came back, and leaving behind a reason the next person can use. That part is still ours. It is also the part that compounds.