A USC team ran a within-subject experiment with 20 programmers, each solving tasks twice: once with an LLM, once without. The LLM-assisted runs had significantly shorter idea-generation periods (p=0.0004) and fewer creative moments (p=0.002). The number of distinct ideas in the final solutions was roughly the same in both conditions.
The output did not move. The thinking that produced it collapsed.
Why I read this twice
Most leaders I work with are measuring AI adoption by the wrong instrument. They look at deliverables. The deck shipped. The model compiled. The memo reads well. The output is the part AI is protecting. The process is the part it is quietly removing.
If the ideas-per-solution count is flat and the ideas-per-minute count went up, your team is moving faster across the exact same ground. That is not productivity. That is foreshortening. And the paper title is not the researchers' phrase. It is a participant's own description of working with the model: like taking the path of least resistance.
People notice it in their own work and keep doing it anyway. That is what makes this hard.
A coaching call from last week
A VP of product at a Series C analytics company. Smart, calm, hires well. He was complaining that his PMs' specs had started reading like each other. Same structure. Same edge cases. Same blind spots. He asked me what changed.
I asked him a different question. How long does your team sit with a problem before they open Claude or ChatGPT?
He went quiet. Then he said, "I do not actually know. But I do it in under a minute now myself."
There it is. Fatjon Tony Kalemaj would tell you the same thing I told him on that call: the model is not the issue. The two-minute pause your team used to take, sitting with a half-formed idea before reaching for anything, is the asset that is leaving the building. Senior people protect their thinking time. Mid-level people optimize it away.
It does not show up in the next sprint review. It shows up two quarters later, when the work all looks the same and nobody can explain why the differentiated insights have stopped arriving.
The coaching move is not to ban the tool. It is to put the friction back. A timer on the wall. A rule that the first draft of any spec, memo, or plan is written longhand. A standing question in 1:1s: what did you think before you asked the model? If your team cannot answer, you have your data.
"LLM-assisted participants produced solutions containing roughly the same number of ideas as participant-generated ones."
The deliverable is fine. The muscle is not.