This one is one of my favorites because it directly tackles the issue of ai coding productivity in the real world: https://arxiv.org/abs/2507.09089

Interestingly, the same group tried to repeat the study and found that it was difficult to find ai coders willing to try doing things without AI. Those that did join told the researchers they selectively chose tasks they knew they could handle if AI was taken away, avoiding complexity. In just one year, self confidence has plummeted among most developers who have experience using AI coding tools.

https://metr.org/blog/2026-02-24-uplift-update/#wider-adoption-of-ai-has-made-it-more-difficult-to-measure-task-level-productivity

The blog frames this as saying, without evidence, that maybe the avoided tasks would be even quicker with ai than without, and that avoiding more complex work is hiding an even bigger increase in productivity, but I’m skeptical. The first study showed that simpler tasks were easier to automate, but that the slow down experienced during longer more complex tasks, and the frequency of such tasks occuring in the normal course of software development, meant that ai made developers slower overall.