Does splitting work with an AI, blending with it step by step, or just handing it off change what skills you actually keep?

This explores whether the way you split work with AI decides what you end up learning. The terms come from field studies of consultants using GPT-4. Centaurs keep a clear line between their own tasks and the AI's. Cyborgs interleave with the model sentence by sentence. Self-Automators hand over the whole task and accept the result. The short version usually reported is that Centaurs deepen their domain expertise, Cyborgs gain new skills in steering and questioning AI, and Self-Automators gain little of either. The corpus has no paper on those field studies, though, so treat that comparison as background rather than something this collection shows.

What the corpus does have is the mechanism that explains why the modes would diverge. Does passive GenAI use risk causing worker deskilling? draws on an older idea about tools. Some tools are complementary: you get better at the task even after you put the tool down, the way an abacus trains mental arithmetic. Others are competitive: you only perform well while you hold them, the way GPS can erode your sense of direction. Krakowski argues that generative AI can be either one, and that what decides it is how actively the person engages, not the model itself. Passive use makes everyone's output look alike and deskills junior professionals in particular. Dialogue-based use, where you question and push back on the model, keeps skills that last beyond the tool.

Mapped onto the three modes, this explains the pattern. Self-Automators are the clearest case of passive use: the AI does the thinking, and they get the output without the practice. Centaurs keep the judgment-heavy parts for themselves, so their domain skill keeps getting exercised. Cyborgs are in constant back-and-forth, the kind of dialogue Krakowski says keeps skills alive. The catch is that much of what Cyborgs learn is about working with the AI rather than about the underlying subject.

The less obvious lesson concerns who is at risk. Deskilling falls hardest on junior people. Experts who drift into self-automating already have skills to fall back on. Novices who start that way may never build those skills, because the AI does the very tasks that used to train them. So the question for a team is not whether people use AI but which mode its juniors default to. The rest of this retrieval set is about AI agents learning skills, not humans, so it doesn't add to this question. A search for 'deskilling', 'AI at work', or 'expertise' would likely turn up more relevant material in the collection.