Do different GenAI collaboration modes produce different skill gains?

Within the same field study of 244 Boston Consulting Group consultants, the three modes of human-GenAI collaboration (Do humans and AI divide labor differently in knowledge work?) carried three different skilling consequences. The paper states it directly: "Centaurs were upskilling themselves and increasing their current domain expertise. Cyborgs were newskilling themselves in GenAI-related capabilities, acquiring a new area of AI expertise. Self-Automators were not increasing either their domain expertise or AI expertise."

The authors tie this to the same "human-in-the-loop" surface description hiding different underlying work: because the modes differ in who selects the task and who decides the method, each leaves the consultant practicing a different skill. A consultant who keeps directing the AI's method while judging the output is argued to be the one still exercising — and so building — domain judgment; one who merges closely with the tool's own process is argued to be the one building AI-specific capability instead; one who hands off both decisions is argued to be the one left practicing neither.

This sits next to How does generative AI actually change worker skills?, which found four self-reported outcomes (development, maintenance, erosion, revaluation) among 38 Dutch knowledge workers across varied professions and tools. That paper ties outcome to how a worker subjectively uses GenAI in general professional practice; this one ties outcome to a specific, structural variable — which of two decisions (what to do, how to do it) the consultant retains — measured from logged interactions and interviews on one task, not self-report alone. Read together, the earlier paper's binary framing of "upskilling versus deskilling" can be read as particular cases of this paper's mode-driven split: a Centaur's experience would read as development, a Self-Automator's as erosion or stagnation.

The excerpt reports the skilling finding as a summary claim ahead of its results section; it does not yet show how skill change was measured (self-report, interview coding, or logged behavior change), over what timeframe, or whether consultants' modes were stable enough to produce durable skill differences rather than one-task differences. The claim should be read as what this field study found for consultants on one memo-writing task with GPT-4, not as a general law that directing an AI's method always builds domain skill.