Do humans and AI divide labor differently in knowledge work?

A field study of 244 global management consultants at Boston Consulting Group, using a custom GPT-4 platform that logged "4,975 human–AI interactions" across 244 problem-solving sessions, found that professionals "engaged in three distinct modes of human–AI knowledge co-creation: Fused Knowledge Co-Creation, Directed Knowledge Co-Creation, and Abdicated Knowledge Co-Creation." In practice these modes surfaced as three types the researchers named "Cyborgs," "Centaurs," and "Self-Automators." The same superficial setup — a consultant working with a GenAI chatbot on a CEO investment memo — produced three structurally different collaborations, not one generic "human-in-the-loop" pattern.

The paper derives the typology from "two fundamental questions that structure any collaborative problem-solving dynamic between human and machine: Who selects what needs to be done? and Who identifies how it gets done?" The authors argue these questions "reveal the underlying architecture of human–AI collaboration, showing not only what GenAI is used for but also who steers the workflow and how the division of labor unfolds in practice" — the same tool used for the same task can sit at different points on both axes depending on who the consultant lets drive task selection and task method.

This gives a second typology of human-GenAI work modes alongside How are national lab staff actually using generative AI?, which sorts use by what the tool is asked to do (bounded Q&A versus multi-step delegated work) rather than by who controls problem framing versus method. The two frameworks could describe the same population differently: a "workflow-agent" user in the Argonne scheme could be a Centaur, a Cyborg, or a Self-Automator in this one, depending on whether the consultant or the AI is setting the subtask and choosing the approach.

The excerpt (introduction and method, ahead of the results section) does not yet show how the modes were operationalized from the interaction logs, how many of the 244 consultants fell into each mode, or how stable a given consultant's mode was across a session. It is one task (a fictional brand-investment memo), one firm (BCG), and one model generation (GPT-4, April 2023), so the three-mode taxonomy should be read as what emerged in this setting, not as an exhaustive map of how all knowledge workers divide labor with GenAI.