Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise
Citation. Randazzo, S., Lifshitz-Assaf, H., Kellogg, K. C., Dell’Acqua, F., Mollick, E. R., Candelon, F., & Lakhani, K. R. (2025). Cyborgs, centaurs and self-automators: The three modes of human-GenAI knowledge work and their implications for skilling and the future of expertise (Harvard Business School Working Paper No. 26-036). https://doi.org/10.2139/ssrn.4921696
Summary. Randazzo and colleagues analyze how 244 global management consultants used a custom GPT-4 interface during a structured strategic problem-solving task. They identify three modes of human–GenAI co-creation — Fused/Cyborg, Directed/Centaur, and Abdicated/Self-Automator — distinguished by who selects what needs to be done and who identifies how it gets done. The central contribution is that superficially similar “human-in-the-loop” workflows produced different skill trajectories: Cyborgs built AI-related expertise while maintaining task expertise, Centaurs built domain/task expertise, and Self-Automators built neither.
Key claims
- GenAI shifts human–AI collaboration from recommendation-level interaction to continuous conversational co-creation across problem framing, data analysis, recommendation generation, iteration, extension, evaluation, and reporting.
- The study identifies two structuring questions for human–GenAI problem solving: who selects what needs to be done, and who identifies how it gets done.
- In the sample, 60% of professionals (N = 146) used Fused Co-Creation/Cyborg practices, integrating GenAI continuously throughout the workflow through iterative prompting, questioning, and refinement.
- 14% of professionals (N = 34) used Directed Co-Creation/Centaur practices, preserving human control over analytical logic while using GenAI selectively for mapping the problem domain, gathering methods information, extending ideas, and refining human-generated content.
- 27% of professionals (N = 63) used Abdicated Co-Creation/Self-Automator practices, concentrating interaction in later/reporting phases and often allowing GenAI to perform analytical and evaluative work with little substantive human modification.
- Co-creation mode shaped skilling: Fused/Cyborg work supported GenAI-related newskilling while maintaining task expertise, Directed/Centaur work supported task/domain upskilling, and Abdicated/Self-Automator work supported neither task nor AI expertise.
- Directed/Centaur users had the highest accuracy in business recommendations, while Directed/Centaur and Fused/Cyborg users were more persuasive than Abdicated/Self-Automator users.
Evidence & limitations
- Evidence comes from a working paper based on a field experiment with 244 BCG consultants using a custom GPT-4 interface on a fictional strategic-investment recommendation task; the authors analyzed 4,975 interaction logs and follow-up interview data.
- The PDF is Harvard Business School Working Paper 26-036; Crossref/SSRN provides public DOI provenance, while the PDF itself is a working-paper version distributed for comment and discussion.
- The task was a single-session experiment using April 2023 GPT-4, so the evidence cannot show whether co-creation modes remain stable over time, transfer across tasks, or change as AI capability improves.
- The participant pool was junior, high-performing management consultants, which may limit generalization to senior experts, lower-discretion work, or workers who refuse to use GenAI.
- The authors argue that abdication is not inherently problematic; whether it is efficient delegation or harmful deskilling depends on task routineness, time pressure, and whether the work sits inside or outside the model’s capability frontier.